Added GPL license, deb packaging script, complete GUI rewrite, added CLI mode, better default config options, config now lives in user folder, reduced syscalls by 5x.

.gitattributes

@@ -0,0 +1,6 @@
+# cmd.exe is sensitive to line endings — a run.bat checked out with LF can
+# misparse labels and multi-line constructs — while sh requires LF regardless
+# of the platform it is checked out on.
+*.bat text eol=crlf
+*.cmd text eol=crlf
+*.sh  text eol=lf

.gitignore

@@ -1,3 +1,4 @@
 /target
+/dist
 *.db
 config.toml

.vscode/launch.json

@@ -12,7 +12,7 @@
                 "args": [
                     "build",
                     "--bin=quicksearch",
-                    "--package=quicksearch"
+                    "--package=quicksearch-gui"
                 ],
                 "filter": {
                     "name": "quicksearch",
@@ -31,7 +31,7 @@
                     "test",
                     "--no-run",
                     "--bin=quicksearch",
-                    "--package=quicksearch"
+                    "--package=quicksearch-gui"
                 ],
                 "filter": {
                     "name": "quicksearch",

Cargo.toml

@@ -6,5 +6,8 @@ members = [
 ]
 
 [workspace.package]
-version = "0.1.0"
+version = "0.8.0"
 edition = "2021"
+license = "GPL-3.0-or-later"
+authors = ["Jeremy <jeremy@karsttech.com>"]
+repository = "https://code.karsttech.com/jeremy/quick_search.git"

LICENSE

@@ -0,0 +1,674 @@
+                    GNU GENERAL PUBLIC LICENSE
+                       Version 3, 29 June 2007
+
+ Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
+ Everyone is permitted to copy and distribute verbatim copies
+ of this license document, but changing it is not allowed.
+
+                            Preamble
+
+  The GNU General Public License is a free, copyleft license for
+software and other kinds of works.
+
+  The licenses for most software and other practical works are designed
+to take away your freedom to share and change the works.  By contrast,
+the GNU General Public License is intended to guarantee your freedom to
+share and change all versions of a program--to make sure it remains free
+software for all its users.  We, the Free Software Foundation, use the
+GNU General Public License for most of our software; it applies also to
+any other work released this way by its authors.  You can apply it to
+your programs, too.
+
+  When we speak of free software, we are referring to freedom, not
+price.  Our General Public Licenses are designed to make sure that you
+have the freedom to distribute copies of free software (and charge for
+them if you wish), that you receive source code or can get it if you
+want it, that you can change the software or use pieces of it in new
+free programs, and that you know you can do these things.
+
+  To protect your rights, we need to prevent others from denying you
+these rights or asking you to surrender the rights.  Therefore, you have
+certain responsibilities if you distribute copies of the software, or if
+you modify it: responsibilities to respect the freedom of others.
+
+  For example, if you distribute copies of such a program, whether
+gratis or for a fee, you must pass on to the recipients the same
+freedoms that you received.  You must make sure that they, too, receive
+or can get the source code.  And you must show them these terms so they
+know their rights.
+
+  Developers that use the GNU GPL protect your rights with two steps:
+(1) assert copyright on the software, and (2) offer you this License
+giving you legal permission to copy, distribute and/or modify it.
+
+  For the developers' and authors' protection, the GPL clearly explains
+that there is no warranty for this free software.  For both users' and
+authors' sake, the GPL requires that modified versions be marked as
+changed, so that their problems will not be attributed erroneously to
+authors of previous versions.
+
+  Some devices are designed to deny users access to install or run
+modified versions of the software inside them, although the manufacturer
+can do so.  This is fundamentally incompatible with the aim of
+protecting users' freedom to change the software.  The systematic
+pattern of such abuse occurs in the area of products for individuals to
+use, which is precisely where it is most unacceptable.  Therefore, we
+have designed this version of the GPL to prohibit the practice for those
+products.  If such problems arise substantially in other domains, we
+stand ready to extend this provision to those domains in future versions
+of the GPL, as needed to protect the freedom of users.
+
+  Finally, every program is threatened constantly by software patents.
+States should not allow patents to restrict development and use of
+software on general-purpose computers, but in those that do, we wish to
+avoid the special danger that patents applied to a free program could
+make it effectively proprietary.  To prevent this, the GPL assures that
+patents cannot be used to render the program non-free.
+
+  The precise terms and conditions for copying, distribution and
+modification follow.
+
+                       TERMS AND CONDITIONS
+
+  0. Definitions.
+
+  "This License" refers to version 3 of the GNU General Public License.
+
+  "Copyright" also means copyright-like laws that apply to other kinds of
+works, such as semiconductor masks.
+
+  "The Program" refers to any copyrightable work licensed under this
+License.  Each licensee is addressed as "you".  "Licensees" and
+"recipients" may be individuals or organizations.
+
+  To "modify" a work means to copy from or adapt all or part of the work
+in a fashion requiring copyright permission, other than the making of an
+exact copy.  The resulting work is called a "modified version" of the
+earlier work or a work "based on" the earlier work.
+
+  A "covered work" means either the unmodified Program or a work based
+on the Program.
+
+  To "propagate" a work means to do anything with it that, without
+permission, would make you directly or secondarily liable for
+infringement under applicable copyright law, except executing it on a
+computer or modifying a private copy.  Propagation includes copying,
+distribution (with or without modification), making available to the
+public, and in some countries other activities as well.
+
+  To "convey" a work means any kind of propagation that enables other
+parties to make or receive copies.  Mere interaction with a user through
+a computer network, with no transfer of a copy, is not conveying.
+
+  An interactive user interface displays "Appropriate Legal Notices"
+to the extent that it includes a convenient and prominently visible
+feature that (1) displays an appropriate copyright notice, and (2)
+tells the user that there is no warranty for the work (except to the
+extent that warranties are provided), that licensees may convey the
+work under this License, and how to view a copy of this License.  If
+the interface presents a list of user commands or options, such as a
+menu, a prominent item in the list meets this criterion.
+
+  1. Source Code.
+
+  The "source code" for a work means the preferred form of the work
+for making modifications to it.  "Object code" means any non-source
+form of a work.
+
+  A "Standard Interface" means an interface that either is an official
+standard defined by a recognized standards body, or, in the case of
+interfaces specified for a particular programming language, one that
+is widely used among developers working in that language.
+
+  The "System Libraries" of an executable work include anything, other
+than the work as a whole, that (a) is included in the normal form of
+packaging a Major Component, but which is not part of that Major
+Component, and (b) serves only to enable use of the work with that
+Major Component, or to implement a Standard Interface for which an
+implementation is available to the public in source code form.  A
+"Major Component", in this context, means a major essential component
+(kernel, window system, and so on) of the specific operating system
+(if any) on which the executable work runs, or a compiler used to
+produce the work, or an object code interpreter used to run it.
+
+  The "Corresponding Source" for a work in object code form means all
+the source code needed to generate, install, and (for an executable
+work) run the object code and to modify the work, including scripts to
+control those activities.  However, it does not include the work's
+System Libraries, or general-purpose tools or generally available free
+programs which are used unmodified in performing those activities but
+which are not part of the work.  For example, Corresponding Source
+includes interface definition files associated with source files for
+the work, and the source code for shared libraries and dynamically
+linked subprograms that the work is specifically designed to require,
+such as by intimate data communication or control flow between those
+subprograms and other parts of the work.
+
+  The Corresponding Source need not include anything that users
+can regenerate automatically from other parts of the Corresponding
+Source.
+
+  The Corresponding Source for a work in source code form is that
+same work.
+
+  2. Basic Permissions.
+
+  All rights granted under this License are granted for the term of
+copyright on the Program, and are irrevocable provided the stated
+conditions are met.  This License explicitly affirms your unlimited
+permission to run the unmodified Program.  The output from running a
+covered work is covered by this License only if the output, given its
+content, constitutes a covered work.  This License acknowledges your
+rights of fair use or other equivalent, as provided by copyright law.
+
+  You may make, run and propagate covered works that you do not
+convey, without conditions so long as your license otherwise remains
+in force.  You may convey covered works to others for the sole purpose
+of having them make modifications exclusively for you, or provide you
+with facilities for running those works, provided that you comply with
+the terms of this License in conveying all material for which you do
+not control copyright.  Those thus making or running the covered works
+for you must do so exclusively on your behalf, under your direction
+and control, on terms that prohibit them from making any copies of
+your copyrighted material outside their relationship with you.
+
+  Conveying under any other circumstances is permitted solely under
+the conditions stated below.  Sublicensing is not allowed; section 10
+makes it unnecessary.
+
+  3. Protecting Users' Legal Rights From Anti-Circumvention Law.
+
+  No covered work shall be deemed part of an effective technological
+measure under any applicable law fulfilling obligations under article
+11 of the WIPO copyright treaty adopted on 20 December 1996, or
+similar laws prohibiting or restricting circumvention of such
+measures.
+
+  When you convey a covered work, you waive any legal power to forbid
+circumvention of technological measures to the extent such circumvention
+is effected by exercising rights under this License with respect to
+the covered work, and you disclaim any intention to limit operation or
+modification of the work as a means of enforcing, against the work's
+users, your or third parties' legal rights to forbid circumvention of
+technological measures.
+
+  4. Conveying Verbatim Copies.
+
+  You may convey verbatim copies of the Program's source code as you
+receive it, in any medium, provided that you conspicuously and
+appropriately publish on each copy an appropriate copyright notice;
+keep intact all notices stating that this License and any
+non-permissive terms added in accord with section 7 apply to the code;
+keep intact all notices of the absence of any warranty; and give all
+recipients a copy of this License along with the Program.
+
+  You may charge any price or no price for each copy that you convey,
+and you may offer support or warranty protection for a fee.
+
+  5. Conveying Modified Source Versions.
+
+  You may convey a work based on the Program, or the modifications to
+produce it from the Program, in the form of source code under the
+terms of section 4, provided that you also meet all of these conditions:
+
+    a) The work must carry prominent notices stating that you modified
+    it, and giving a relevant date.
+
+    b) The work must carry prominent notices stating that it is
+    released under this License and any conditions added under section
+    7.  This requirement modifies the requirement in section 4 to
+    "keep intact all notices".
+
+    c) You must license the entire work, as a whole, under this
+    License to anyone who comes into possession of a copy.  This
+    License will therefore apply, along with any applicable section 7
+    additional terms, to the whole of the work, and all its parts,
+    regardless of how they are packaged.  This License gives no
+    permission to license the work in any other way, but it does not
+    invalidate such permission if you have separately received it.
+
+    d) If the work has interactive user interfaces, each must display
+    Appropriate Legal Notices; however, if the Program has interactive
+    interfaces that do not display Appropriate Legal Notices, your
+    work need not make them do so.
+
+  A compilation of a covered work with other separate and independent
+works, which are not by their nature extensions of the covered work,
+and which are not combined with it such as to form a larger program,
+in or on a volume of a storage or distribution medium, is called an
+"aggregate" if the compilation and its resulting copyright are not
+used to limit the access or legal rights of the compilation's users
+beyond what the individual works permit.  Inclusion of a covered work
+in an aggregate does not cause this License to apply to the other
+parts of the aggregate.
+
+  6. Conveying Non-Source Forms.
+
+  You may convey a covered work in object code form under the terms
+of sections 4 and 5, provided that you also convey the
+machine-readable Corresponding Source under the terms of this License,
+in one of these ways:
+
+    a) Convey the object code in, or embodied in, a physical product
+    (including a physical distribution medium), accompanied by the
+    Corresponding Source fixed on a durable physical medium
+    customarily used for software interchange.
+
+    b) Convey the object code in, or embodied in, a physical product
+    (including a physical distribution medium), accompanied by a
+    written offer, valid for at least three years and valid for as
+    long as you offer spare parts or customer support for that product
+    model, to give anyone who possesses the object code either (1) a
+    copy of the Corresponding Source for all the software in the
+    product that is covered by this License, on a durable physical
+    medium customarily used for software interchange, for a price no
+    more than your reasonable cost of physically performing this
+    conveying of source, or (2) access to copy the
+    Corresponding Source from a network server at no charge.
+
+    c) Convey individual copies of the object code with a copy of the
+    written offer to provide the Corresponding Source.  This
+    alternative is allowed only occasionally and noncommercially, and
+    only if you received the object code with such an offer, in accord
+    with subsection 6b.
+
+    d) Convey the object code by offering access from a designated
+    place (gratis or for a charge), and offer equivalent access to the
+    Corresponding Source in the same way through the same place at no
+    further charge.  You need not require recipients to copy the
+    Corresponding Source along with the object code.  If the place to
+    copy the object code is a network server, the Corresponding Source
+    may be on a different server (operated by you or a third party)
+    that supports equivalent copying facilities, provided you maintain
+    clear directions next to the object code saying where to find the
+    Corresponding Source.  Regardless of what server hosts the
+    Corresponding Source, you remain obligated to ensure that it is
+    available for as long as needed to satisfy these requirements.
+
+    e) Convey the object code using peer-to-peer transmission, provided
+    you inform other peers where the object code and Corresponding
+    Source of the work are being offered to the general public at no
+    charge under subsection 6d.
+
+  A separable portion of the object code, whose source code is excluded
+from the Corresponding Source as a System Library, need not be
+included in conveying the object code work.
+
+  A "User Product" is either (1) a "consumer product", which means any
+tangible personal property which is normally used for personal, family,
+or household purposes, or (2) anything designed or sold for incorporation
+into a dwelling.  In determining whether a product is a consumer product,
+doubtful cases shall be resolved in favor of coverage.  For a particular
+product received by a particular user, "normally used" refers to a
+typical or common use of that class of product, regardless of the status
+of the particular user or of the way in which the particular user
+actually uses, or expects or is expected to use, the product.  A product
+is a consumer product regardless of whether the product has substantial
+commercial, industrial or non-consumer uses, unless such uses represent
+the only significant mode of use of the product.
+
+  "Installation Information" for a User Product means any methods,
+procedures, authorization keys, or other information required to install
+and execute modified versions of a covered work in that User Product from
+a modified version of its Corresponding Source.  The information must
+suffice to ensure that the continued functioning of the modified object
+code is in no case prevented or interfered with solely because
+modification has been made.
+
+  If you convey an object code work under this section in, or with, or
+specifically for use in, a User Product, and the conveying occurs as
+part of a transaction in which the right of possession and use of the
+User Product is transferred to the recipient in perpetuity or for a
+fixed term (regardless of how the transaction is characterized), the
+Corresponding Source conveyed under this section must be accompanied
+by the Installation Information.  But this requirement does not apply
+if neither you nor any third party retains the ability to install
+modified object code on the User Product (for example, the work has
+been installed in ROM).
+
+  The requirement to provide Installation Information does not include a
+requirement to continue to provide support service, warranty, or updates
+for a work that has been modified or installed by the recipient, or for
+the User Product in which it has been modified or installed.  Access to a
+network may be denied when the modification itself materially and
+adversely affects the operation of the network or violates the rules and
+protocols for communication across the network.
+
+  Corresponding Source conveyed, and Installation Information provided,
+in accord with this section must be in a format that is publicly
+documented (and with an implementation available to the public in
+source code form), and must require no special password or key for
+unpacking, reading or copying.
+
+  7. Additional Terms.
+
+  "Additional permissions" are terms that supplement the terms of this
+License by making exceptions from one or more of its conditions.
+Additional permissions that are applicable to the entire Program shall
+be treated as though they were included in this License, to the extent
+that they are valid under applicable law.  If additional permissions
+apply only to part of the Program, that part may be used separately
+under those permissions, but the entire Program remains governed by
+this License without regard to the additional permissions.
+
+  When you convey a copy of a covered work, you may at your option
+remove any additional permissions from that copy, or from any part of
+it.  (Additional permissions may be written to require their own
+removal in certain cases when you modify the work.)  You may place
+additional permissions on material, added by you to a covered work,
+for which you have or can give appropriate copyright permission.
+
+  Notwithstanding any other provision of this License, for material you
+add to a covered work, you may (if authorized by the copyright holders of
+that material) supplement the terms of this License with terms:
+
+    a) Disclaiming warranty or limiting liability differently from the
+    terms of sections 15 and 16 of this License; or
+
+    b) Requiring preservation of specified reasonable legal notices or
+    author attributions in that material or in the Appropriate Legal
+    Notices displayed by works containing it; or
+
+    c) Prohibiting misrepresentation of the origin of that material, or
+    requiring that modified versions of such material be marked in
+    reasonable ways as different from the original version; or
+
+    d) Limiting the use for publicity purposes of names of licensors or
+    authors of the material; or
+
+    e) Declining to grant rights under trademark law for use of some
+    trade names, trademarks, or service marks; or
+
+    f) Requiring indemnification of licensors and authors of that
+    material by anyone who conveys the material (or modified versions of
+    it) with contractual assumptions of liability to the recipient, for
+    any liability that these contractual assumptions directly impose on
+    those licensors and authors.
+
+  All other non-permissive additional terms are considered "further
+restrictions" within the meaning of section 10.  If the Program as you
+received it, or any part of it, contains a notice stating that it is
+governed by this License along with a term that is a further
+restriction, you may remove that term.  If a license document contains
+a further restriction but permits relicensing or conveying under this
+License, you may add to a covered work material governed by the terms
+of that license document, provided that the further restriction does
+not survive such relicensing or conveying.
+
+  If you add terms to a covered work in accord with this section, you
+must place, in the relevant source files, a statement of the
+additional terms that apply to those files, or a notice indicating
+where to find the applicable terms.
+
+  Additional terms, permissive or non-permissive, may be stated in the
+form of a separately written license, or stated as exceptions;
+the above requirements apply either way.
+
+  8. Termination.
+
+  You may not propagate or modify a covered work except as expressly
+provided under this License.  Any attempt otherwise to propagate or
+modify it is void, and will automatically terminate your rights under
+this License (including any patent licenses granted under the third
+paragraph of section 11).
+
+  However, if you cease all violation of this License, then your
+license from a particular copyright holder is reinstated (a)
+provisionally, unless and until the copyright holder explicitly and
+finally terminates your license, and (b) permanently, if the copyright
+holder fails to notify you of the violation by some reasonable means
+prior to 60 days after the cessation.
+
+  Moreover, your license from a particular copyright holder is
+reinstated permanently if the copyright holder notifies you of the
+violation by some reasonable means, this is the first time you have
+received notice of violation of this License (for any work) from that
+copyright holder, and you cure the violation prior to 30 days after
+your receipt of the notice.
+
+  Termination of your rights under this section does not terminate the
+licenses of parties who have received copies or rights from you under
+this License.  If your rights have been terminated and not permanently
+reinstated, you do not qualify to receive new licenses for the same
+material under section 10.
+
+  9. Acceptance Not Required for Having Copies.
+
+  You are not required to accept this License in order to receive or
+run a copy of the Program.  Ancillary propagation of a covered work
+occurring solely as a consequence of using peer-to-peer transmission
+to receive a copy likewise does not require acceptance.  However,
+nothing other than this License grants you permission to propagate or
+modify any covered work.  These actions infringe copyright if you do
+not accept this License.  Therefore, by modifying or propagating a
+covered work, you indicate your acceptance of this License to do so.
+
+  10. Automatic Licensing of Downstream Recipients.
+
+  Each time you convey a covered work, the recipient automatically
+receives a license from the original licensors, to run, modify and
+propagate that work, subject to this License.  You are not responsible
+for enforcing compliance by third parties with this License.
+
+  An "entity transaction" is a transaction transferring control of an
+organization, or substantially all assets of one, or subdividing an
+organization, or merging organizations.  If propagation of a covered
+work results from an entity transaction, each party to that
+transaction who receives a copy of the work also receives whatever
+licenses to the work the party's predecessor in interest had or could
+give under the previous paragraph, plus a right to possession of the
+Corresponding Source of the work from the predecessor in interest, if
+the predecessor has it or can get it with reasonable efforts.
+
+  You may not impose any further restrictions on the exercise of the
+rights granted or affirmed under this License.  For example, you may
+not impose a license fee, royalty, or other charge for exercise of
+rights granted under this License, and you may not initiate litigation
+(including a cross-claim or counterclaim in a lawsuit) alleging that
+any patent claim is infringed by making, using, selling, offering for
+sale, or importing the Program or any portion of it.
+
+  11. Patents.
+
+  A "contributor" is a copyright holder who authorizes use under this
+License of the Program or a work on which the Program is based.  The
+work thus licensed is called the contributor's "contributor version".
+
+  A contributor's "essential patent claims" are all patent claims
+owned or controlled by the contributor, whether already acquired or
+hereafter acquired, that would be infringed by some manner, permitted
+by this License, of making, using, or selling its contributor version,
+but do not include claims that would be infringed only as a
+consequence of further modification of the contributor version.  For
+purposes of this definition, "control" includes the right to grant
+patent sublicenses in a manner consistent with the requirements of
+this License.
+
+  Each contributor grants you a non-exclusive, worldwide, royalty-free
+patent license under the contributor's essential patent claims, to
+make, use, sell, offer for sale, import and otherwise run, modify and
+propagate the contents of its contributor version.
+
+  In the following three paragraphs, a "patent license" is any express
+agreement or commitment, however denominated, not to enforce a patent
+(such as an express permission to practice a patent or covenant not to
+sue for patent infringement).  To "grant" such a patent license to a
+party means to make such an agreement or commitment not to enforce a
+patent against the party.
+
+  If you convey a covered work, knowingly relying on a patent license,
+and the Corresponding Source of the work is not available for anyone
+to copy, free of charge and under the terms of this License, through a
+publicly available network server or other readily accessible means,
+then you must either (1) cause the Corresponding Source to be so
+available, or (2) arrange to deprive yourself of the benefit of the
+patent license for this particular work, or (3) arrange, in a manner
+consistent with the requirements of this License, to extend the patent
+license to downstream recipients.  "Knowingly relying" means you have
+actual knowledge that, but for the patent license, your conveying the
+covered work in a country, or your recipient's use of the covered work
+in a country, would infringe one or more identifiable patents in that
+country that you have reason to believe are valid.
+
+  If, pursuant to or in connection with a single transaction or
+arrangement, you convey, or propagate by procuring conveyance of, a
+covered work, and grant a patent license to some of the parties
+receiving the covered work authorizing them to use, propagate, modify
+or convey a specific copy of the covered work, then the patent license
+you grant is automatically extended to all recipients of the covered
+work and works based on it.
+
+  A patent license is "discriminatory" if it does not include within
+the scope of its coverage, prohibits the exercise of, or is
+conditioned on the non-exercise of one or more of the rights that are
+specifically granted under this License.  You may not convey a covered
+work if you are a party to an arrangement with a third party that is
+in the business of distributing software, under which you make payment
+to the third party based on the extent of your activity of conveying
+the work, and under which the third party grants, to any of the
+parties who would receive the covered work from you, a discriminatory
+patent license (a) in connection with copies of the covered work
+conveyed by you (or copies made from those copies), or (b) primarily
+for and in connection with specific products or compilations that
+contain the covered work, unless you entered into that arrangement,
+or that patent license was granted, prior to 28 March 2007.
+
+  Nothing in this License shall be construed as excluding or limiting
+any implied license or other defenses to infringement that may
+otherwise be available to you under applicable patent law.
+
+  12. No Surrender of Others' Freedom.
+
+  If conditions are imposed on you (whether by court order, agreement or
+otherwise) that contradict the conditions of this License, they do not
+excuse you from the conditions of this License.  If you cannot convey a
+covered work so as to satisfy simultaneously your obligations under this
+License and any other pertinent obligations, then as a consequence you may
+not convey it at all.  For example, if you agree to terms that obligate you
+to collect a royalty for further conveying from those to whom you convey
+the Program, the only way you could satisfy both those terms and this
+License would be to refrain entirely from conveying the Program.
+
+  13. Use with the GNU Affero General Public License.
+
+  Notwithstanding any other provision of this License, you have
+permission to link or combine any covered work with a work licensed
+under version 3 of the GNU Affero General Public License into a single
+combined work, and to convey the resulting work.  The terms of this
+License will continue to apply to the part which is the covered work,
+but the special requirements of the GNU Affero General Public License,
+section 13, concerning interaction through a network will apply to the
+combination as such.
+
+  14. Revised Versions of this License.
+
+  The Free Software Foundation may publish revised and/or new versions of
+the GNU General Public License from time to time.  Such new versions will
+be similar in spirit to the present version, but may differ in detail to
+address new problems or concerns.
+
+  Each version is given a distinguishing version number.  If the
+Program specifies that a certain numbered version of the GNU General
+Public License "or any later version" applies to it, you have the
+option of following the terms and conditions either of that numbered
+version or of any later version published by the Free Software
+Foundation.  If the Program does not specify a version number of the
+GNU General Public License, you may choose any version ever published
+by the Free Software Foundation.
+
+  If the Program specifies that a proxy can decide which future
+versions of the GNU General Public License can be used, that proxy's
+public statement of acceptance of a version permanently authorizes you
+to choose that version for the Program.
+
+  Later license versions may give you additional or different
+permissions.  However, no additional obligations are imposed on any
+author or copyright holder as a result of your choosing to follow a
+later version.
+
+  15. Disclaimer of Warranty.
+
+  THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
+APPLICABLE LAW.  EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
+HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
+OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
+THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+PURPOSE.  THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
+IS WITH YOU.  SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
+ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
+
+  16. Limitation of Liability.
+
+  IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
+WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
+THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
+GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
+USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
+DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
+PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
+EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
+SUCH DAMAGES.
+
+  17. Interpretation of Sections 15 and 16.
+
+  If the disclaimer of warranty and limitation of liability provided
+above cannot be given local legal effect according to their terms,
+reviewing courts shall apply local law that most closely approximates
+an absolute waiver of all civil liability in connection with the
+Program, unless a warranty or assumption of liability accompanies a
+copy of the Program in return for a fee.
+
+                     END OF TERMS AND CONDITIONS
+
+            How to Apply These Terms to Your New Programs
+
+  If you develop a new program, and you want it to be of the greatest
+possible use to the public, the best way to achieve this is to make it
+free software which everyone can redistribute and change under these terms.
+
+  To do so, attach the following notices to the program.  It is safest
+to attach them to the start of each source file to most effectively
+state the exclusion of warranty; and each file should have at least
+the "copyright" line and a pointer to where the full notice is found.
+
+    <one line to give the program's name and a brief idea of what it does.>
+    Copyright (C) <year>  <name of author>
+
+    This program is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation, either version 3 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+Also add information on how to contact you by electronic and paper mail.
+
+  If the program does terminal interaction, make it output a short
+notice like this when it starts in an interactive mode:
+
+    <program>  Copyright (C) <year>  <name of author>
+    This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
+    This is free software, and you are welcome to redistribute it
+    under certain conditions; type `show c' for details.
+
+The hypothetical commands `show w' and `show c' should show the appropriate
+parts of the General Public License.  Of course, your program's commands
+might be different; for a GUI interface, you would use an "about box".
+
+  You should also get your employer (if you work as a programmer) or school,
+if any, to sign a "copyright disclaimer" for the program, if necessary.
+For more information on this, and how to apply and follow the GNU GPL, see
+<https://www.gnu.org/licenses/>.
+
+  The GNU General Public License does not permit incorporating your program
+into proprietary programs.  If your program is a subroutine library, you
+may consider it more useful to permit linking proprietary applications with
+the library.  If this is what you want to do, use the GNU Lesser General
+Public License instead of this License.  But first, please read
+<https://www.gnu.org/licenses/why-not-lgpl.html>.

README.md

@@ -0,0 +1,312 @@
+# QuickSearch
+
+A fast local file indexer and search tool. QuickSearch walks your chosen
+folders into a compact SQLite index (FTS5 full-text + zstd-compressed text
+sidecar), keeps it fresh automatically with filesystem watchers and
+periodic reindexing, and serves ranked search-as-you-type results in a
+compact egui desktop app, or straight to your terminal.
+
+## Build & run
+
+Requirements: a Rust toolchain (edition 2021). The old WebKit/WebView
+dependencies (`setup.sh`) are gone; the GUI renders with OpenGL via egui.
+
+SQLite and zstd are compiled from bundled C sources, so a C toolchain is
+required on every platform:
+
+- Linux: working OpenGL 3.3 drivers; `xdg-desktop-portal` (present on all
+  mainstream desktops) provides the native folder picker. On minimal
+  images you may need `build-essential pkg-config libxkbcommon-dev`.
+- Windows: Visual Studio 2022 Build Tools with the "Desktop development
+  with C++" workload (MSVC v143 plus a Windows SDK). For the GNU target
+  instead, `rustup target add x86_64-pc-windows-gnu` and a mingw-w64
+  toolchain. Note that Windows ships only a software OpenGL 1.1 driver, so
+  a bare VM or an RDP session without a vendor GPU driver cannot create a
+  context and the window will fail to open.
+- macOS: Xcode command line tools.
+
+```sh
+cargo build --release -p quicksearch-gui   # binaries: target/release/quicksearch{,-cli}
+cargo run -p quicksearch-gui               # or just run it
+cargo test -p quicksearch-core             # backend test suite
+```
+
+`run.sh` / `run.bat` wrap the same commands.
+
+Two binaries are produced. `quicksearch` is the desktop app; on Windows it
+is built as a window-subsystem app so no console appears behind it.
+`quicksearch-cli` is terminal search — a console app, so pipes, redirection
+and exit codes behave normally. On Unix `quicksearch` also does both, and
+`quicksearch-cli` is simply the same tool under a clearer name.
+
+## Install (Debian / Ubuntu)
+
+```sh
+./packaging/build-deb.sh
+sudo apt install ./dist/quicksearch_0.1.0-1_amd64.deb
+```
+
+The script builds the release binary, strips it, and assembles a `.deb` with
+`dpkg-deb`. It needs no `cargo-deb`, no `debhelper` and no SVG rasteriser —
+only `dpkg-deb` and `desktop-file-utils`, both standard on Debian and Ubuntu.
+Useful flags: `--no-build` to package a binary you already built, `--no-strip`
+to keep debug symbols, `-o DIR` to write elsewhere. `DEB_REVISION` and
+`DEB_MAINTAINER` override the packaging revision and maintainer.
+
+The package installs:
+
+| Path | Contents |
+| --- | --- |
+| `/usr/bin/quicksearch` | the desktop app, which also does terminal search |
+| `/usr/bin/quicksearch-cli` | terminal search only |
+| `/usr/share/applications/quicksearch.desktop` | menu entry, so QuickSearch appears in the app launcher |
+| `/usr/share/icons/hicolor/{16,22,24,32,48,64,128,256}x*/apps/` | icons at each size |
+| `/usr/share/icons/hicolor/scalable/apps/quicksearch.svg` | the source icon |
+| `/usr/share/man/man1/quicksearch{,-cli}.1.gz` | `man quicksearch`; the `-cli` page is a `.so` stub pointing at it |
+| `/usr/share/doc/quicksearch/` | copyright, changelog, README, `config_example.toml` |
+
+Installing registers the menu entry and the icon: dpkg triggers owned by
+`desktop-file-utils` and `hicolor-icon-theme` refresh both caches, so no
+maintainer scripts are involved and `apt remove` reverses it cleanly.
+
+No `config.toml` is installed. One placed next to the executable would put
+every user into portable mode (see [Configuration](#configuration)); instead
+the app writes `~/.config/quicksearch/config.toml` on first run.
+
+### Icons
+
+`crates/quicksearch-gui/assets/icons/` holds `quicksearch_icon.svg` and the
+PNGs rasterised from it. The PNGs are committed rather than generated, so an
+ordinary `cargo build` needs no image tooling — the 256px one is compiled into
+the binary with `include_bytes!` and becomes the window icon. Editing the SVG
+means re-rendering the PNGs; `packaging/build-deb.sh` documents how in a
+comment at the top.
+
+X11 takes the window icon from the embedded PNG. Wayland ignores it and
+matches the app id (`quicksearch`) against the installed
+`quicksearch.desktop`, so under Wayland the titlebar icon appears only once
+the package is installed.
+
+## Usage
+
+### GUI
+
+`quicksearch` with no query arguments opens the app:
+
+- **Search**: results appear as you type; every keystroke cancels the
+  previous search. One checkbox enables the two fuzzy passes. Sort by
+  rank, name, path, size, or modified; right-click a result to open it,
+  reveal it in the file manager, or build an ignore filter from it
+  (session-only by default, optionally persisted to the config). Matches
+  in file contents show highlighted snippets.
+- **Manage Index**: full indexing status, Start/Stop/Automatic controls,
+  indexed folder list, full-text extension filters, ignore patterns, and
+  the indexing options.
+- **Duplicates**: files sharing a content hash, grouped.
+- **Logs**: the lines the app would have printed to a terminal — warnings
+  from indexing, folder watching and opening files, newest last, with a
+  filter box and Copy button. Launched from a desktop launcher (or on
+  Windows, where the app has no console at all) this is the only place
+  they are visible.
+
+The bottom status bar always shows what the indexer is doing (phase,
+percent, files/sec) or the total indexed file count when idle.
+
+### Terminal
+
+```sh
+quicksearch report type:Document modified:">=2024-01-01"
+quicksearch --long --limit 20 "quarterly budget"
+quicksearch --fuzzy repot          # tolerates typos
+```
+
+Prints rank-ordered paths (pipe-friendly); `--long` adds rank, size,
+mtime, and highlighted snippets. `quicksearch --help` shows all flags.
+
+On Windows use `quicksearch-cli` for all of the above — `quicksearch.exe`
+opens the app, and any query given to it seeds the search box instead of
+printing. Colour in `--long` output needs a console with virtual-terminal
+processing; Windows Terminal has it, and older consoles get plain text.
+
+### Query syntax
+
+This section is the complete reference (the in-app "?" popup shows a
+condensed version of the same rules). Everything that isn't a filter is
+matched as one phrase, in order. Filters combine freely with the search
+text:
+
+| Syntax | Meaning |
+|---|---|
+| `budget report` | names, contents, and paths containing the phrase `budget report` |
+| `"exact phrase"` | quotes keep spaces, stars, and filter-like words literal; `""` escapes a quote |
+| `bud*port` | `*` matches any run of characters (it stays on one line of content); `%` and `_` are always literal |
+| `regex:"(foo|bar)\d+"` | regular expression matched against names, contents, and paths; case-insensitive by default (`(?-i:…)` overrides); quote patterns containing spaces or `( ) : = < > "` |
+| `type:Audio` | one of Audio, Image, Video, Document, Text, Archive, Spreadsheet, Presentation, Folder |
+| `modified:>=2024-01-01` | also `<`, `<=`, `>`, `=` (dates are `yyyy-mm-dd`; `mtime:` is an alias) |
+| `path:/home/me/docs` | restrict to a folder and its subfolders (`folder:` and `includefolder:` are aliases); `*` is literal here |
+| `path:C:\Users\me\docs` | the same on Windows — drive letters and backslashes need no quoting |
+| `mime:application/pdf` | exact MIME type |
+| `name:re*.txt` | filename contains (as a filter, unranked; `filename:` is an alias); unquoted `*` globs |
+
+Unrecognized `key:value` text (like `12:30`) stays part of the search
+phrase, and a half-typed quote never errors while you type. `AND`, `OR`
+and parentheses are treated as plain words. A term of only stars matches
+nothing, and a regex that could match the empty string is rejected rather
+than matching every file. `regex:` bypasses the trigram index entirely
+and combines with filters; alongside search text it acts as an extra
+requirement on those results.
+
+The search box highlights this syntax as you type: recognized filter
+keywords in red, their arguments in blue, syntax characters (operators,
+quotes, live wildcards) in green, on a tinted chip per complete filter.
+An argument the engine would reject — unknown `type:` name, bad date,
+invalid regex — switches to the error color immediately.
+
+Results are ranked: exact filename matches (case-sensitive first), then
+filename substrings, then full-text matches ordered by occurrence count,
+then fuzzy filename/full-text matches when enabled, and last the files
+matched somewhere else in their path. Later, weaker matches only ever
+append to the bottom of the list. Wildcard terms rank through the same
+tiers (an "exact" match means the whole name matches the pattern) but
+skip the fuzzy passes; regex-only queries reuse the substring, full-text,
+and path tiers. Path matching needs at least three characters, and terms
+may span separators (`docs/report`). Full-text matching also needs at
+least three characters of literal text (the trigram floor). The fuzzy
+passes tolerate typos with a budget of one edit per three characters,
+capped by `[search] fuzzy_max_edits` (default 2; 0 turns fuzzy off).
+
+### Configuration
+
+`config.toml` lives at `~/.config/quicksearch/config.toml` (Windows:
+`%APPDATA%\quicksearch\config.toml`) and is created on first run; the
+default index goes to `~/.local/share/quicksearch/index.sqlite`
+(Windows: `%LOCALAPPDATA%\quicksearch\index.sqlite`). See
+`config_example.toml` for every option.
+
+Defaults follow the platform. The first indexing root is your home
+directory or `%USERPROFILE%`; `include_hidden = false` skips dot-files
+everywhere and additionally anything marked Hidden or System on Windows,
+which is what keeps `AppData`, `$RECYCLE.BIN` and `System Volume
+Information` out of the index; and ignore patterns are matched
+case-insensitively on Windows and macOS, matching the filesystem.
+
+**Portable mode**: a `config.toml` sitting next to the `quicksearch`
+binary overrides the user config entirely, and relative paths inside any
+config resolve against the config file's own directory, so a folder
+containing the binary, its config, and its index can be moved wholesale.
+
+The GUI edits the config live; external edits apply on next start.
+
+## Engineering overview
+
+Two crates:
+
+```
+crates/quicksearch-core   library: indexing, storage, search
+crates/quicksearch-gui    binary "quicksearch": egui app + terminal mode
+```
+
+### Backend (`quicksearch-core`)
+
+Synchronous Rust: `std::thread` + `mpsc` channels, no async runtime.
+
+- **Storage** (`db/`): SQLite via rusqlite (bundled), WAL mode so the
+  single writer never blocks streaming read-only searches. `files` holds
+  metadata (name, path, size, mtime, hash, MIME/type bitmask, per-row
+  index state); `searchabletext` is a *contentless* FTS5 table (postings
+  only, configurable tokenizer, trigram by default); canonical extracted
+  text lives zstd-compressed in `documents_text`, which powers snippets,
+  occurrence ranking, and fuzzy full-text search. Schema changes wipe and
+  rebuild by policy; the indexer (`open_or_recreate`) is the only code
+  allowed to do that; every consumer uses `open_existing`, which treats
+  drift as an error, never data loss.
+- **Indexing** (`indexing.rs`, `file_handling.rs`): full runs walk each
+  root (`filtered_walk` prunes hidden/ignored subtrees before descending),
+  classify files by mtime into insert/update/skip, batch-write metadata,
+  sweep stale rows, then extract content (plaintext, Office, PDF, audio
+  tags, EXIF; see `extract/`) for FTS. Files no larger than
+  `processing.hash_length` skip that second pass entirely: the head the walk
+  reads to hash them is already their whole content, so a plaintext body is
+  extracted in the same `read` and stored complete. Progress streams through
+  a polled `IndexingStatus`.
+- **Coordinator** (`coordinator.rs`): the object binaries construct.
+  Owns the `IndexingService`, the debouncing filesystem watcher
+  (`watcher.rs`), and the mode state machine (Auto / Manual). Watcher
+  events become single-file transactions (`incremental.rs`) that keep
+  `files`, FTS, and the text sidecar consistent per commit; a full
+  reindex runs on a configurable interval. Incremental writes defer while
+  a full run is active, so there is exactly one writer at a time.
+  Registration follows what the platform's notification API can do:
+  inotify covers one directory per watch, so the roots are walked and each
+  surviving directory registered individually (skipping `.git`,
+  `node_modules` and hidden subtrees, which is what keeps the watch count
+  affordable), while `ReadDirectoryChangesW` covers a whole tree from one
+  handle and takes a single watch per root, filtering the events instead.
+  Either way a tree too large to watch degrades to periodic reindexing
+  rather than going silently stale.
+- **Search** (`search/`): `SearchService` runs one worker thread; each
+  query is a *generation*. New queries interrupt the in-flight SQLite
+  statement (`InterruptHandle`) and stale generations stop cooperatively,
+  so typing never waits. The cascade streams rank-ordered batches: one
+  `files` scan classifies exact/case/substring filename matches (ranks
+  1–4) and, since a path contains its own name, sets aside full-path
+  matches from the same rows (ranks 9–10); one FTS phrase probe verified
+  against the decompressed text yields full-text ranks 5–6 ordered by
+  occurrence count; and the opt-in fuzzy passes run a bitap (Wu–Manber)
+  matcher over filenames (rank 7), document text (rank 8) and paths
+  (rank 11), with a configurable edit budget. The deferred path tiers
+  flush last, so weaker matches only ever append. All SQL is
+  parameterized; structured filters from the query language (`query/`)
+  are ANDed onto every pass.
+- **Baloo compatibility** (`cli.rs`, `mime.rs`): read-only endpoints
+  (`status_for_path`, `list_failed`, `index_size_breakdown`, …) and a
+  Baloo-shaped type model, groundwork for a future `balooctl`-compatible
+  layer.
+- **Logging** (`log.rs`): background reporting goes through `log_info!` /
+  `log_warn!` rather than `println!`/`eprintln!`. Each writes its line to
+  stderr *and* appends it to a bounded in-memory ring (newest 5000 lines,
+  with a count of what was dropped) that the GUI's Logs tab reads, so a
+  windowed run with no terminal still surfaces them. Command output —
+  search hits, usage, the error a command exits with — stays on stdio.
+- **Platform differences** (`platform.rs`): the single home for `#[cfg]`.
+  Home directory lookup, what counts as a hidden entry (dot-prefix, plus
+  the Hidden/System attributes on Windows), network-filesystem detection
+  (`/proc/mounts` against `GetDriveTypeW`), path collation, and the
+  watch-registration strategy all live here, so the rest of the crate can
+  ask a question rather than test a target. Anything decidable from a
+  string alone is split out so its tests run on every platform.
+
+### Frontend (`quicksearch-gui`)
+
+Immediate-mode egui/eframe app, one UI thread:
+
+```
+UI thread ──SearchRequest──▶ search worker ──SearchUpdate (mpsc)──▶ drained per frame
+UI thread ──commands──────▶ IndexCoordinator ──state──▶ polled per frame
+core threads ─────────────▶ ctx.request_repaint() (wake the UI)
+```
+
+Modules map one-to-one onto what you see: `app.rs` (shell, status bar,
+config routing), `search_tab.rs` (query strip, virtualized results table,
+snippet highlighting via `LayoutJob` byte ranges, ignore dialog),
+`manage_tab.rs` (status detail + `tracker.rs` rate estimation, roots and
+filter editors), `duplicates_tab.rs`, `logs_tab.rs` (a virtualized view of
+the core log ring), `options.rs` (draft-based settings
+editor shared between the window and the Manage tab), `platform.rs`
+(open / reveal-in-file-manager, and the Windows stdio setup a
+window-subsystem process needs before anything prints), `cli.rs` (terminal
+mode, shared with the `quicksearch-cli` binary). There is no
+pagination: the table is virtualized, so a single scroll list capped at
+`display_limit` renders in microseconds regardless of row count.
+
+## Development
+
+- `cargo test -p quicksearch-core`: unit + integration suites (cascade
+  ranking, cancellation, incremental indexing, coordinator modes, config
+  resolution, fuzzy matcher vs. brute-force oracle).
+- `cargo test -p quicksearch-gui`: formatter/tracker/CLI-parsing units.
+- `QSB_SNIPPET_PERF=1 cargo test --release -p quicksearch-core --test
+  snippet_perf -- --nocapture`: snippet pipeline benchmark.
+- New extractors: implement `extract::Extractor` and register it in
+  `Registry::default_set()`. New cascade behavior: `search/cascade.rs`
+  documents the rank invariants that keep streamed results append-only.

config_example.toml

@@ -1,37 +1,112 @@
+# QuickSearch configuration reference.
+#
+# The live config is auto-created at ~/.config/quicksearch/config.toml
+# (Windows: %APPDATA%\quicksearch\config.toml). A config.toml placed next
+# to the quicksearch binary overrides it entirely (portable mode).
+# Relative paths resolve against the directory containing the config
+# file, so a portable folder can be moved wholesale.
+#
+# Every key is optional; missing keys take the defaults shown here.
+
 [paths]
 # One or more directory roots to index. Walked in order; duplicate and
-# nested roots are de-duplicated automatically.
+# nested roots are de-duplicated automatically. `~` expands to home.
-indexing_paths = ["/"]
+indexing_paths = ["~"]
-database_path = "QuickSearch.db"
+# SQLite index location. Default: ~/.local/share/quicksearch/index.sqlite
+# On Windows the default is %LOCALAPPDATA%\quicksearch\index.sqlite. Write
+# Windows paths as TOML *literal* strings (single quotes) so the
+# backslashes need no escaping, and keep the index out of a roaming
+# profile — it is far too large to synchronise:
+#   database_path = 'C:\Users\you\AppData\Local\quicksearch\index.sqlite'
+database_path = "~/.local/share/quicksearch/index.sqlite"
+
+[indexing]
+# Automatic mode: filesystem watchers apply changes as they happen and a
+# full reindex runs every reindex_interval_minutes. The watcher catches
+# changes as they happen, so this only needs to be often enough to cover
+# whatever the watcher missed.
+auto_index = true
+reindex_interval_minutes = 1440
+# Follow symbolic links during directory walks.
+follow_symlinks = false
+# Index hidden files and directories. That means dot-files everywhere, and
+# additionally anything carrying the Hidden or System attribute on Windows
+# (AppData, $RECYCLE.BIN, System Volume Information, pagefile.sys ...).
+include_hidden = false
+# Empty = extract text from every supported format. Non-empty = content
+# indexing only for these extensions; other files are still listed for
+# filename search. Entries are case-insensitive, leading dot optional.
+content_extensions = []
+# Excluded from the index entirely. A pattern without a separator matches
+# any single path component (so ".git" prunes whole subtrees); patterns
+# containing one match full paths. Glob syntax (*, ?, [..]). Matching is
+# case-insensitive on Windows and macOS, case-sensitive elsewhere,
+# following the filesystem.
+#
+# The Windows defaults add: "$RECYCLE.BIN", "System Volume Information",
+# "pagefile.sys", "hiberfil.sys", "swapfile.sys", "Thumbs.db",
+# "desktop.ini".
+ignore_patterns = [".git", "node_modules", "*.tmp", ".venv", "venv"]
+
+# Worth adding by hand if you index a whole Windows drive rather than just
+# your profile. Neither is excluded by default, because the default root
+# is your profile and a bare "Windows" pattern would also match a folder
+# of your own with that name:
+#   'C:\Windows'        — system files, nothing you would search for
+#   'C:\Windows\WinSxS' — a hardlink farm that floods the Duplicates tab
+
+# Walker threads per root, keyed by the exact root string from
+# indexing_paths. Absent or 0 = auto (4 on local storage, 16 on network
+# mounts, detected per root). Applies at the start of the next run.
+# root_workers = { "/media/share" = 24 }
 
 [processing]
-# Amount of data in bytes read from start/end of files used to calculate hash
+# Bytes read from the start of each file for its content hash, which is
+# `sha256(size || first hash_length bytes)` and backs duplicate detection.
+# Only the head is read: seeking to the end for a second block costs an
+# extra round trip per file on network shares.
+#
+# Known limitation: files of identical size whose heads match will be
+# reported as duplicates. In practice that means pre-allocated VM disk
+# images: a fixed-size VHD stores its unique footer at the end of the
+# file, and a freshly pre-allocated raw/qcow2/VMDK image is all zeros at
+# the head until it is partitioned.
 hash_length = 8192
-# Maximum text content to index per file (bytes) 
+# Maximum extracted text stored per file (bytes).
 maximum_text_size = 262144
-# Maximum file size to process for text extraction (bytes)
+# Files larger than this skip text extraction entirely (bytes).
 maximum_text_file_size = 2097152
-# Number of files to process in each batch (directory walk / inserts / text extraction batches)
+# Files per batch during walks / inserts / extraction.
 batch_size = 200
-# Files per transaction for incremental UPDATE files + DELETE from searchabletext (FTS); larger = fewer commits, more RAM per chunk
+# Files per transaction for incremental FTS updates.
 fts_update_batch_size = 1000
-# If true, run a fast shell-backed tree count before Phase 1 (enables % progress; Linux uses GNU find -printf '\n' | wc -l when available).
+# FTS5 tokenizer: 'trigram' (substring matching, the default; gets
-# If false, Phase 1 shows file counts without a percentage.
+# remove_diacritics 1 appended), 'unicode61', 'porter', or a full FTS5
-precount_files_for_progress = false
+# option string. See https://www.sqlite.org/fts5.html#tokenizers
-# If true, follow symbolic links during directory walks (indexing only; shell precount unchanged).
-follow_symlinks = false
-# If true, hidden files and directories will be indexed.
-include_hidden = false
-# FTS5 tokenization method (e.g., 'trigram', 'porter', 'unicode61')
-# Look here for more information https://www.sqlite.org/fts5.html#tokenizers
 tokenize = "trigram"
-# If true (default), extracted text is stored zstd-compressed in a sidecar
+# Store extracted text (zstd-compressed) alongside the FTS index. Off:
-# table so the GUI's search results can render snippet previews with the
+# the index shrinks to roughly stock-Baloo size, but search loses snippet
-# query terms highlighted. If false, the inverted FTS5 index is still
+# previews, occurrence ranking, case verification, and fuzzy full-text.
-# populated (so queries match the same files) but nothing is stored
-# alongside; result rows carry no snippet. Turning this off drops the
-# on-disk footprint to roughly what stock Baloo uses, useful for
-# apples-to-apples comparisons and for users who never read snippet text.
-# Changing this only affects files indexed *after* the change; existing
-# sidecar rows are kept until re-indexed.
 store_text_for_snippets = true
+
+[ui]
+# Zoom factor for the whole GUI: fonts, spacing, and widgets scale
+# together (0.5 – 2.5). Ctrl +/- and Ctrl 0 adjust it temporarily at
+# runtime; this value is the persistent baseline.
+scale = 1.1
+
+[search]
+# Start with the fuzzy passes enabled.
+fuzzy_default = false
+# Ceiling on the fuzzy stages' typo budget. The allowance grows with the
+# search term, one edit per three characters, up to this value, so 2
+# means "1 edit for 3-5 character terms, 2 for anything longer". 0 turns
+# the fuzzy stages off. Above 3 is allowed but not recommended: matches
+# become dominated by coincidence and every fuzzy pass slows down.
+fuzzy_max_edits = 2
+# Hard cap on results per search (the GUI's scroll list length).
+display_limit = 1000
+# Results per streamed batch (latency/overhead knob, not a page size).
+results_per_page = 100
+# How long the GUI waits after the last keystroke before searching (ms).
+debounce_ms = 150

crates/quicksearch-core/Cargo.toml

@@ -2,6 +2,10 @@
 name = "quicksearch-core"
 version.workspace = true
 edition.workspace = true
+license.workspace = true
+authors.workspace = true
+repository.workspace = true
+description = "Indexing, storage and search engine behind QuickSearch."
 
 [lib]
 name = "quicksearch_core"
@@ -17,10 +21,25 @@ serde = { version = "1.0", features = ["derive"] }
 toml = "0.8"
 mime_guess = "2.0"
 infer = "0.15"
-pdf-extract = "0.7"
+pdf-extract = "0.12"
 lopdf = "0.32"
 lofty = "0.19"
 kamadak-exif = "0.5"
 notify = "6.1"
 ctrlc = "3.4"
 zstd = "0.13"
+globset = "0.4"
+regex = "1"
+
+# `GetDriveTypeW` (a mapped drive letter is the only way to spot an SMB share
+# that isn't written as UNC) plus the FILE_ATTRIBUTE_* constants for hidden
+# detection. Pinned to 0.52 deliberately: walkdir → winapi-util already
+# resolves exactly that version, so this adds no new crate compilations.
+[target.'cfg(windows)'.dependencies]
+# GetDriveTypeW and the FILE_ATTRIBUTE_* constants live in
+# Win32_Storage_FileSystem; DRIVE_REMOTE, oddly, is filed under
+# Win32_System_WindowsProgramming.
+windows-sys = { version = "0.52", features = [
+    "Win32_Storage_FileSystem",
+    "Win32_System_WindowsProgramming",
+] }

crates/quicksearch-core/examples/indexprobe.rs

@@ -0,0 +1,216 @@
+//! End-to-end timing and syscall accounting for a full indexing run.
+//!
+//! [`walkprobe`](walkprobe.rs) covers phase 1 alone, without a database. This
+//! covers the whole pipeline — parallel walk, `files` writes, and content
+//! extraction — because the interesting redundancy lives *between* the two
+//! phases: the walk reads a file's head to hash it and sniff its MIME, and
+//! extraction then reopens the same file and reads it again.
+//!
+//! ```text
+//! cargo build -p quicksearch-core --example indexprobe --release
+//! ./target/release/examples/indexprobe gen  /tmp/qs-bench
+//! ./target/release/examples/indexprobe cold /tmp/qs-bench /tmp/qs-bench.db
+//! ./target/release/examples/indexprobe warm /tmp/qs-bench /tmp/qs-bench.db
+//! ```
+//!
+//! `cold` deletes the database first, so every file is new: the walk hashes
+//! it and extraction reads it. `warm` re-runs over the existing database with
+//! the tree untouched, which is the case that has to stay at one `stat` per
+//! file — see [`crate::file_handling::classify_for_indexing`].
+//!
+//! For syscalls per file, trace a run and bucket by the tree's paths:
+//!
+//! ```text
+//! strace -f -y -o /tmp/t.log \
+//!     -e trace=openat,statx,newfstatat,fstat,read,pread64,readlink,close,getdents64,lseek \
+//!     ./target/release/examples/indexprobe cold /tmp/qs-bench /tmp/qs-bench.db
+//! grep -oP '^\d+ \K[a-z0-9_]+' <(grep '/tmp/qs-bench/' /tmp/t.log) | sort | uniq -c
+//! ```
+//!
+//! Group by thread id instead (`grep -oP '^\d+ [a-z0-9_]+'`) to see the split
+//! between the walk workers and the extraction thread.
+//!
+//! The run modes deliberately do no filesystem inspection of their own — no
+//! progress walk, no size survey — so that every syscall the trace attributes
+//! to the tree came from the indexer. The size histogram is printed by `gen`.
+
+use std::path::{Path, PathBuf};
+use std::time::{Duration, Instant};
+
+use quicksearch_core::config::Config;
+use quicksearch_core::indexing::{IndexingService, IndexingStatus};
+
+/// Files whose head the walk reads in full at the default 8 KiB
+/// `hash_length`, i.e. the ones extraction never needs to reopen.
+const SMALL_TEXT: usize = 800;
+/// Text files past `hash_length`, which extraction must still read.
+const LARGE_TEXT: usize = 100;
+/// No extractor claims these, so extraction resolves them without touching
+/// the disk. A control group: their cost must not move.
+const BINARY: usize = 100;
+
+const WORDS: &[&str] = &[
+    "alpha", "beta", "gamma", "delta", "epsilon", "zeta", "eta", "theta",
+    "quick", "brown", "fox", "jumps", "over", "lazy", "dog", "indexer",
+    "rust", "cargo", "sqlite", "baloo", "tokenizer", "trigram", "snippet",
+    "ocean", "forest", "mountain", "river", "valley", "bridge", "tunnel",
+    "morning", "afternoon", "evening", "midnight", "yesterday", "today",
+];
+
+/// Deterministic so two runs index byte-identical trees and their timings are
+/// comparable. Plain LCG — this only has to spread, not to be random.
+struct Rng(u64);
+
+impl Rng {
+    fn next(&mut self) -> u64 {
+        self.0 = self.0.wrapping_mul(6364136223846793005).wrapping_add(1442695040888963407);
+        self.0 >> 33
+    }
+
+    fn in_range(&mut self, lo: usize, hi: usize) -> usize {
+        lo + (self.next() as usize) % (hi - lo)
+    }
+}
+
+fn main() {
+    let mode = std::env::args().nth(1).unwrap_or_default();
+    let tree = PathBuf::from(std::env::args().nth(2).expect("usage: indexprobe <gen|cold|warm> <tree> [db]"));
+
+    match mode.as_str() {
+        "gen" => generate(&tree),
+        "cold" | "warm" => {
+            let db = PathBuf::from(std::env::args().nth(3).expect("usage: indexprobe <cold|warm> <tree> <db>"));
+            if mode == "cold" {
+                for suffix in ["", "-wal", "-shm"] {
+                    let _ = std::fs::remove_file(format!("{}{}", db.display(), suffix));
+                }
+            }
+            run(&mode, &tree, &db);
+        }
+        _ => {
+            eprintln!("usage: indexprobe <gen|cold|warm> <tree> [db]");
+            std::process::exit(2);
+        }
+    }
+}
+
+/// Build a tree with a size mix that separates the three code paths, and
+/// report it so results are self-describing.
+fn generate(tree: &Path) {
+    let _ = std::fs::remove_dir_all(tree);
+    std::fs::create_dir_all(tree).expect("create tree");
+
+    let mut rng = Rng(0x5eed);
+    let (mut small_bytes, mut large_bytes, mut bin_bytes) = (0usize, 0usize, 0usize);
+
+    // Spread across subdirectories so the walk does real directory work
+    // rather than one enormous readdir.
+    for i in 0..SMALL_TEXT {
+        let dir = tree.join(format!("src/mod{}", i % 40));
+        std::fs::create_dir_all(&dir).expect("mkdir");
+        let ext = ["txt", "md", "rs", "json"][i % 4];
+        let size = rng.in_range(200, 8 * 1024);
+        let body = prose(&mut rng, size);
+        small_bytes += body.len();
+        std::fs::write(dir.join(format!("f{}.{}", i, ext)), body).expect("write");
+    }
+
+    for i in 0..LARGE_TEXT {
+        let dir = tree.join(format!("docs/set{}", i % 10));
+        std::fs::create_dir_all(&dir).expect("mkdir");
+        let size = rng.in_range(8 * 1024 + 1, 200 * 1024);
+        let body = prose(&mut rng, size);
+        large_bytes += body.len();
+        std::fs::write(dir.join(format!("doc{}.md", i)), body).expect("write");
+    }
+
+    for i in 0..BINARY {
+        let dir = tree.join(format!("assets/set{}", i % 10));
+        std::fs::create_dir_all(&dir).expect("mkdir");
+        let n = rng.in_range(1024, 50 * 1024);
+        let blob: Vec<u8> = (0..n).map(|_| (rng.next() & 0xff) as u8).collect();
+        bin_bytes += blob.len();
+        std::fs::write(dir.join(format!("blob{}.bin", i)), blob).expect("write");
+    }
+
+    let total = SMALL_TEXT + LARGE_TEXT + BINARY;
+    eprintln!("generated {} files under {}", total, tree.display());
+    eprintln!(
+        "  text <= 8 KiB : {:5} files, {:8.1} MiB  (head covers the whole file)",
+        SMALL_TEXT,
+        small_bytes as f64 / (1024.0 * 1024.0)
+    );
+    eprintln!(
+        "  text >  8 KiB : {:5} files, {:8.1} MiB  (extraction must read it)",
+        LARGE_TEXT,
+        large_bytes as f64 / (1024.0 * 1024.0)
+    );
+    eprintln!(
+        "  binary        : {:5} files, {:8.1} MiB  (no extractor; control group)",
+        BINARY,
+        bin_bytes as f64 / (1024.0 * 1024.0)
+    );
+}
+
+fn prose(rng: &mut Rng, target: usize) -> String {
+    let mut s = String::with_capacity(target + 16);
+    while s.len() < target {
+        s.push_str(WORDS[rng.next() as usize % WORDS.len()]);
+        s.push(if rng.next() % 12 == 0 { '\n' } else { ' ' });
+    }
+    s.truncate(target);
+    s
+}
+
+fn run(mode: &str, tree: &Path, db: &Path) {
+    let config = Config::default();
+
+    // `run_indexing` writes this marker only on a successful finish, so it is
+    // the one unambiguous completion signal — polling the status enum races,
+    // because a small tree finishes between two polls and `Idle` then means
+    // both "not started" and "already done".
+    if db.exists() {
+        let conn = rusqlite::Connection::open(db).expect("open db");
+        conn.execute("DELETE FROM schema_info WHERE key = 'last_full_index'", [])
+            .expect("clear marker");
+    }
+
+    let service = IndexingService::new();
+    let start = Instant::now();
+    service
+        .start_indexing(
+            vec![tree.to_string_lossy().into_owned()],
+            db.to_string_lossy().into_owned(),
+            config,
+        )
+        .expect("start indexing");
+
+    let deadline = Instant::now() + Duration::from_secs(600);
+    let mut done = false;
+    while Instant::now() < deadline {
+        if let IndexingStatus::Error(e) = service.get_status() {
+            panic!("indexing failed: {}", e);
+        }
+        if db.exists() {
+            if let Ok(conn) = rusqlite::Connection::open(db) {
+                if quicksearch_core::db::repo::get_last_full_index(&conn).is_some() {
+                    done = true;
+                    break;
+                }
+            }
+        }
+        std::thread::sleep(Duration::from_millis(5));
+    }
+    let elapsed = start.elapsed();
+    assert!(done, "indexing did not finish within the timeout");
+    service.stop_indexing().expect("stop");
+
+    let total = SMALL_TEXT + LARGE_TEXT + BINARY;
+    eprintln!(
+        "{}: {:?} ({:.0} files/sec over {} files)",
+        mode,
+        elapsed,
+        total as f64 / elapsed.as_secs_f64(),
+        total
+    );
+}

crates/quicksearch-core/examples/walkprobe.rs

@@ -0,0 +1,126 @@
+//! Times and syscall-counts the phase-1 walk, without touching a database.
+//!
+//! ```text
+//! cargo build -p quicksearch-core --example walkprobe --release
+//! ./target/release/examples/walkprobe <root> parallel   # the threaded walker
+//! ./target/release/examples/walkprobe <root> serial     # one thread, for comparison
+//! ```
+//!
+//! Indexing a network share is bound by round trips, not bandwidth: every
+//! metadata operation that misses the client cache costs one, and throughput
+//! is round-trips-in-flight divided by latency. So the number that matters is
+//! syscalls per file, which this makes directly visible:
+//!
+//! ```text
+//! strace -f -c -e trace=openat,statx,newfstatat,readlink,getdents64,read,lseek,close \
+//!     ./target/release/examples/walkprobe <root> parallel
+//! ```
+//!
+//! Expect roughly one `statx` per unchanged file, plus open/read/close for
+//! files that are new or modified, and `readlink` only for resolving the roots
+//! themselves — a per-file `readlink` count means a `canonicalize` has crept
+//! back into the hot path.
+//!
+//! Those four syscalls also now cover the *whole* cost of a small text file:
+//! the head read for the hash is the file's entire contents, so the walk
+//! extracts its text there and the content pass never opens it again. That
+//! work is CPU, not syscalls, so it shows up in files/sec here and not in the
+//! trace. Use [`indexprobe`](indexprobe.rs) to see both phases together.
+//!
+//! Both modes report files/sec. Run each twice: the first pass warms the page
+//! cache (or, on a share, the client's attribute cache), so the second is the
+//! one to compare.
+use std::collections::HashMap;
+use std::sync::atomic::AtomicBool;
+use std::sync::{Arc, Mutex};
+use std::time::{Instant, UNIX_EPOCH};
+
+use quicksearch_core::config::{Config, IgnoreSet};
+use quicksearch_core::extract::Registry;
+use quicksearch_core::file_handling::{
+    classify_for_indexing, filtered_walk, prepare_file_record, ExistingFileEntry, FileIndexAction,
+    UnreadableDirs,
+};
+use quicksearch_core::walk::walk_indexable_files;
+
+fn main() {
+    let root = std::env::args().nth(1).unwrap();
+    let mode = std::env::args().nth(2).unwrap_or_else(|| "parallel".into());
+    let config = Config::default();
+    let existing: HashMap<String, ExistingFileEntry> = HashMap::new();
+
+    let start = Instant::now();
+    let (seen, prepared) = match mode.as_str() {
+        "serial" => serial(&root, &config, &existing),
+        _ => parallel(&root, &config, existing),
+    };
+    let elapsed = start.elapsed();
+
+    eprintln!(
+        "{mode}: {seen} files, {prepared} prepared in {:?} ({:.0} files/sec)",
+        elapsed,
+        seen as f64 / elapsed.as_secs_f64()
+    );
+}
+
+fn serial(
+    root: &str,
+    config: &Config,
+    existing: &HashMap<String, ExistingFileEntry>,
+) -> (usize, usize) {
+    let ignore = IgnoreSet::compile(&[]).unwrap();
+    let registry = Registry::default_set();
+    let (mut seen, mut prepared) = (0, 0);
+    for entry in filtered_walk(root, false, false, &ignore, &UnreadableDirs::default()) {
+        seen += 1;
+        // Same rule as the real walk: a name that is not valid UTF-8 cannot be
+        // stored in `files.path` and reopened by it, so it is skipped before
+        // anything tries to hash it. Counted as seen, never prepared.
+        let Some(path) = entry.path().to_str().map(str::to_owned) else {
+            continue;
+        };
+        let Ok(meta) = std::fs::metadata(entry.path()) else {
+            continue;
+        };
+        let Some(mtime) = meta
+            .modified()
+            .ok()
+            .and_then(|t| t.duration_since(UNIX_EPOCH).ok())
+            .map(|d| d.as_secs())
+        else {
+            continue;
+        };
+        if classify_for_indexing(&path, mtime, existing) != FileIndexAction::Skip
+            && prepare_file_record(&path, &meta, config, &registry).is_some()
+        {
+            prepared += 1;
+        }
+    }
+    (seen, prepared)
+}
+
+fn parallel(
+    root: &str,
+    config: &Config,
+    existing: HashMap<String, ExistingFileEntry>,
+) -> (usize, usize) {
+    let (mut seen, mut prepared) = (0, 0);
+    for file in walk_indexable_files(
+        &[root.to_string()],
+        false,
+        false,
+        IgnoreSet::compile(&[]).unwrap(),
+        Arc::new(existing),
+        config.clone(),
+        Arc::new(Registry::default_set()),
+        Arc::new(Mutex::new(false)),
+        Arc::new(AtomicBool::new(false)),
+        4,
+    ) {
+        seen += 1;
+        if file.record.is_some() {
+            prepared += 1;
+        }
+    }
+    (seen, prepared)
+}

crates/quicksearch-core/src/cli.rs

@@ -8,7 +8,7 @@
 
 use rusqlite::{params, OptionalExtension};
 
-use crate::db::open_or_recreate;
+use crate::db::open_existing;
 use crate::db::repo::{STATE_DONE, STATE_FAILED, STATE_NA, STATE_PENDING};
 
 /// Per-file indexing status, mirroring Baloo's multi-state reporting.
@@ -86,7 +86,7 @@ impl SizeReport {
 /// Query the per-file indexing status. Returns `FileStatus` with
 /// `basic == NotIndexed` if the path isn't in the database.
 pub fn status_for_path(db_path: &str, path: &str) -> Result<FileStatus, String> {
-    let conn = open_or_recreate(db_path, "trigram")?;
+    let conn = open_existing(db_path, false)?;
     let row: Option<(i64, i64, Option<String>)> = conn
         .query_row(
             "SELECT basic_state, content_state, failure_msg FROM files WHERE path = ?1",
@@ -113,7 +113,7 @@ pub fn status_for_path(db_path: &str, path: &str) -> Result<FileStatus, String>
 
 /// Return every file that failed content extraction, newest first.
 pub fn list_failed(db_path: &str, limit: Option<u32>) -> Result<Vec<FailedEntry>, String> {
-    let conn = open_or_recreate(db_path, "trigram")?;
+    let conn = open_existing(db_path, false)?;
     let limit_sql = match limit {
         Some(n) => format!(" LIMIT {}", n),
         None => String::new(),
@@ -147,7 +147,7 @@ pub fn index_size_breakdown(db_path: &str) -> Result<SizeReport, String> {
     let file_size_bytes = std::fs::metadata(db_path)
         .map(|m| m.len())
         .unwrap_or(0);
-    let conn = open_or_recreate(db_path, "trigram")?;
+    let conn = open_existing(db_path, false)?;
     let count = |table: &str| -> Result<i64, String> {
         conn.query_row(&format!("SELECT COUNT(*) FROM {}", table), [], |r| r.get(0))
             .map_err(|e| format!("count {}: {}", table, e))
@@ -181,7 +181,7 @@ pub fn index_size_breakdown(db_path: &str) -> Result<SizeReport, String> {
 /// Used by the Baloo compat daemon to report the "Files waiting for content
 /// indexing" figure both to balooctl and to the LMDB mirror.
 pub fn pending_content_count(db_path: &str) -> Result<i64, String> {
-    let conn = open_or_recreate(db_path, "trigram")?;
+    let conn = open_existing(db_path, false)?;
     conn.query_row(
         "SELECT COUNT(*) FROM files WHERE content_state = ?1",
         rusqlite::params![crate::db::repo::STATE_PENDING],
@@ -190,10 +190,33 @@ pub fn pending_content_count(db_path: &str) -> Result<i64, String> {
     .map_err(|e| format!("pending_content_count: {}", e))
 }
 
+/// Cheap aggregate counts for the GUI's idle status bar ("N files
+/// indexed"). Callers cache the result; it's three COUNT scans, not
+/// something to run per frame.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub struct IndexCounts {
+    pub files: i64,
+    pub content_done: i64,
+    pub content_pending: i64,
+}
+
+pub fn index_counts(db_path: &str) -> Result<IndexCounts, String> {
+    let conn = open_existing(db_path, false)?;
+    let count = |sql: &str| -> Result<i64, String> {
+        conn.query_row(sql, [], |r| r.get(0))
+            .map_err(|e| format!("index_counts: {}", e))
+    };
+    Ok(IndexCounts {
+        files: count("SELECT COUNT(*) FROM files")?,
+        content_done: count("SELECT COUNT(*) FROM files WHERE content_state = 1")?,
+        content_pending: count("SELECT COUNT(*) FROM files WHERE content_state = 0")?,
+    })
+}
+
 /// Remove a single file from the index. Returns whether a row was deleted.
 /// Keeps FTS/documents/properties in sync via the repo helpers.
 pub fn clear_path(db_path: &str, path: &str) -> Result<bool, String> {
-    let mut conn = open_or_recreate(db_path, "trigram")?;
+    let mut conn = open_existing(db_path, true)?;
     let tx = conn
         .transaction()
         .map_err(|e| format!("clear_path begin tx: {}", e))?;
@@ -206,6 +229,7 @@ pub fn clear_path(db_path: &str, path: &str) -> Result<bool, String> {
 #[cfg(test)]
 mod tests {
     use super::*;
+    use crate::db::open_or_recreate;
     use crate::db::repo::{insert_file, set_content_done, set_content_failed, NewFile};
     use crate::mime::FileType;
 
@@ -374,4 +398,51 @@ mod tests {
 
         std::fs::remove_file(&p).ok();
     }
+
+    #[test]
+    fn clear_path_on_nondefault_tokenizer_db_removes_only_target() {
+        // Regression: clear_path used to open with a hardcoded "trigram", so on
+        // an index built with a non-default tokenizer the schema-mismatch wipe
+        // destroyed the WHOLE index instead of deleting one row. With
+        // open_existing it must delete only the target and leave the rest.
+        let p = tmp_path();
+        let dbp = p.to_str().unwrap();
+        {
+            let mut conn = open_or_recreate(dbp, "unicode61").unwrap();
+            let tx = conn.transaction().unwrap();
+            for (name, path) in [("a.txt", "/tmp/a.txt"), ("b.txt", "/tmp/b.txt")] {
+                insert_file(
+                    &tx,
+                    &NewFile {
+                        name,
+                        path,
+                        parent: "/tmp",
+                        size: 1,
+                        mtime: 1,
+                        inode: None,
+                        device_id: None,
+                        mime: Some("text/plain"),
+                        ftype: FileType::TEXT,
+                        hash: None,
+                    },
+                )
+                .unwrap()
+                .expect("unique path");
+            }
+            tx.commit().unwrap();
+        }
+
+        assert!(clear_path(dbp, "/tmp/a.txt").unwrap());
+        // The other row must survive — proof we deleted one row, not wiped.
+        assert_eq!(
+            status_for_path(dbp, "/tmp/b.txt").unwrap().basic,
+            IndexState::Done
+        );
+        assert_eq!(
+            status_for_path(dbp, "/tmp/a.txt").unwrap().basic,
+            IndexState::NotIndexed
+        );
+
+        std::fs::remove_file(&p).ok();
+    }
 }

crates/quicksearch-core/src/db/mod.rs

@@ -1,13 +1,15 @@
 //! SQLite schema, on-disk open/recreate, and row-level repository helpers.
 //!
-//! Policy: a single [`open::open_or_recreate`] is the only entry point. Any
+//! Policy: the indexer (owner) opens via [`open::open_or_recreate`], which on
-//! schema mismatch — wrong version, drifted tokenizer, absent `schema_info`
+//! any schema mismatch — wrong version, drifted tokenizer, absent
-//! — wipes the DB and rebuilds from [`schema::SCHEMA_CURRENT`]. There are
+//! `schema_info` — wipes the DB and rebuilds from [`schema::SCHEMA_CURRENT`].
-//! no in-place migrations by design; re-indexing is accepted as the cost
+//! There are no in-place migrations by design; re-indexing is accepted as the
-//! of avoiding migration-path complexity.
+//! cost of avoiding migration-path complexity. *Consumers* (search, status,
+//! size, `clear`) instead use [`open::open_existing`], which never creates or
+//! wipes — a tokenizer difference or stale version is an error, not data loss.
 
 pub mod open;
 pub mod repo;
 pub mod schema;
 
-pub use open::{open_or_recreate, CURRENT_SCHEMA_VERSION};
+pub use open::{open_existing, open_or_recreate, CURRENT_SCHEMA_VERSION};

crates/quicksearch-core/src/db/open.rs

@@ -15,9 +15,11 @@
 
 use std::path::Path;
 
-use rusqlite::{params, Connection, OptionalExtension};
+use rusqlite::{params, Connection, OpenFlags, OptionalExtension};
 
-use super::schema::{effective_tokenizer, fts_create_sql, PRAGMAS_FAST, SCHEMA_CURRENT};
+use super::schema::{
+    effective_tokenizer, fts_create_sql, PRAGMAS_FAST, PRAGMAS_READONLY, SCHEMA_CURRENT,
+};
 
 /// Bump this whenever [`SCHEMA_CURRENT`] or [`fts_create_sql`] changes in
 /// a way that makes an old DB unreadable by new code. Any such bump
@@ -34,6 +36,14 @@ pub const CURRENT_SCHEMA_VERSION: u32 = 3;
 /// schema mismatch and triggers the wipe-and-recreate path.
 pub fn open_or_recreate(db_path: &str, tokenizer: &str) -> Result<Connection, String> {
     let path = Path::new(db_path).to_path_buf();
+    // The owner creates the directory too — a fresh install's default
+    // XDG data dir doesn't exist until first use.
+    if let Some(dir) = path.parent() {
+        if !dir.as_os_str().is_empty() {
+            std::fs::create_dir_all(dir)
+                .map_err(|e| format!("Failed to create database dir {}: {}", dir.display(), e))?;
+        }
+    }
     let conn = Connection::open(db_path)
         .map_err(|e| format!("Failed to open database at {}: {}", db_path, e))?;
     conn.execute_batch(PRAGMAS_FAST)
@@ -46,8 +56,8 @@ pub fn open_or_recreate(db_path: &str, tokenizer: &str) -> Result<Connection, St
     // Schema is present but stale, or pre-existing rows belong to an
     // older layout, or the tokenizer drifted. Log once so the rebuild
     // isn't silent, then wipe + recreate.
-    eprintln!(
+    crate::log_warn!(
-        "QuickSearch: database at {} does not match current schema; rebuilding. \
+        "database at {} does not match current schema; rebuilding. \
          Existing rows will be re-scanned on next indexing run.",
         db_path
     );
@@ -56,10 +66,47 @@ pub fn open_or_recreate(db_path: &str, tokenizer: &str) -> Result<Connection, St
     Ok(conn)
 }
 
-/// True iff the DB has `schema_info` with the current version *and* the
+/// Open an *existing* index without ever recreating it. Verifies the schema
-/// effective-tokenizer string this caller asked for. Anything else —
+/// version matches this build; on any mismatch — missing file, no
-/// missing table, wrong version, different tokenizer — returns false.
+/// `schema_info`, wrong version — returns an error instead of wiping. The
-fn db_matches_current(conn: &Connection, tokenizer: &str) -> Result<bool, String> {
+/// on-disk FTS tokenizer is used as-is: a tokenizer difference is never a
+/// reason to destroy a readable index.
+///
+/// `write == false` opens read-only; `write == true` opens read-write (for
+/// row-level deletes like `clear`) but still never creates or wipes — there
+/// is no `SQLITE_OPEN_CREATE`, so a missing file is a clean error.
+///
+/// Use this for every *consumer* (search, status, size, `clear`). Only the
+/// indexer's own write path uses [`open_or_recreate`], which may wipe on a
+/// genuine schema/tokenizer change it owns.
+pub fn open_existing(db_path: &str, write: bool) -> Result<Connection, String> {
+    let flags = OpenFlags::SQLITE_OPEN_NO_MUTEX
+        | if write {
+            OpenFlags::SQLITE_OPEN_READ_WRITE
+        } else {
+            OpenFlags::SQLITE_OPEN_READ_ONLY
+        };
+    let conn = Connection::open_with_flags(db_path, flags)
+        .map_err(|e| format!("Failed to open database at {}: {}", db_path, e))?;
+    let pragmas = if write { PRAGMAS_FAST } else { PRAGMAS_READONLY };
+    conn.execute_batch(pragmas)
+        .map_err(|e| format!("Failed to apply pragmas: {}", e))?;
+
+    if !schema_version_current(&conn)? {
+        return Err(format!(
+            "index at {} is not a compatible QuickSearch index (schema v{} expected); \
+             refusing to modify it. Re-index to rebuild.",
+            db_path, CURRENT_SCHEMA_VERSION
+        ));
+    }
+    Ok(conn)
+}
+
+/// True iff the DB has a `schema_info` table whose `version` equals
+/// [`CURRENT_SCHEMA_VERSION`]. Shared by the wipe decision
+/// ([`db_matches_current`]) and the non-destructive [`open_existing`] path.
+/// Deliberately ignores the tokenizer — that's only the owner's concern.
+fn schema_version_current(conn: &Connection) -> Result<bool, String> {
     let has_info: bool = conn
         .query_row(
             "SELECT 1 FROM sqlite_master WHERE type='table' AND name='schema_info'",
@@ -81,8 +128,14 @@ fn db_matches_current(conn: &Connection, tokenizer: &str) -> Result<bool, String
         )
         .optional()
         .map_err(|e| format!("read schema_info.version: {}", e))?;
-    let version_ok = version.as_deref() == Some(&CURRENT_SCHEMA_VERSION.to_string());
+    Ok(version.as_deref() == Some(&CURRENT_SCHEMA_VERSION.to_string()))
-    if !version_ok {
+}
+
+/// True iff the DB has `schema_info` with the current version *and* the
+/// effective-tokenizer string this caller asked for. Anything else —
+/// missing table, wrong version, different tokenizer — returns false.
+fn db_matches_current(conn: &Connection, tokenizer: &str) -> Result<bool, String> {
+    if !schema_version_current(conn)? {
         return Ok(false);
     }
 
@@ -104,10 +157,22 @@ fn wipe_and_reopen(conn: Connection, path: &Path) -> Result<Connection, String>
     drop(conn);
     // Primary file may already be absent (fresh open that just needed
     // the table applied). Ignore NotFound; anything else is an error.
-    match std::fs::remove_file(path) {
+    //
+    // `remove_file_retrying` matters on Windows, where a delete fails while
+    // *any* handle is open — most often an antivirus scanner reading the file
+    // in the moment after we closed it. Unix `unlink` never hits this, so the
+    // retry costs nothing there.
+    match crate::platform::remove_file_retrying(path) {
         Ok(()) => {}
         Err(e) if e.kind() == std::io::ErrorKind::NotFound => {}
-        Err(e) => return Err(format!("Failed to remove old database: {}", e)),
+        Err(e) => {
+            return Err(format!(
+                "Failed to remove old database at {}: {}. \
+                 Another QuickSearch instance may have the index open.",
+                path.display(),
+                e
+            ))
+        }
     }
     // Sidecars are optional — delete best-effort.
     for suffix in ["-wal", "-shm", "-journal"] {
@@ -116,7 +181,7 @@ fn wipe_and_reopen(conn: Connection, path: &Path) -> Result<Connection, String>
             path.file_name().and_then(|s| s.to_str()).unwrap_or(""),
             suffix
         ));
-        let _ = std::fs::remove_file(sidecar);
+        let _ = crate::platform::remove_file_retrying(&sidecar);
     }
     let conn = Connection::open(path)
         .map_err(|e| format!("Failed to reopen database after rebuild: {}", e))?;
@@ -326,4 +391,166 @@ mod tests {
         drop(conn);
         std::fs::remove_file(&p).ok();
     }
+
+    #[test]
+    fn open_existing_reads_nondefault_tokenizer_without_wiping() {
+        // The exact scenario that previously caused data loss: an index built
+        // with a non-default tokenizer, then opened by a *consumer* that only
+        // knows "trigram". `open_existing` must read it as-is and never wipe.
+        let p = tmp_db_path();
+        {
+            let conn = open_or_recreate(p.to_str().unwrap(), "unicode61").unwrap();
+            conn.execute(
+                "INSERT INTO files (name, path, parent, size, mtime) \
+                 VALUES ('note', '/note.txt', '/', 0, 0)",
+                [],
+            )
+            .unwrap();
+            // Seed the FTS index (rowid = the files row we just inserted) so a
+            // MATCH query can be exercised against the on-disk tokenizer.
+            conn.execute(
+                "INSERT INTO searchabletext (rowid, name, text, properties) \
+                 VALUES (last_insert_rowid(), 'note', 'hello world', '')",
+                [],
+            )
+            .unwrap();
+        }
+
+        let conn = open_existing(p.to_str().unwrap(), false).unwrap();
+        let files: i64 = conn
+            .query_row("SELECT COUNT(*) FROM files", [], |r| r.get(0))
+            .unwrap();
+        assert_eq!(
+            files, 1,
+            "open_existing must not wipe a non-default-tokenizer DB"
+        );
+        // The on-disk tokenizer is used as-is: a MATCH against the stored term
+        // returns the row.
+        let hits: i64 = conn
+            .query_row(
+                "SELECT COUNT(*) FROM searchabletext WHERE searchabletext MATCH 'hello'",
+                [],
+                |r| r.get(0),
+            )
+            .unwrap();
+        assert_eq!(hits, 1);
+        // And the stored tokenizer is still the non-default one — proof we
+        // neither rewrote the FTS table nor reset schema_info.
+        let tok: String = conn
+            .query_row(
+                "SELECT value FROM schema_info WHERE key='tokenize'",
+                [],
+                |r| r.get(0),
+            )
+            .unwrap();
+        assert_eq!(tok, "unicode61");
+        drop(conn);
+        std::fs::remove_file(&p).ok();
+    }
+
+    #[test]
+    fn open_or_recreate_creates_missing_parent_dirs() {
+        // Fresh installs point at ~/.local/share/quicksearch/… which
+        // doesn't exist yet; the owner open must create it.
+        let mut dir = std::env::temp_dir();
+        dir.push(format!(
+            "qs-mkdir-{}-{}",
+            std::process::id(),
+            std::time::SystemTime::now()
+                .duration_since(std::time::UNIX_EPOCH)
+                .unwrap()
+                .as_nanos()
+        ));
+        let db = dir.join("nested/deeper/index.sqlite");
+        let conn = open_or_recreate(db.to_str().unwrap(), "trigram").unwrap();
+        drop(conn);
+        assert!(db.exists());
+        std::fs::remove_dir_all(&dir).ok();
+    }
+
+    #[test]
+    fn writable_opens_use_wal_and_it_persists() {
+        let p = tmp_db_path();
+        {
+            let conn = open_or_recreate(p.to_str().unwrap(), "trigram").unwrap();
+            let mode: String = conn
+                .query_row("PRAGMA journal_mode", [], |r| r.get(0))
+                .unwrap();
+            assert_eq!(mode.to_lowercase(), "wal");
+        }
+        // WAL is persistent in the file: a later read-only consumer sees it
+        // without being able to (or needing to) set it.
+        let conn = open_existing(p.to_str().unwrap(), false).unwrap();
+        let mode: String = conn
+            .query_row("PRAGMA journal_mode", [], |r| r.get(0))
+            .unwrap();
+        assert_eq!(mode.to_lowercase(), "wal");
+        drop(conn);
+        std::fs::remove_file(&p).ok();
+    }
+
+    #[test]
+    fn open_existing_errors_on_missing_file() {
+        let p = tmp_db_path();
+        assert!(!p.exists());
+        let res = open_existing(p.to_str().unwrap(), false);
+        assert!(res.is_err(), "missing file must error, not be created");
+        assert!(!p.exists(), "open_existing must not create the file");
+    }
+
+    #[test]
+    fn open_existing_errors_on_version_mismatch_without_wiping() {
+        // A DB from a prior schema version. A consumer opening it must get an
+        // error and leave the file untouched — the data is the owner's to
+        // rebuild, never a reader's to destroy.
+        let p = tmp_db_path();
+        {
+            let conn = Connection::open(&p).unwrap();
+            conn.execute(
+                "CREATE TABLE schema_info (key TEXT PRIMARY KEY, value TEXT NOT NULL)",
+                [],
+            )
+            .unwrap();
+            conn.execute(
+                "INSERT INTO schema_info(key,value) VALUES('version','1')",
+                [],
+            )
+            .unwrap();
+            conn.execute("CREATE TABLE files (id INTEGER PRIMARY KEY, name TEXT)", [])
+                .unwrap();
+            conn.execute("INSERT INTO files(name) VALUES('sentinel')", [])
+                .unwrap();
+        }
+        let res = open_existing(p.to_str().unwrap(), false);
+        assert!(res.is_err(), "stale schema version must error");
+        // Sentinel row still present → the file was not wiped.
+        let conn = Connection::open(&p).unwrap();
+        let n: i64 = conn
+            .query_row("SELECT COUNT(*) FROM files", [], |r| r.get(0))
+            .unwrap();
+        assert_eq!(n, 1, "open_existing must never delete on version mismatch");
+        drop(conn);
+        std::fs::remove_file(&p).ok();
+    }
+
+    #[test]
+    fn open_existing_rw_allows_delete() {
+        let p = tmp_db_path();
+        {
+            let conn = open_or_recreate(p.to_str().unwrap(), "trigram").unwrap();
+            conn.execute(
+                "INSERT INTO files (name, path, parent, size, mtime) \
+                 VALUES ('a', '/a', '/', 0, 0)",
+                [],
+            )
+            .unwrap();
+        }
+        let conn = open_existing(p.to_str().unwrap(), true).unwrap();
+        let removed = conn
+            .execute("DELETE FROM files WHERE path = '/a'", [])
+            .unwrap();
+        assert_eq!(removed, 1);
+        drop(conn);
+        std::fs::remove_file(&p).ok();
+    }
 }

crates/quicksearch-core/src/db/repo.rs

@@ -278,17 +278,41 @@ fn encode_properties_for_fts(props: &[(String, String)]) -> String {
     buf
 }
 
-/// Flush and close a connection, restoring durable PRAGMAs. Call on clean
+/// Flush the WAL into the main DB file and close. Call on clean shutdown so
-/// shutdown so the next open sees a consistent DB.
+/// the next open starts with an empty log. WAL mode itself is persistent in
+/// the file — deliberately left on.
 pub fn checkpoint_and_close(conn: Connection) {
-    let _ = conn.execute_batch(
+    let _ = conn.execute_batch("PRAGMA wal_checkpoint(TRUNCATE);");
-        "PRAGMA journal_mode = DELETE; \
-         PRAGMA synchronous = FULL; \
-         PRAGMA wal_checkpoint(FULL);",
-    );
     drop(conn);
 }
 
+/// Read the `last_full_index` marker (unix seconds of the last *successful*
+/// full indexing run) from `schema_info`. Absent key — fresh DB, or a DB
+/// from before this marker existed — means "never".
+pub fn get_last_full_index(conn: &Connection) -> Option<u64> {
+    conn.query_row(
+        "SELECT value FROM schema_info WHERE key = 'last_full_index'",
+        [],
+        |r| r.get::<_, String>(0),
+    )
+    .optional()
+    .ok()
+    .flatten()
+    .and_then(|v| v.parse().ok())
+}
+
+/// Stamp `last_full_index` with `ts` (unix seconds). Called at the end of
+/// every successful full indexing run; the coordinator reads it to schedule
+/// periodic reindexing.
+pub fn set_last_full_index(conn: &Connection, ts: u64) -> Result<(), String> {
+    conn.execute(
+        "INSERT OR REPLACE INTO schema_info(key, value) VALUES ('last_full_index', ?1)",
+        params![ts.to_string()],
+    )
+    .map_err(|e| format!("write last_full_index: {}", e))?;
+    Ok(())
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
@@ -480,6 +504,59 @@ mod tests {
         std::fs::remove_file(&p).ok();
     }
 
+    #[test]
+    fn last_full_index_round_trip() {
+        let p = tmp_path();
+        let conn = open_or_recreate(p.to_str().unwrap(), "trigram").unwrap();
+        assert_eq!(get_last_full_index(&conn), None, "fresh DB has no marker");
+        set_last_full_index(&conn, 1_700_000_123).unwrap();
+        assert_eq!(get_last_full_index(&conn), Some(1_700_000_123));
+        // Overwrite, not accumulate.
+        set_last_full_index(&conn, 1_700_000_999).unwrap();
+        assert_eq!(get_last_full_index(&conn), Some(1_700_000_999));
+        drop(conn);
+        std::fs::remove_file(&p).ok();
+    }
+
+    #[test]
+    fn checkpoint_and_close_truncates_wal() {
+        let p = tmp_path();
+        let mut conn = open_or_recreate(p.to_str().unwrap(), "trigram").unwrap();
+        {
+            let tx = conn.transaction().unwrap();
+            insert_file(
+                &tx,
+                &NewFile {
+                    name: "w.txt",
+                    path: "/tmp/w.txt",
+                    parent: "/tmp",
+                    size: 1,
+                    mtime: 1,
+                    inode: None,
+                    device_id: None,
+                    mime: None,
+                    ftype: FileType::EMPTY,
+                    hash: None,
+                },
+            )
+            .unwrap();
+            tx.commit().unwrap();
+        }
+        checkpoint_and_close(conn);
+        // After a TRUNCATE checkpoint + close of the last connection the WAL
+        // sidecar is gone or empty; the row lives in the main file.
+        let wal = std::path::PathBuf::from(format!("{}-wal", p.display()));
+        let wal_len = std::fs::metadata(&wal).map(|m| m.len()).unwrap_or(0);
+        assert_eq!(wal_len, 0, "WAL should be truncated on clean close");
+        let conn = crate::db::open_existing(p.to_str().unwrap(), false).unwrap();
+        let n: i64 = conn
+            .query_row("SELECT COUNT(*) FROM files", [], |r| r.get(0))
+            .unwrap();
+        assert_eq!(n, 1);
+        drop(conn);
+        std::fs::remove_file(&p).ok();
+    }
+
     #[test]
     fn set_content_failed_writes_failed_table() {
         let p = tmp_path();

crates/quicksearch-core/src/db/schema.rs

@@ -1,12 +1,32 @@
 //! SQL strings for the current schema. Versioned; [`migrate`](super::migrate)
 //! drives the upgrade path.
 
-/// Pragmas applied on every connection open. Tuned for write throughput during
+/// Pragmas applied on every writable connection open.
-/// indexing; a clean shutdown re-enables journal_mode/synchronous via
+///
+/// WAL, not journal-off: auto-indexing writes continuously while searches
+/// stream from their own read-only connections, and WAL is what lets those
+/// readers proceed without ever blocking the writer (or vice versa).
+/// `synchronous = NORMAL` under WAL risks only the last commit on power
+/// loss — acceptable for an index that is re-derivable from disk. Only two
+/// writers exist (full index runs and the coordinator's incremental
+/// updates) and they're serialized by design; `busy_timeout` is a backstop,
+/// not a coordination mechanism. A clean shutdown truncates the log via
 /// [`super::repo::checkpoint_and_close`].
 pub const PRAGMAS_FAST: &str = "
-    PRAGMA journal_mode = OFF;
+    PRAGMA journal_mode = WAL;
-    PRAGMA synchronous = 0;
+    PRAGMA synchronous = NORMAL;
+    PRAGMA busy_timeout = 5000;
+    PRAGMA cache_size = 10000;
+    PRAGMA temp_store = MEMORY;
+    PRAGMA foreign_keys = ON;
+";
+
+/// Pragmas safe to apply on a read-only connection, where `journal_mode`
+/// and `synchronous` can't be changed on the file. Used by
+/// [`super::open::open_existing`] for read-only opens; write paths get the
+/// full [`PRAGMAS_FAST`] set.
+pub const PRAGMAS_READONLY: &str = "
+    PRAGMA busy_timeout = 5000;
     PRAGMA cache_size = 10000;
     PRAGMA temp_store = MEMORY;
     PRAGMA foreign_keys = ON;

crates/quicksearch-core/src/extract/mod.rs

@@ -2,8 +2,15 @@
 //!
 //! An [`Extractor`] decides whether it can handle a given MIME type and, if
 //! so, produces [`ExtractedContent`] for the file. The [`Registry`] picks the
-//! first registered extractor that accepts the MIME and runs it. Callers can
+//! first registered extractor that accepts the MIME and runs it.
-//! also fall back to an extension-based match for files with no detected MIME.
+//!
+//! Dispatch is by MIME only — a file with no detected type is recorded as
+//! "not applicable" rather than guessed at again here. Extensions that
+//! `mime_guess` misses or mistypes are corrected upstream instead, in
+//! [`crate::mime::guess_mime_from_head`], so there is one place where "what is
+//! this file" gets decided, and it is decided once: the walk sniffs the head it
+//! already read, stores the answer, and nothing downstream reopens the file to
+//! ask again.
 
 use std::collections::HashMap;
 use std::path::Path;
@@ -70,6 +77,32 @@ pub trait Extractor: Send + Sync {
     /// [`ExtractError`] to mark the file's content state as failed (so it
     /// won't be retried every run).
     fn extract(&self, path: &Path) -> Result<ExtractedContent, ExtractError>;
+
+    /// Extract from bytes the caller already holds, when those bytes are the
+    /// file's *entire* contents.
+    ///
+    /// Indexing hashes the head of every new or changed file, so for anything
+    /// no larger than `hash_length` the whole file is already in memory by the
+    /// time the walk classifies it. An extractor that can work from that buffer
+    /// saves the content pass an open/read/close — several round trips per
+    /// file on a network share — and closes a consistency gap, because the
+    /// text then comes from the same `read` as the size, mtime and hash stored
+    /// alongside it.
+    ///
+    /// The default is `None`: "I need the file on disk." Formats that seek,
+    /// or that read a central directory at the end of the file, must keep it.
+    /// Returning `Some(Err(_))` is a real extraction failure, recorded like
+    /// any other; returning `None` simply defers to [`Extractor::extract`].
+    ///
+    /// `path` is passed only so failures name the same file the on-disk path
+    /// would — nothing here may open it.
+    fn extract_from_head(
+        &self,
+        _path: &Path,
+        _head: &[u8],
+    ) -> Option<Result<ExtractedContent, ExtractError>> {
+        None
+    }
 }
 
 /// An ordered dispatch table of extractors. The first extractor whose
@@ -105,6 +138,28 @@ impl Registry {
         Ok(None)
     }
 
+    /// [`Registry::extract`] for a file whose complete contents the caller
+    /// already holds. `Ok(None)` when no extractor claims the MIME, and
+    /// `None` when the one that does needs the file on disk after all —
+    /// both mean "leave this to the content pass".
+    ///
+    /// Dispatch stays here rather than at the call site so there is still
+    /// exactly one place that decides what an extractor sees for a given MIME.
+    pub fn extract_complete_head(
+        &self,
+        path: &Path,
+        mime: &str,
+        head: &[u8],
+    ) -> Option<Result<ExtractedContent, ExtractError>> {
+        let lower = mime.to_ascii_lowercase();
+        for e in &self.extractors {
+            if e.supports(&lower) {
+                return e.extract_from_head(path, head);
+            }
+        }
+        None
+    }
+
     /// The default set wired up for Set A: plaintext, office docs, PDF,
     /// audio tags, image EXIF.
     pub fn default_set() -> Self {
@@ -136,6 +191,38 @@ mod tests {
         assert!(out.is_none());
     }
 
+    #[test]
+    fn complete_head_extraction_dispatches_only_to_extractors_that_opt_in() {
+        let r = Registry::default_set();
+        let p = Path::new("/tmp/whatever");
+
+        // Plaintext opts in, so a small text file never reaches the disk pass.
+        let out = r.extract_complete_head(p, "text/plain", b"hello");
+        assert!(matches!(out, Some(Ok(ref c)) if c.text == "hello"));
+
+        // A format that seeks or reads a trailer must not be handed a buffer.
+        // `None` here is what routes it back to the on-disk extractor.
+        assert!(r.extract_complete_head(p, "application/pdf", b"%PDF-1.4").is_none());
+        assert!(r.extract_complete_head(p, "image/png", b"\x89PNG").is_none());
+
+        // No extractor claims the MIME at all.
+        assert!(r.extract_complete_head(p, "application/x-nonesuch", b"..").is_none());
+    }
+
+    #[test]
+    fn complete_head_extraction_matches_the_on_disk_dispatch() {
+        // Both entry points must pick the same extractor for a MIME, or a
+        // file's text would depend on which pass happened to handle it.
+        let r = Registry::default_set();
+        let p = Path::new("/tmp/whatever");
+        for mime in ["text/plain", "TEXT/PLAIN", "application/json", "application/x-sql"] {
+            assert!(
+                r.extract_complete_head(p, mime, b"x").is_some(),
+                "{} should extract from a head", mime
+            );
+        }
+    }
+
     #[test]
     fn properties_sorted_is_deterministic() {
         let c = ExtractedContent::with_text("hi")

crates/quicksearch-core/src/extract/pdf.rs

@@ -6,12 +6,48 @@
 //! failed. Properties (title, author, etc.) from the PDF `Info` dictionary
 //! are pulled via `lopdf` where available.
 
+use std::cell::Cell;
 use std::path::Path;
+use std::sync::OnceLock;
 
 use lopdf::{Document as LopdfDocument, Object};
 
 use super::{ExtractError, ExtractedContent, Extractor};
 
+thread_local! {
+    /// True while this thread is inside a contained `pdf_extract` call.
+    static SUPPRESS_PANIC_PRINT: Cell<bool> = const { Cell::new(false) };
+}
+
+/// Chain a process panic hook (once) that swallows the default
+/// "thread panicked at …" report while this thread is inside a *contained*
+/// PDF extraction — those panics are expected on malformed PDFs, caught,
+/// and recorded as the file's failure reason, so printing each one is pure
+/// console spam. Panics anywhere else print exactly as before.
+fn install_quiet_panic_hook() {
+    static INSTALLED: OnceLock<()> = OnceLock::new();
+    INSTALLED.get_or_init(|| {
+        let previous = std::panic::take_hook();
+        std::panic::set_hook(Box::new(move |info| {
+            if !SUPPRESS_PANIC_PRINT.with(|flag| flag.get()) {
+                previous(info);
+            }
+        }));
+    });
+}
+
+/// Human-readable message from a caught panic payload; lands in
+/// `failed_files.reason`.
+fn panic_message(payload: &(dyn std::any::Any + Send)) -> String {
+    if let Some(s) = payload.downcast_ref::<&str>() {
+        (*s).to_string()
+    } else if let Some(s) = payload.downcast_ref::<String>() {
+        s.clone()
+    } else {
+        "unknown panic".to_string()
+    }
+}
+
 pub struct PdfExtractor;
 
 impl Extractor for PdfExtractor {
@@ -20,10 +56,15 @@ impl Extractor for PdfExtractor {
     }
 
     fn extract(&self, path: &Path) -> Result<ExtractedContent, ExtractError> {
-        // Text. Catch panics from pdf_extract (some PDFs crash its parser).
+        // Text. Catch panics from pdf_extract (some PDFs crash its parser)
+        // and keep the default hook from spamming stderr about them.
+        install_quiet_panic_hook();
         let path_buf = path.to_path_buf();
-        let text = std::panic::catch_unwind(move || pdf_extract::extract_text(&path_buf))
+        SUPPRESS_PANIC_PRINT.with(|flag| flag.set(true));
-            .map_err(|_| "pdf_extract panicked".to_string())?
+        let result = std::panic::catch_unwind(move || pdf_extract::extract_text(&path_buf));
+        SUPPRESS_PANIC_PRINT.with(|flag| flag.set(false));
+        let text = result
+            .map_err(|panic| format!("pdf_extract panicked: {}", panic_message(&*panic)))?
             .map_err(|e| format!("pdf_extract: {}", e))?;
 
         let mut out = ExtractedContent::with_text(text);
@@ -69,6 +110,20 @@ fn object_to_string(obj: &Object) -> Option<String> {
 mod tests {
     use super::*;
 
+    #[test]
+    fn contained_panics_are_caught_quietly_with_reason() {
+        install_quiet_panic_hook();
+        SUPPRESS_PANIC_PRINT.with(|flag| flag.set(true));
+        let result = std::panic::catch_unwind(|| panic!("synthetic pdf failure"));
+        SUPPRESS_PANIC_PRINT.with(|flag| flag.set(false));
+        let payload = result.expect_err("must panic");
+        assert_eq!(panic_message(&*payload), "synthetic pdf failure");
+        // Panics outside the suppression window keep printing: the flag is
+        // thread-local and cleared, so nothing here can silence other
+        // threads or later tests.
+        assert!(!SUPPRESS_PANIC_PRINT.with(|flag| flag.get()));
+    }
+
     #[test]
     fn supports_pdf_mime() {
         assert!(PdfExtractor.supports("application/pdf"));

crates/quicksearch-core/src/extract/plaintext.rs

@@ -1,10 +1,21 @@
 //! Read the file as UTF-8 text. Handles text/plain, text/x-*, application/json
 //! and most source-code MIMEs.
 
+use std::fs::File;
+use std::io::Read;
 use std::path::Path;
 
 use super::{ExtractError, ExtractedContent, Extractor};
 
+/// Decode bytes that are known to be a complete file. Shared by both entry
+/// points so on-disk and already-in-memory extraction cannot drift apart.
+fn decode(bytes: Vec<u8>, path: &Path) -> Result<ExtractedContent, ExtractError> {
+    match String::from_utf8(bytes) {
+        Ok(text) => Ok(ExtractedContent::with_text(text)),
+        Err(e) => Err(format!("plaintext read {}: {}", path.display(), e.utf8_error())),
+    }
+}
+
 pub struct PlaintextExtractor;
 
 impl Extractor for PlaintextExtractor {
@@ -22,13 +33,66 @@ impl Extractor for PlaintextExtractor {
                 | "application/toml"
                 | "application/yaml"
                 | "application/x-yaml"
+                // `.sql` resolves here rather than to `text/*`, so without it
+                // schema dumps are listed by name but never full-text indexed.
+                | "application/x-sql"
         )
     }
 
+    /// Read the whole file, sized from the handle we just opened.
+    ///
+    /// `std::fs::read_to_string` would cost two extra syscalls here: a
+    /// path-based `statx` to size its buffer, and a second `read` returning 0,
+    /// because "read to EOF" can only observe EOF that way — `read_to_end`
+    /// terminates on `Ok(0)` alone, so a short read does not end it. Sizing
+    /// the buffer ourselves lets the loop finish on `filled == size` and issue
+    /// exactly one `read` for a file that fits.
+    ///
+    /// A file that shrank between the `fstat` and the `read` keeps its prefix
+    /// rather than failing. A file that grew is read up to the size we saw;
+    /// its mtime moved, so the next run reclassifies it as changed and
+    /// re-extracts (see [`crate::file_handling::classify_for_indexing`]).
+    /// Neither case was ever atomic — a concurrent writer can tear a file
+    /// across any read sequence, including `read_to_string`'s.
     fn extract(&self, path: &Path) -> Result<ExtractedContent, ExtractError> {
-        let text = std::fs::read_to_string(path)
+        let mut f = File::open(path)
             .map_err(|e| format!("plaintext read {}: {}", path.display(), e))?;
-        Ok(ExtractedContent::with_text(text))
+        let size = f
+            .metadata()
+            .map_err(|e| format!("plaintext read {}: {}", path.display(), e))?
+            .len() as usize;
+
+        // procfs, sysfs and some FUSE mounts report zero for files that do
+        // have content, so a sized read would store nothing. Only these pay
+        // the read-to-EOF probe — which is what a genuinely empty file cost
+        // before anyway.
+        if size == 0 {
+            let mut buf = Vec::new();
+            f.read_to_end(&mut buf)
+                .map_err(|e| format!("plaintext read {}: {}", path.display(), e))?;
+            return decode(buf, path);
+        }
+
+        let mut buf = vec![0u8; size];
+        let mut filled = 0;
+        while filled < size {
+            match f.read(&mut buf[filled..]) {
+                Ok(0) => break,
+                Ok(n) => filled += n,
+                Err(ref e) if e.kind() == std::io::ErrorKind::Interrupted => {}
+                Err(e) => return Err(format!("plaintext read {}: {}", path.display(), e)),
+            }
+        }
+        buf.truncate(filled);
+        decode(buf, path)
+    }
+
+    fn extract_from_head(
+        &self,
+        path: &Path,
+        head: &[u8],
+    ) -> Option<Result<ExtractedContent, ExtractError>> {
+        Some(decode(head.to_vec(), path))
     }
 }
 
@@ -36,23 +100,115 @@ impl Extractor for PlaintextExtractor {
 mod tests {
     use super::*;
 
-    #[test]
+    fn tmp(tag: &str, body: &[u8]) -> std::path::PathBuf {
-    fn reads_utf8_file() {
         let mut p = std::env::temp_dir();
         p.push(format!(
-            "qs-plaintext-{}-{}.txt",
+            "qs-plaintext-{}-{}-{}.txt",
+            tag,
             std::process::id(),
             std::time::SystemTime::now()
                 .duration_since(std::time::UNIX_EPOCH)
                 .unwrap()
                 .as_nanos()
         ));
-        std::fs::write(&p, "hello world").unwrap();
+        std::fs::write(&p, body).unwrap();
+        p
+    }
+
+    #[test]
+    fn reads_utf8_file() {
+        let p = tmp("basic", b"hello world");
         let c = PlaintextExtractor.extract(&p).unwrap();
         assert_eq!(c.text, "hello world");
         std::fs::remove_file(&p).ok();
     }
 
+    #[test]
+    fn head_extraction_matches_reading_the_file() {
+        let p = tmp("agree", b"shared body with unicode: caf\xc3\xa9 \xe2\x9c\x93");
+        let from_disk = PlaintextExtractor.extract(&p).unwrap();
+        let bytes = std::fs::read(&p).unwrap();
+        let from_head = PlaintextExtractor.extract_from_head(&p, &bytes).unwrap().unwrap();
+        assert_eq!(from_disk.text, from_head.text);
+        assert_eq!(from_disk.properties, from_head.properties);
+        std::fs::remove_file(&p).ok();
+    }
+
+    #[test]
+    fn both_paths_reject_invalid_utf8_and_name_the_file() {
+        let p = tmp("badutf8", &[0x68, 0x69, 0xff, 0xfe]);
+        let disk_err = PlaintextExtractor.extract(&p).unwrap_err();
+        let head_err = PlaintextExtractor
+            .extract_from_head(&p, &[0x68, 0x69, 0xff, 0xfe])
+            .unwrap()
+            .unwrap_err();
+        assert_eq!(disk_err, head_err, "one decode path, one message");
+        assert!(disk_err.contains("badutf8"), "the failure names the file: {}", disk_err);
+        std::fs::remove_file(&p).ok();
+    }
+
+    #[test]
+    fn reads_a_file_larger_than_one_buffer_completely() {
+        // Past any plausible head window, so the read loop has to iterate if
+        // the kernel returns a short read.
+        let body = "abcdefgh".repeat(200 * 1024 / 8);
+        let p = tmp("large", body.as_bytes());
+        let c = PlaintextExtractor.extract(&p).unwrap();
+        assert_eq!(c.text.len(), body.len());
+        assert_eq!(c.text, body);
+        std::fs::remove_file(&p).ok();
+    }
+
+    #[test]
+    fn an_empty_file_extracts_to_empty_text() {
+        let p = tmp("empty", b"");
+        assert_eq!(PlaintextExtractor.extract(&p).unwrap().text, "");
+        assert_eq!(
+            PlaintextExtractor.extract_from_head(&p, &[]).unwrap().unwrap().text,
+            ""
+        );
+        std::fs::remove_file(&p).ok();
+    }
+
+    /// A file whose reported size is a lie in the "there is more than this"
+    /// direction — the shape procfs and sysfs have. Sizing the buffer from
+    /// `st_size` alone would store nothing, so `extract` must fall back to
+    /// reading until EOF.
+    #[test]
+    fn a_file_reporting_zero_size_is_still_read_to_eof() {
+        let p = Path::new("/proc/self/status");
+        if !p.exists() {
+            return; // not Linux; the guard is only reachable there
+        }
+        assert_eq!(
+            std::fs::metadata(p).unwrap().len(),
+            0,
+            "precondition: procfs reports zero size"
+        );
+        let c = PlaintextExtractor.extract(p).unwrap();
+        assert!(
+            c.text.contains("Name:"),
+            "content must survive a zero st_size, got {} bytes",
+            c.text.len()
+        );
+    }
+
+    /// The same lie in the other direction, which the sized read handles by
+    /// keeping whatever was actually there.
+    #[test]
+    fn a_file_that_shrank_after_sizing_keeps_its_prefix() {
+        let p = tmp("shrink", &vec![b'x'; 4096]);
+        let f = File::options().write(true).open(&p).unwrap();
+        // Truncate behind `extract`'s back is not reproducible, so assert the
+        // property directly: a buffer sized larger than the file yields the
+        // file, not an error.
+        f.set_len(10).unwrap();
+        drop(f);
+        let c = PlaintextExtractor.extract(&p).unwrap();
+        assert_eq!(c.text, "xxxxxxxxxx", "a shrunk file reads short, not fatal");
+        std::fs::remove_file(&p).ok();
+    }
+
     #[test]
     fn supports_text_mimes() {
         let e = PlaintextExtractor;

crates/quicksearch-core/src/incremental.rs

@@ -0,0 +1,486 @@
+//! Incremental single-path index updates, driven by watcher events.
+//!
+//! One [`FsEvent`] becomes one (or a few) small transactions: files row,
+//! `documents_text`, and FTS are updated together, so the index is
+//! consistent after every commit. The same filters as the full walk apply
+//! ([`IgnoreSet`], hidden components, `content_extensions`, size caps) —
+//! a watcher event for something the walker would have skipped is a no-op.
+//!
+//! Renames are handled as remove + re-add: they're rare, and rewriting
+//! `path`/`parent` strings plus re-tokenizing the FTS `name` column in
+//! place is more machinery than re-extracting one file.
+//!
+//! Scope note: the watcher only reports paths under the configured roots,
+//! so no root containment check is repeated here.
+
+use std::path::Path;
+
+use rusqlite::{Connection, OptionalExtension};
+
+use crate::config::{content_allowed, Config, IgnoreSet};
+use crate::db::repo;
+use crate::extract::Registry;
+use crate::file_handling::{
+    db_key_for_missing_path, extract_and_store, filtered_walk, UnreadableDirs,
+    prepare_file_record_from_path, store_inline_text,
+};
+use crate::platform::path_has_hidden_component_under;
+use crate::query::translator::like_subtree_pattern;
+use crate::watcher::FsEvent;
+
+/// Apply one filesystem event to the index. Missing files are treated as
+/// no-ops (a Create followed by a quick delete resolves via the Remove
+/// event); unchanged mtimes short-circuit without touching the DB.
+pub fn apply_fs_event(
+    conn: &mut Connection,
+    event: &FsEvent,
+    config: &Config,
+    ignore: &IgnoreSet,
+    registry: &Registry,
+) -> Result<(), String> {
+    match event {
+        FsEvent::Create(p) | FsEvent::Modify(p) => upsert_path(conn, p, config, ignore, registry),
+        FsEvent::Remove(p) => remove_path(conn, p),
+        FsEvent::Rename { from, to } => {
+            remove_path(conn, from)?;
+            upsert_path(conn, to, config, ignore, registry)
+        }
+    }
+}
+
+fn upsert_path(
+    conn: &mut Connection,
+    path: &Path,
+    config: &Config,
+    ignore: &IgnoreSet,
+    registry: &Registry,
+) -> Result<(), String> {
+    if ignore.matches_path(path) {
+        return Ok(());
+    }
+    // Measured from the innermost configured root: the walk never filters the
+    // root it was handed, so a root that is itself hidden (`~/.config/app`, or
+    // anything under `%LOCALAPPDATA%` on Windows) must not be rejected here —
+    // that disagreement is what makes the index churn every cycle.
+    if !config.indexing.include_hidden
+        && path_has_hidden_component_under(path, &config.resolved_indexing_paths())
+    {
+        return Ok(());
+    }
+    let Ok(meta) = std::fs::metadata(path) else {
+        // Already gone again — the pending Remove event handles it.
+        return Ok(());
+    };
+    if meta.is_dir() {
+        // A moved-in tree surfaces as one directory event; walk it with
+        // the same filters as a full run.
+        let Some(root) = path.to_str() else {
+            return Ok(());
+        };
+        let entries: Vec<_> = filtered_walk(
+            root,
+            config.indexing.follow_symlinks,
+            config.indexing.include_hidden,
+            ignore,
+            &UnreadableDirs::default(),
+        )
+        .collect();
+        for entry in entries {
+            upsert_file(conn, entry.path(), config, registry)?;
+        }
+        Ok(())
+    } else {
+        upsert_file(conn, path, config, registry)
+    }
+}
+
+fn upsert_file(
+    conn: &mut Connection,
+    path: &Path,
+    config: &Config,
+    registry: &Registry,
+) -> Result<(), String> {
+    let Some(rec) = prepare_file_record_from_path(path, config, registry) else {
+        return Ok(());
+    };
+
+    let tx = conn
+        .transaction()
+        .map_err(|e| format!("begin incremental tx: {}", e))?;
+
+    let existing: Option<(i64, i64)> = tx
+        .query_row(
+            "SELECT id, mtime FROM files WHERE path = ?1",
+            rusqlite::params![rec.path],
+            |r| Ok((r.get(0)?, r.get(1)?)),
+        )
+        .optional()
+        .map_err(|e| format!("lookup {}: {}", rec.path, e))?;
+
+    let file_id = match existing {
+        Some((_, mtime)) if mtime.max(0) as u64 == rec.mtime => return Ok(()),
+        Some((id, _)) => {
+            repo::update_file_basic(
+                &tx,
+                &rec.path,
+                rec.size,
+                rec.mtime,
+                Some(&rec.hash),
+                rec.mime.as_deref(),
+                rec.ftype,
+            )?;
+            id
+        }
+        None => match repo::insert_file(&tx, &rec.as_new_file())? {
+            Some(id) => id,
+            // Lost a race with another writer on the same path; the row
+            // that won is current enough.
+            None => return Ok(()),
+        },
+    };
+
+    if rec.size > config.processing.maximum_text_file_size
+        || !content_allowed(Path::new(&rec.path), config)
+    {
+        repo::set_content_na(&tx, file_id)?;
+    } else if let Some(text) = rec.inline_text.as_deref() {
+        // Small enough that `prepare_file_record_from_path` already read the
+        // whole file; reopening it here would be the same bytes twice.
+        store_inline_text(&tx, file_id, &rec, text, config)?;
+    } else {
+        extract_and_store(
+            &tx,
+            file_id,
+            &rec.name,
+            &rec.path,
+            rec.mime.as_deref(),
+            registry,
+            config,
+        )?;
+    }
+
+    tx.commit().map_err(|e| format!("commit incremental tx: {}", e))
+}
+
+fn remove_path(conn: &mut Connection, path: &Path) -> Result<(), String> {
+    // The insert side stores a canonicalized path, so the raw event spelling
+    // is not a usable key — but the file is already gone, so `canonicalize`
+    // cannot be called on it directly either.
+    let path_str = db_key_for_missing_path(path);
+    let tx = conn
+        .transaction()
+        .map_err(|e| format!("begin incremental tx: {}", e))?;
+
+    repo::delete_file_by_path(&tx, &path_str)?;
+
+    // Directory removals surface as one event for the directory itself —
+    // sweep everything indexed beneath it.
+    let subtree: Vec<String> = {
+        let mut stmt = tx
+            .prepare("SELECT path FROM files WHERE path LIKE ?1 ESCAPE '\\'")
+            .map_err(|e| e.to_string())?;
+        let rows = stmt
+            .query_map(rusqlite::params![like_subtree_pattern(&path_str)], |r| {
+                r.get::<_, String>(0)
+            })
+            .map_err(|e| e.to_string())?;
+        rows.collect::<Result<Vec<_>, _>>().map_err(|e| e.to_string())?
+    };
+    for p in &subtree {
+        repo::delete_file_by_path(&tx, p)?;
+    }
+
+    tx.commit().map_err(|e| format!("commit incremental tx: {}", e))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::db::open_or_recreate;
+
+    struct Fixture {
+        conn: Connection,
+        dir: std::path::PathBuf,
+        db: std::path::PathBuf,
+        config: Config,
+        ignore: IgnoreSet,
+        registry: Registry,
+    }
+
+    impl Fixture {
+        fn new() -> Fixture {
+            let stamp = format!(
+                "{}-{}",
+                std::process::id(),
+                std::time::SystemTime::now()
+                    .duration_since(std::time::UNIX_EPOCH)
+                    .unwrap()
+                    .as_nanos()
+            );
+            let dir = std::env::temp_dir().join(format!("qs-incr-{}", stamp));
+            std::fs::create_dir_all(&dir).unwrap();
+            let db = std::env::temp_dir().join(format!("qs-incr-{}.sqlite", stamp));
+            let conn = open_or_recreate(db.to_str().unwrap(), "trigram").unwrap();
+            let config = Config::default();
+            let ignore = IgnoreSet::compile(&config.indexing.ignore_patterns).unwrap();
+            Fixture {
+                conn,
+                dir,
+                db,
+                config,
+                ignore,
+                registry: Registry::default_set(),
+            }
+        }
+
+        fn apply(&mut self, event: &FsEvent) {
+            apply_fs_event(&mut self.conn, event, &self.config, &self.ignore, &self.registry)
+                .unwrap();
+        }
+
+        fn write(&self, name: &str, content: &str) -> std::path::PathBuf {
+            let p = self.dir.join(name);
+            std::fs::create_dir_all(p.parent().unwrap()).unwrap();
+            std::fs::write(&p, content).unwrap();
+            p
+        }
+
+        /// The key the index actually stores. Must go through
+        /// `path_to_db_string`, or every lookup here misses the
+        /// `\\?\`-stripped spelling on Windows.
+        fn canonical(&self, p: &Path) -> String {
+            crate::file_handling::path_to_db_string(&p.canonicalize().unwrap())
+        }
+
+        fn row(&self, path: &str) -> Option<(i64, i64, i64)> {
+            self.conn
+                .query_row(
+                    "SELECT id, mtime, content_state FROM files WHERE path = ?1",
+                    rusqlite::params![path],
+                    |r| Ok((r.get(0)?, r.get(1)?, r.get(2)?)),
+                )
+                .optional()
+                .unwrap()
+        }
+
+        fn counts(&self) -> (i64, i64, i64) {
+            let files = self
+                .conn
+                .query_row("SELECT COUNT(*) FROM files", [], |r| r.get(0))
+                .unwrap();
+            let fts = self
+                .conn
+                .query_row("SELECT COUNT(*) FROM searchabletext", [], |r| r.get(0))
+                .unwrap();
+            let texts = self
+                .conn
+                .query_row("SELECT COUNT(*) FROM documents_text", [], |r| r.get(0))
+                .unwrap();
+            (files, fts, texts)
+        }
+
+        fn fts_hits(&self, term: &str) -> i64 {
+            self.conn
+                .query_row(
+                    "SELECT COUNT(*) FROM searchabletext WHERE searchabletext MATCH ?1",
+                    rusqlite::params![format!("\"{}\"", term)],
+                    |r| r.get(0),
+                )
+                .unwrap()
+        }
+    }
+
+    impl Drop for Fixture {
+        fn drop(&mut self) {
+            std::fs::remove_dir_all(&self.dir).ok();
+            std::fs::remove_file(&self.db).ok();
+        }
+    }
+
+    #[test]
+    fn create_indexes_file_and_content() {
+        let mut f = Fixture::new();
+        let p = f.write("hello.txt", "greetings earthling");
+        f.apply(&FsEvent::Create(p.clone()));
+
+        let canonical = f.canonical(&p);
+        let (_, _, content_state) = f.row(&canonical).expect("row exists");
+        assert_eq!(content_state, repo::STATE_DONE);
+        assert_eq!(f.counts(), (1, 1, 1), "files + FTS + text all written");
+        assert_eq!(f.fts_hits("earthling"), 1);
+    }
+
+    #[test]
+    fn modify_with_same_mtime_is_noop_and_changed_mtime_reextracts() {
+        let mut f = Fixture::new();
+        let p = f.write("doc.txt", "first version");
+        f.apply(&FsEvent::Create(p.clone()));
+        let canonical = f.canonical(&p);
+        let (id1, mtime1, _) = f.row(&canonical).unwrap();
+
+        // Same mtime → no-op (id unchanged, no re-extraction).
+        f.apply(&FsEvent::Modify(p.clone()));
+        let (id2, mtime2, _) = f.row(&canonical).unwrap();
+        assert_eq!((id1, mtime1), (id2, mtime2));
+
+        // Bump mtime and content → re-extracted, FTS follows.
+        std::fs::write(&p, "second edition entirely").unwrap();
+        let newer = std::time::SystemTime::now() + std::time::Duration::from_secs(5);
+        let file = std::fs::File::options().write(true).open(&p).unwrap();
+        file.set_modified(newer).unwrap();
+        drop(file);
+        f.apply(&FsEvent::Modify(p.clone()));
+        assert_eq!(f.fts_hits("edition"), 1);
+        assert_eq!(f.fts_hits("version"), 0, "stale tokens removed");
+        assert_eq!(f.counts(), (1, 1, 1), "still exactly one of everything");
+    }
+
+    #[test]
+    fn remove_file_cleans_all_tables() {
+        let mut f = Fixture::new();
+        let p = f.write("bye.txt", "ephemeral text");
+        f.apply(&FsEvent::Create(p.clone()));
+        let canonical = f.canonical(&p);
+        std::fs::remove_file(&p).unwrap();
+        f.apply(&FsEvent::Remove(canonical.clone().into()));
+        assert!(f.row(&canonical).is_none());
+        assert_eq!(f.counts(), (0, 0, 0));
+    }
+
+    #[test]
+    fn directory_create_and_remove_walks_subtree() {
+        let mut f = Fixture::new();
+        f.write("tree/a.txt", "alpha content");
+        f.write("tree/nested/b.txt", "beta content");
+        f.write("tree/.hidden.txt", "should not index");
+        f.write("tree/junk.tmp", "ignored pattern");
+        let tree = f.dir.join("tree");
+        f.apply(&FsEvent::Create(tree.clone()));
+        assert_eq!(f.counts().0, 2, "hidden + ignored excluded");
+
+        let canonical_tree = f.canonical(&tree);
+        std::fs::remove_dir_all(&tree).unwrap();
+        f.apply(&FsEvent::Remove(canonical_tree.into()));
+        assert_eq!(f.counts(), (0, 0, 0), "subtree swept");
+    }
+
+    #[test]
+    fn rename_moves_the_row() {
+        let mut f = Fixture::new();
+        let from = f.write("old-name.txt", "movable feast");
+        f.apply(&FsEvent::Create(from.clone()));
+        let canonical_from = f.canonical(&from);
+
+        let to = f.dir.join("new-name.txt");
+        std::fs::rename(&from, &to).unwrap();
+        f.apply(&FsEvent::Rename {
+            from: canonical_from.clone().into(),
+            to: to.clone(),
+        });
+
+        assert!(f.row(&canonical_from).is_none());
+        let canonical_to = f.canonical(&to);
+        assert!(f.row(&canonical_to).is_some());
+        assert_eq!(f.counts(), (1, 1, 1));
+        assert_eq!(f.fts_hits("feast"), 1);
+    }
+
+    #[test]
+    fn ignored_and_hidden_events_are_noops() {
+        let mut f = Fixture::new();
+        let ignored = f.write("junk.tmp", "x");
+        let hidden = f.write(".secret", "x");
+        f.apply(&FsEvent::Create(ignored));
+        f.apply(&FsEvent::Create(hidden));
+        // Missing file too.
+        f.apply(&FsEvent::Create(f.dir.join("never-existed.txt")));
+        assert_eq!(f.counts(), (0, 0, 0));
+    }
+
+    #[test]
+    fn content_extension_filter_gates_extraction() {
+        let mut f = Fixture::new();
+        f.config.indexing.content_extensions = vec!["md".into()];
+        let txt = f.write("listed-only.txt", "text body here");
+        f.apply(&FsEvent::Create(txt.clone()));
+
+        let canonical = f.canonical(&txt);
+        let (_, _, content_state) = f.row(&canonical).expect("row listed");
+        assert_eq!(
+            content_state,
+            repo::STATE_NA,
+            "filename indexed, content skipped"
+        );
+        assert_eq!(f.counts(), (1, 0, 0));
+    }
+
+    /// A Remove event whose path is spelled differently from the stored key
+    /// must still delete the row. `dir/./f.txt` and `dir/f.txt` are the same
+    /// file; only the canonicalized spelling is in the index.
+    #[test]
+    fn remove_with_a_non_canonical_spelling_still_deletes() {
+        let mut f = Fixture::new();
+        let p = f.write("sub/gone.txt", "vanishing text");
+        f.apply(&FsEvent::Create(p.clone()));
+        assert_eq!(f.counts(), (1, 1, 1));
+
+        std::fs::remove_file(&p).unwrap();
+        // Same file, spelled with a redundant `.` component.
+        let odd = f.dir.join("sub").join(".").join("gone.txt");
+        f.apply(&FsEvent::Remove(odd));
+        assert_eq!(f.counts(), (0, 0, 0), "row removed despite the spelling");
+    }
+
+    /// The subtree sweep must not take siblings whose names merely share a
+    /// string prefix — `tree2` is not inside `tree`.
+    #[test]
+    fn subtree_sweep_spares_prefix_siblings() {
+        let mut f = Fixture::new();
+        f.write("tree/a.txt", "alpha content");
+        f.write("tree2/b.txt", "beta content");
+        let tree = f.dir.join("tree");
+        f.apply(&FsEvent::Create(tree.clone()));
+        f.apply(&FsEvent::Create(f.dir.join("tree2")));
+        assert_eq!(f.counts().0, 2);
+
+        let canonical_tree = f.canonical(&tree);
+        std::fs::remove_dir_all(&tree).unwrap();
+        f.apply(&FsEvent::Remove(canonical_tree.into()));
+
+        assert_eq!(f.counts().0, 1, "only tree/ was swept");
+        let survivor = f.canonical(&f.dir.join("tree2").join("b.txt"));
+        assert!(f.row(&survivor).is_some(), "tree2 untouched");
+    }
+
+    /// A directory whose name contains a LIKE metacharacter must be swept
+    /// literally, not as a wildcard.
+    #[test]
+    fn subtree_sweep_treats_like_metacharacters_literally() {
+        let mut f = Fixture::new();
+        f.write("a_b/inside.txt", "underscore dir");
+        f.write("axb/other.txt", "wildcard bait");
+        f.apply(&FsEvent::Create(f.dir.join("a_b")));
+        f.apply(&FsEvent::Create(f.dir.join("axb")));
+        assert_eq!(f.counts().0, 2);
+
+        let target = f.dir.join("a_b");
+        let canonical = f.canonical(&target);
+        std::fs::remove_dir_all(&target).unwrap();
+        f.apply(&FsEvent::Remove(canonical.into()));
+
+        assert_eq!(f.counts().0, 1, "`_` must not match `x`");
+        let survivor = f.canonical(&f.dir.join("axb").join("other.txt"));
+        assert!(f.row(&survivor).is_some());
+    }
+
+    #[test]
+    fn oversize_files_get_content_na() {
+        let mut f = Fixture::new();
+        f.config.processing.maximum_text_file_size = 4;
+        let p = f.write("big.txt", "way more than four bytes");
+        f.apply(&FsEvent::Create(p.clone()));
+        let canonical = f.canonical(&p);
+        let (_, _, content_state) = f.row(&canonical).unwrap();
+        assert_eq!(content_state, repo::STATE_NA);
+    }
+}

crates/quicksearch-core/src/lib.rs

@@ -1,13 +1,18 @@
 pub mod cli;
 pub mod config;
+pub mod coordinator;
 pub mod db;
 pub mod document_extraction;
 pub mod extract;
 pub mod file_handling;
+pub mod incremental;
 pub mod indexing;
+pub mod log;
 pub mod mime;
+pub mod platform;
 pub mod query;
-pub mod search_sql;
+pub mod search;
 pub mod shutdown;
 pub mod snippet;
+pub mod walk;
 pub mod watcher;

crates/quicksearch-core/src/log.rs

@@ -0,0 +1,241 @@
+//! The process log: every line that would go to the terminal, kept in
+//! memory so a windowed run can show it.
+//!
+//! Launched from a desktop launcher — or on Windows, where the GUI binary is
+//! built for the window subsystem and has no console at all — the process has
+//! nowhere to print. The warnings the walker, indexer and watcher emit are
+//! exactly the ones a user needs when something looks wrong, and they were
+//! going nowhere.
+//!
+//! So background reporting goes through [`log_info!`] and [`log_warn!`]
+//! instead of `println!`/`eprintln!`: each writes the same line to stderr
+//! *and* appends it to a bounded ring the GUI's Logs tab reads. A terminal
+//! run looks exactly as it did; a windowed run gains the tab.
+//!
+//! Command output — search hits from `quicksearch-cli`, usage text, the
+//! errors a command exits with — is not logged. That is a program's answer to
+//! what it was asked, not a background event, and it belongs on stdout.
+
+use std::collections::VecDeque;
+use std::io::Write;
+use std::sync::{LazyLock, Mutex, MutexGuard};
+use std::time::{SystemTime, UNIX_EPOCH};
+
+/// Lines retained before the oldest are dropped.
+///
+/// A run over a tree full of unreadable files can log per file, so this is
+/// bounded rather than complete: the newest few thousand lines are what
+/// diagnosing anything actually needs, and the ring holds the count it threw
+/// away so the tab can say so instead of quietly lying.
+pub const CAPACITY: usize = 5_000;
+
+/// How loud a line is. The GUI colors by this; stderr gets the `Warning:`
+/// prefix that the same messages carried when they were `eprintln!`s.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum Level {
+    Info,
+    Warn,
+}
+
+/// One recorded line.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct LogLine {
+    /// Unix seconds when it was recorded.
+    pub at: u64,
+    pub level: Level,
+    /// Exactly the text written to stderr, prefix included.
+    pub text: String,
+}
+
+/// Record an informational line: `println!`-style formatting.
+#[macro_export]
+macro_rules! log_info {
+    ($($arg:tt)*) => {
+        $crate::log::record($crate::log::Level::Info, ::std::format!($($arg)*))
+    };
+}
+
+/// Record a warning. The stored and printed text gains a `Warning: ` prefix,
+/// so call sites pass the message alone.
+#[macro_export]
+macro_rules! log_warn {
+    ($($arg:tt)*) => {
+        $crate::log::record($crate::log::Level::Warn, ::std::format!($($arg)*))
+    };
+}
+
+/// Write `message` to stderr and to the ring.
+///
+/// Prefer the [`log_info!`] / [`log_warn!`] macros; this is what they call.
+///
+/// A failed stderr write is ignored rather than propagated: `eprintln!`
+/// *panics* when the handle is unwritable, which on a process launched
+/// without stdio would take down whichever background thread happened to
+/// report something. Losing the terminal copy is acceptable — that is
+/// precisely the case where the in-memory copy is the one that matters.
+pub fn record(level: Level, message: String) {
+    let text = match level {
+        Level::Warn => format!("Warning: {}", message),
+        Level::Info => message,
+    };
+    let _ = writeln!(std::io::stderr(), "{}", text);
+    lock().push(level, text);
+}
+
+/// Every retained line, oldest first.
+pub fn snapshot() -> Vec<LogLine> {
+    lock().lines.iter().cloned().collect()
+}
+
+/// How many lines have been recorded since the process started, including
+/// ones since dropped. Only ever grows, so a poll of this is the cheap way
+/// to ask "anything new?" without copying the ring.
+pub fn recorded() -> u64 {
+    lock().recorded
+}
+
+/// How many lines the ring has evicted since the last [`clear`].
+pub fn dropped() -> u64 {
+    lock().dropped
+}
+
+/// Forget every retained line. [`recorded`] keeps counting.
+pub fn clear() {
+    lock().clear();
+}
+
+static LOG: LazyLock<Mutex<Ring>> = LazyLock::new(|| Mutex::new(Ring::new(CAPACITY)));
+
+/// Logging must not turn one panic into a cascade of them: a thread that
+/// died mid-push would otherwise poison the lock and take down every later
+/// logger. The ring is a `VecDeque` of owned strings, so the worst a poisoned
+/// guard can hold is a line that was half-added.
+fn lock() -> MutexGuard<'static, Ring> {
+    LOG.lock().unwrap_or_else(|e| e.into_inner())
+}
+
+/// The bounded line buffer. Split from the global so it can be tested on its
+/// own instance — every other test in the process shares the global one.
+struct Ring {
+    lines: VecDeque<LogLine>,
+    capacity: usize,
+    recorded: u64,
+    dropped: u64,
+}
+
+impl Ring {
+    fn new(capacity: usize) -> Ring {
+        Ring {
+            lines: VecDeque::new(),
+            capacity: capacity.max(1),
+            recorded: 0,
+            dropped: 0,
+        }
+    }
+
+    fn push(&mut self, level: Level, text: String) {
+        while self.lines.len() >= self.capacity {
+            self.lines.pop_front();
+            self.dropped += 1;
+        }
+        self.lines.push_back(LogLine {
+            at: now_unix(),
+            level,
+            text,
+        });
+        self.recorded += 1;
+    }
+
+    fn clear(&mut self) {
+        self.lines.clear();
+        self.dropped = 0;
+    }
+}
+
+fn now_unix() -> u64 {
+    SystemTime::now()
+        .duration_since(UNIX_EPOCH)
+        .map(|d| d.as_secs())
+        .unwrap_or(0)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn texts(ring: &Ring) -> Vec<&str> {
+        ring.lines.iter().map(|l| l.text.as_str()).collect()
+    }
+
+    #[test]
+    fn lines_come_back_oldest_first() {
+        let mut ring = Ring::new(8);
+        ring.push(Level::Info, "one".into());
+        ring.push(Level::Warn, "two".into());
+        assert_eq!(texts(&ring), vec!["one", "two"]);
+        assert_eq!(ring.lines[1].level, Level::Warn);
+        assert_eq!(ring.recorded, 2);
+        assert_eq!(ring.dropped, 0);
+    }
+
+    #[test]
+    fn the_oldest_lines_are_dropped_at_capacity() {
+        let mut ring = Ring::new(3);
+        for i in 0..5 {
+            ring.push(Level::Info, format!("line {}", i));
+        }
+        assert_eq!(
+            texts(&ring),
+            vec!["line 2", "line 3", "line 4"],
+            "only the newest `capacity` lines survive"
+        );
+        assert_eq!(ring.dropped, 2, "and the count of the lost ones is kept");
+        assert_eq!(ring.recorded, 5, "recorded counts everything ever pushed");
+    }
+
+    /// A zero capacity would spin the eviction loop forever on the first
+    /// push; hand-configuring one is not possible today, but the ring should
+    /// not depend on that staying true.
+    #[test]
+    fn a_zero_capacity_still_holds_one_line() {
+        let mut ring = Ring::new(0);
+        ring.push(Level::Info, "kept".into());
+        assert_eq!(texts(&ring), vec!["kept"]);
+    }
+
+    #[test]
+    fn clearing_empties_the_ring_but_not_the_total() {
+        let mut ring = Ring::new(2);
+        for i in 0..4 {
+            ring.push(Level::Info, format!("line {}", i));
+        }
+        ring.clear();
+        assert!(ring.lines.is_empty());
+        assert_eq!(ring.dropped, 0, "dropped counts against what is shown");
+        assert_eq!(ring.recorded, 4, "the running total survives a clear");
+    }
+
+    /// Through the global: the macros must land in the snapshot, and a
+    /// warning must carry the prefix its terminal line has. Written to
+    /// tolerate lines from tests running in parallel in this process.
+    #[test]
+    fn recorded_lines_reach_the_snapshot() {
+        let before = recorded();
+        crate::log_info!("test-marker info {}", 1);
+        crate::log_warn!("test-marker warn {}", 2);
+        assert!(recorded() >= before + 2);
+
+        let lines = snapshot();
+        let info = lines.iter().find(|l| l.text == "test-marker info 1");
+        let warn = lines
+            .iter()
+            .find(|l| l.text == "Warning: test-marker warn 2");
+        assert_eq!(info.map(|l| l.level), Some(Level::Info));
+        assert_eq!(
+            warn.map(|l| l.level),
+            Some(Level::Warn),
+            "a warning is stored with the prefix it printed with"
+        );
+        assert!(info.unwrap().at > 0, "timestamped when recorded");
+    }
+}

crates/quicksearch-core/src/mime.rs

@@ -1,11 +1,16 @@
 //! MIME type guessing and `FileType` bitmask classification.
 //!
 //! Two stages:
-//! 1. [`guess_mime`] infers a MIME type from a path — extension first via
+//! 1. [`guess_mime_from_head`] infers a MIME type — extension first via
 //!    `mime_guess`, falling back to magic-byte sniffing via `infer` for files
 //!    whose extension is missing or ambiguous.
 //! 2. [`mime_to_type`] maps a MIME string to a [`FileType`] bitmask so a single
 //!    file can belong to multiple categories (e.g. a `.docx` is Document|Text).
+//!
+//! The magic bytes are always ones the caller already holds. Indexing reads
+//! the head of every new or changed file to hash it, and those are the same
+//! bytes `infer` wants, so there is no path-based variant that goes back to
+//! disk for them — that was a second open/read/close per undetectable file.
 
 use std::path::Path;
 
@@ -65,26 +70,72 @@ impl std::ops::BitOrAssign for FileType {
     }
 }
 
-/// Guess a MIME type for a path on disk.
+/// Extensions `mime_guess` gets wrong or does not know, and what they really
+/// are.
 ///
-/// Tries extension-based lookup via `mime_guess` first (cheap, no I/O). If
+/// Consulted *before* `mime_guess`, because for these the table is not a
-/// that returns nothing or a generic `application/octet-stream`, and the file
+/// fallback but a correction. Everything here is plain text that would
-/// is readable, falls back to `infer` magic-byte detection (reads a small
+/// otherwise get no content indexing at all:
-/// prefix of the file).
 ///
-/// Returns `None` if no guess can be made.
+/// - `.ps1`/`.psm1`/`.psd1` and `.url` are simply absent from `mime_guess`,
-pub fn guess_mime(path: &Path) -> Option<String> {
+///   and `infer` only knows binary magic, so they end up with no MIME — and
+///   [`crate::extract::Registry`] has no extractor to offer, so the file is
+///   marked "not applicable". PowerShell is the most common script type on a
+///   Windows machine.
+/// - `.bat` maps to `application/x-msdownload`, i.e. an executable. It is a
+///   text file, and the plaintext extractor rightly refuses the executable
+///   type. (`.cmd` already resolves to `text/plain`; it is listed so the pair
+///   cannot drift.)
+///
+/// Platform-neutral on purpose: a `.ps1` copied to a Linux box should classify
+/// the same way.
+const EXTENSION_OVERRIDES: &[(&str, &str)] = &[
+    ("bat", "text/plain"),
+    ("cmd", "text/plain"),
+    ("inf", "text/plain"),
+    ("ps1", "text/plain"),
+    ("psd1", "text/plain"),
+    ("psm1", "text/plain"),
+    ("url", "text/plain"),
+];
+
+/// Look up [`EXTENSION_OVERRIDES`] for `path`. Extension comparison is
+/// ASCII-case-insensitive, which matters more on Windows where `REPORT.BAT` is
+/// as common as the lowercase spelling.
+fn extension_override(path: &Path) -> Option<&'static str> {
+    let ext = path.extension()?.to_str()?.to_ascii_lowercase();
+    EXTENSION_OVERRIDES
+        .iter()
+        .find(|(e, _)| *e == ext)
+        .map(|(_, mime)| *mime)
+}
+
+/// Infer a MIME type from a path plus the file's leading bytes.
+///
+/// Extension first — an override table, then `mime_guess` — and magic bytes
+/// only when those come up empty or say `application/octet-stream`.
+///
+/// `head` is whatever the caller already read; indexing passes the same buffer
+/// it hashes. It bounds magic-byte detection, so a caller that supplies fewer
+/// than 262 bytes (`infer`'s longest signature) can get `None` where a longer
+/// head would have matched. The indexer's `hash_length` defaults to 8 KiB —
+/// exactly what `infer` itself reads from a path — so at default config this
+/// is as good as opening the file, and strictly cheaper.
+///
+/// A `None` result is a real answer, not a "don't know": the content pass
+/// stores it and does not re-derive it (see
+/// [`crate::file_handling::extract_and_store`]).
+pub fn guess_mime_from_head(path: &Path, head: &[u8]) -> Option<String> {
+    if let Some(m) = extension_override(path) {
+        return Some(m.to_string());
+    }
     if let Some(g) = mime_guess::from_path(path).first() {
         let s = g.essence_str();
         if !s.is_empty() && s != "application/octet-stream" {
             return Some(s.to_string());
         }
     }
-    // Magic-byte fallback. `infer::get_from_path` handles errors by returning None.
+    infer::get(head).map(|t| t.mime_type().to_string())
-    if let Ok(Some(t)) = infer::get_from_path(path) {
-        return Some(t.mime_type().to_string());
-    }
-    None
 }
 
 /// Map a MIME string to a [`FileType`] bitmask. Ported from Baloo's
@@ -131,7 +182,13 @@ pub fn mime_to_type(mime: &str) -> FileType {
         | "vnd.oasis.opendocument.presentation" => {
             t |= FileType::DOCUMENT | FileType::PRESENTATION;
         }
-        // Archives
+        // Outlook saved messages and compiled HTML help are documents; both
+        // are ordinary things to find in a Windows home directory.
+        "vnd.ms-outlook" | "vnd.ms-htmlhelp" => {
+            t |= FileType::DOCUMENT;
+        }
+        // Archives. The Windows installer/cabinet formats are containers in
+        // exactly the same sense as the rest of this list.
         "zip"
         | "x-tar"
         | "x-7z-compressed"
@@ -142,7 +199,9 @@ pub fn mime_to_type(mime: &str) -> FileType {
         | "x-bzip2"
         | "x-xz"
         | "vnd.debian.binary-package"
-        | "x-rpm" => {
+        | "x-rpm"
+        | "vnd.ms-cab-compressed"
+        | "x-msi" => {
             t |= FileType::ARCHIVE;
         }
         // application/xml is structured text
@@ -228,11 +287,137 @@ mod tests {
         assert_eq!(FileType::from_name("Weird"), FileType::EMPTY);
     }
 
+    /// Extension resolution happens before magic bytes are consulted, so an
+    /// empty head is enough to exercise it.
     #[test]
     fn guess_mime_by_extension() {
         use std::path::PathBuf;
-        assert_eq!(guess_mime(&PathBuf::from("a.txt")).as_deref(), Some("text/plain"));
+        let by_ext = |n: &str| guess_mime_from_head(&PathBuf::from(n), b"").unwrap_or_default();
-        assert_eq!(guess_mime(&PathBuf::from("a.png")).as_deref(), Some("image/png"));
+        assert_eq!(by_ext("a.txt"), "text/plain");
-        assert_eq!(guess_mime(&PathBuf::from("a.mp3")).as_deref(), Some("audio/mpeg"));
+        assert_eq!(by_ext("a.png"), "image/png");
+        assert_eq!(by_ext("a.mp3"), "audio/mpeg");
+    }
+
+    /// Every override must land on a type the plaintext extractor accepts —
+    /// the point of the table is that these files get their contents indexed.
+    #[test]
+    fn windows_script_types_reach_the_plaintext_extractor() {
+        use crate::extract::{plaintext::PlaintextExtractor, Extractor};
+        use std::path::PathBuf;
+
+        for name in [
+            "deploy.ps1",
+            "Module.psm1",
+            "Module.psd1",
+            "build.bat",
+            "build.cmd",
+            "driver.inf",
+            "bookmark.url",
+        ] {
+            let mime = guess_mime_from_head(&PathBuf::from(name), b"")
+                .unwrap_or_else(|| panic!("{} has no MIME", name));
+            assert!(
+                PlaintextExtractor.supports(&mime),
+                "{} -> {} is not extractable as text",
+                name,
+                mime
+            );
+        }
+    }
+
+    #[test]
+    fn extension_overrides_are_case_insensitive() {
+        use std::path::PathBuf;
+        // Uppercase extensions are ordinary on Windows.
+        assert_eq!(
+            guess_mime_from_head(&PathBuf::from("DEPLOY.PS1"), b"").as_deref(),
+            Some("text/plain")
+        );
+        assert_eq!(
+            guess_mime_from_head(&PathBuf::from("Build.Bat"), b"").as_deref(),
+            Some("text/plain")
+        );
+    }
+
+    /// The override table must win over the file's actual content: a `.ps1`
+    /// holding something `infer` would recognise is still a script.
+    #[test]
+    fn extension_overrides_beat_magic_bytes() {
+        use std::path::PathBuf;
+        assert_eq!(
+            guess_mime_from_head(&PathBuf::from("a.ps1"), b"Write-Host hi").as_deref(),
+            Some("text/plain")
+        );
+        assert_eq!(
+            guess_mime_from_head(&PathBuf::from("a.ps1"), b"%PDF-1.7").as_deref(),
+            Some("text/plain")
+        );
+    }
+
+    #[test]
+    fn sql_dumps_are_extractable() {
+        use crate::extract::{plaintext::PlaintextExtractor, Extractor};
+        use std::path::PathBuf;
+        let mime = guess_mime_from_head(&PathBuf::from("schema.sql"), b"").unwrap();
+        assert!(PlaintextExtractor.supports(&mime), "{}", mime);
+    }
+
+    /// The content pass trusts the MIME the walk stored, including `None`, and
+    /// never reopens the file to second-guess it. That is only sound if a
+    /// `hash_length`-sized head is enough to recognise a format from its magic
+    /// bytes — `infer`'s longest signature is 262 bytes and the default head is
+    /// 8 KiB, so it is by a wide margin. This pins that for extensionless
+    /// files, where magic bytes are the only signal there is.
+    #[test]
+    fn a_default_sized_head_is_enough_for_magic_byte_detection() {
+        use std::path::PathBuf;
+        let head_bytes = crate::config::ProcessingConfig::default().hash_length;
+
+        let samples: &[(&str, &[u8], &str)] = &[
+            ("png", &[0x89, b'P', b'N', b'G', 0x0d, 0x0a, 0x1a, 0x0a], "image/png"),
+            ("gif", b"GIF89a", "image/gif"),
+            ("pdf", b"%PDF-1.7", "application/pdf"),
+            ("zip", &[0x50, 0x4b, 0x03, 0x04], "application/zip"),
+            ("gz", &[0x1f, 0x8b, 0x08], "application/gzip"),
+        ];
+
+        for (tag, magic, expected) in samples {
+            // No extension at all, so nothing but the bytes can answer.
+            let path = PathBuf::from(format!("/tmp/qs-sniff-{}", tag));
+            let mut body = magic.to_vec();
+            body.resize(head_bytes, 0);
+            assert_eq!(
+                guess_mime_from_head(&path, &body).as_deref(),
+                Some(*expected),
+                "{} must be detectable from a default-sized head",
+                tag
+            );
+        }
+    }
+
+    /// The other side of that bound: starve the head below `infer`'s longest
+    /// signature and detection legitimately degrades. Documented behaviour of
+    /// a non-default `hash_length`, not a bug — but it must stay a `None`
+    /// rather than a wrong guess.
+    #[test]
+    fn a_head_shorter_than_the_signature_declines_rather_than_guessing() {
+        use std::path::PathBuf;
+        let path = PathBuf::from("/tmp/qs-sniff-truncated");
+        assert_eq!(guess_mime_from_head(&path, b"").as_deref(), None);
+        assert_eq!(guess_mime_from_head(&path, &[0x89]).as_deref(), None);
+        // Enough bytes, and it resolves.
+        assert_eq!(
+            guess_mime_from_head(&path, &[0x89, b'P', b'N', b'G', 0x0d, 0x0a, 0x1a, 0x0a])
+                .as_deref(),
+            Some("image/png")
+        );
+    }
+
+    #[test]
+    fn windows_container_and_document_types_classify() {
+        assert!(mime_to_type("application/vnd.ms-cab-compressed").contains(FileType::ARCHIVE));
+        assert!(mime_to_type("application/x-msi").contains(FileType::ARCHIVE));
+        assert!(mime_to_type("application/vnd.ms-outlook").contains(FileType::DOCUMENT));
+        assert!(mime_to_type("application/vnd.ms-htmlhelp").contains(FileType::DOCUMENT));
     }
 }

crates/quicksearch-core/src/platform.rs

@@ -0,0 +1,421 @@
+//! The one place `#[cfg]` lives.
+//!
+//! Everything here answers "what does this platform do differently", so no
+//! other module has to ask. Two rules keep it honest:
+//!
+//! - Every function is defined for every target. Callers never wrap a call
+//!   site in `#[cfg]`; if a platform has nothing to do, its arm is the
+//!   trivial one.
+//! - Anything that can be decided from a string rather than a syscall is
+//!   split out and made testable everywhere ([`is_unc_string`],
+//!   [`PATH_COLLATION`]), because the test suite runs on Linux.
+
+use std::ffi::OsString;
+use std::path::{Component, Path, PathBuf};
+
+/// The user's home directory.
+///
+/// On Windows `%USERPROFILE%` is checked **first**. Git Bash and MSYS2 export
+/// `HOME` as a POSIX path (`/c/Users/me`) that no Win32 API can open, and
+/// preferring it would point the config file, the index, and the default
+/// indexing root at a directory that does not exist.
+pub fn home_dir() -> Option<OsString> {
+    #[cfg(windows)]
+    {
+        if let Some(profile) = std::env::var_os("USERPROFILE") {
+            return Some(profile);
+        }
+    }
+    std::env::var_os("HOME").or_else(|| std::env::var_os("USERPROFILE"))
+}
+
+/// Whether a directory entry counts as hidden.
+///
+/// Unix: a leading dot. Windows: a leading dot **or** `FILE_ATTRIBUTE_HIDDEN`
+/// / `FILE_ATTRIBUTE_SYSTEM` — without which `include_hidden = false` hides
+/// nothing on Windows, and `$RECYCLE.BIN`, `System Volume Information`,
+/// `pagefile.sys` and `AppData` all get indexed.
+///
+/// `meta` is a closure because on Unix it is never called: the walkers
+/// deliberately avoid `metadata()`, which would cost an extra `lstat` per
+/// entry and a full round trip on a network share. On Windows the cost is
+/// zero anyway — both `std::fs::DirEntry::metadata` and
+/// `walkdir::DirEntry::metadata` hand back data already cached from
+/// `FindNextFileW`.
+pub fn entry_is_hidden<F>(name: &str, meta: F) -> bool
+where
+    F: FnOnce() -> Option<std::fs::Metadata>,
+{
+    if name.starts_with('.') {
+        return true;
+    }
+    #[cfg(windows)]
+    {
+        use std::os::windows::fs::MetadataExt;
+        use windows_sys::Win32::Storage::FileSystem::{
+            FILE_ATTRIBUTE_HIDDEN, FILE_ATTRIBUTE_SYSTEM,
+        };
+        if let Some(m) = meta() {
+            return m.file_attributes() & (FILE_ATTRIBUTE_HIDDEN | FILE_ATTRIBUTE_SYSTEM) != 0;
+        }
+    }
+    #[cfg(not(windows))]
+    {
+        let _ = meta;
+    }
+    false
+}
+
+/// Whether `path` has a hidden component *below* the root that contains it.
+///
+/// Components at or above a root are exempt, because the walkers exempt their
+/// root too (depth 0 is always kept — users explicitly chose their roots).
+/// The two must agree: if they disagree, a full run indexes a file that the
+/// watcher then refuses to update, and the index churns on every cycle.
+///
+/// That is a latent bug on Unix (`~/.config/app` as a root) and a certainty on
+/// Windows, where `AppData` carries `FILE_ATTRIBUTE_HIDDEN` and
+/// `std::env::temp_dir()` lives underneath it.
+///
+/// `roots` are matched by whole path components, so `/a/bc` is not treated as
+/// living under `/a/b`. A path under no known root is checked in full.
+pub fn path_has_hidden_component_under(path: &Path, roots: &[PathBuf]) -> bool {
+    // Innermost containing root wins: with both `/data` and `/data/.cache`
+    // configured, a file under the latter is only judged below `.cache`.
+    let base = roots
+        .iter()
+        .filter(|r| path.starts_with(r))
+        .max_by_key(|r| r.components().count());
+
+    let (mut current, tail) = match base {
+        Some(root) => match path.strip_prefix(root) {
+            Ok(tail) => (root.clone(), tail),
+            Err(_) => (PathBuf::new(), path),
+        },
+        None => (PathBuf::new(), path),
+    };
+
+    // Rebuild the absolute path as we descend: a bare tail component cannot
+    // be stat'd on its own, and the attribute check needs a real path.
+    for component in tail.components() {
+        current.push(component);
+        if let Component::Normal(name) = component {
+            let name = name.to_string_lossy();
+            if entry_is_hidden(&name, || std::fs::metadata(&current).ok()) {
+                return true;
+            }
+        }
+    }
+    false
+}
+
+/// Whether `s` names a UNC path, in either spelling.
+///
+/// Split out from [`is_network_path`] so the string half is testable on every
+/// platform, and written with explicit parentheses — the precedence of `&&`
+/// against `||` is exactly the kind of thing that silently disables the
+/// network thread pool.
+///
+/// Only *called* on Windows; compiled everywhere so its tests run everywhere,
+/// which is the point of splitting it out.
+#[cfg_attr(not(windows), allow(dead_code))]
+pub(crate) fn is_unc_string(s: &str) -> bool {
+    s.starts_with(r"\\?\UNC\") || (s.starts_with(r"\\") && !s.starts_with(r"\\?\"))
+}
+
+/// Filesystem types whose operations are network round trips.
+#[cfg(target_os = "linux")]
+const NETWORK_FS_TYPES: [&str; 8] = [
+    "cifs", "smb3", "smbfs", "nfs", "nfs4", "afs", "fuse.sshfs", "9p",
+];
+
+/// Whether `path` lives on a network filesystem.
+///
+/// Reads `/proc/mounts` and takes the longest mount point that is a prefix of
+/// `path` — the innermost mount is the one that actually serves it.
+#[cfg(target_os = "linux")]
+pub(crate) fn is_network_path(path: &Path) -> bool {
+    let Ok(mounts) = std::fs::read_to_string("/proc/mounts") else {
+        return false;
+    };
+    let target = std::fs::canonicalize(path).unwrap_or_else(|_| path.to_path_buf());
+
+    let mut best: Option<(usize, bool)> = None;
+    for line in mounts.lines() {
+        let mut fields = line.split_whitespace();
+        let (Some(_dev), Some(point), Some(fstype)) = (fields.next(), fields.next(), fields.next())
+        else {
+            continue;
+        };
+        // `/proc/mounts` octal-escapes spaces and a few other characters.
+        let point = point.replace("\\040", " ");
+        let point = Path::new(&point);
+        if !target.starts_with(point) {
+            continue;
+        }
+        let depth = point.components().count();
+        let is_network = NETWORK_FS_TYPES.contains(&fstype);
+        if best.is_none_or(|(d, _)| depth > d) {
+            best = Some((depth, is_network));
+        }
+    }
+    best.is_some_and(|(_, is_network)| is_network)
+}
+
+/// Whether `path` is served by a network redirector.
+///
+/// UNC needs no syscall. A *mapped drive letter* does: `Z:\` backed by an SMB
+/// share is indistinguishable from a local disk by string inspection, and it
+/// is the common case — asking `GetDriveTypeW` is the only way to tell. Left
+/// undetected it walks with `LOCAL_THREADS` instead of `NETWORK_THREADS`,
+/// which is the exact failure the threading design exists to prevent.
+#[cfg(windows)]
+pub(crate) fn is_network_path(path: &Path) -> bool {
+    use std::os::windows::ffi::OsStrExt;
+    use windows_sys::Win32::Storage::FileSystem::GetDriveTypeW;
+    use windows_sys::Win32::System::WindowsProgramming::DRIVE_REMOTE;
+
+    let s = path.to_string_lossy();
+    if is_unc_string(&s) {
+        return true;
+    }
+
+    // GetDriveTypeW wants a root ("Z:\"), not an arbitrary path.
+    let Some(root) = path.components().next() else {
+        return false;
+    };
+    let Component::Prefix(prefix) = root else {
+        return false;
+    };
+    let mut wide: Vec<u16> = prefix.as_os_str().encode_wide().collect();
+    wide.push(b'\\' as u16);
+    wide.push(0);
+    unsafe { GetDriveTypeW(wide.as_ptr()) == DRIVE_REMOTE }
+}
+
+#[cfg(not(any(target_os = "linux", windows)))]
+pub(crate) fn is_network_path(_path: &Path) -> bool {
+    false
+}
+
+/// Whether the filesystem-notification backend covers a whole tree from one
+/// watch on its root.
+///
+/// `false` (inotify): one watch descriptor covers exactly one directory's
+/// entries, so the caller must walk the tree and register every directory
+/// itself — which is what lets it skip `.git`, `node_modules` and hidden
+/// subtrees instead of spending a scarce descriptor on each.
+///
+/// `true` (`ReadDirectoryChangesW`): one handle covers the subtree, and
+/// directories created later are included automatically. Registering
+/// per-directory here would be actively harmful rather than merely wasteful —
+/// notify allocates a 16 KiB buffer *inline per watch* plus a directory
+/// handle, so a large tree would ask for gigabytes of buffers and tens of
+/// thousands of handles. The pruning moves to the event path instead.
+///
+/// macOS FSEvents is also natively recursive, but it is left on the
+/// per-directory path here because that path works there and is the one under
+/// test.
+pub const WATCH_ROOTS_RECURSIVELY: bool = cfg!(windows);
+
+/// SQLite collation for comparing stored path strings.
+///
+/// Windows filesystems are case-insensitive, and SQLite's `LIKE` already folds
+/// ASCII case by default. A path filter that compares one half with `=` and the
+/// other with `LIKE` would otherwise disagree with itself. `NOCASE` folds ASCII
+/// only, which matches what `LIKE` does — non-ASCII paths stay case-sensitive
+/// on both sides, consistently.
+pub const PATH_COLLATION: &str = if cfg!(windows) { "NOCASE" } else { "BINARY" };
+
+/// How long to keep retrying a delete that fails because something else holds
+/// the file open.
+#[cfg(windows)]
+const REMOVE_RETRY_BUDGET: std::time::Duration = std::time::Duration::from_millis(500);
+
+/// `fs::remove_file`, retried briefly on Windows.
+///
+/// Unix `unlink` succeeds even with the file open, so this is a single call
+/// there. Windows returns a sharing violation while *any* handle is open —
+/// most often an antivirus scanner reading the file microseconds after we
+/// closed it. The retry turns a spurious hard failure into a short pause.
+pub fn remove_file_retrying(path: &Path) -> std::io::Result<()> {
+    #[cfg(not(windows))]
+    {
+        std::fs::remove_file(path)
+    }
+    #[cfg(windows)]
+    {
+        let deadline = std::time::Instant::now() + REMOVE_RETRY_BUDGET;
+        loop {
+            match std::fs::remove_file(path) {
+                Ok(()) => return Ok(()),
+                Err(e) if e.kind() == std::io::ErrorKind::NotFound => return Err(e),
+                Err(e) => {
+                    if std::time::Instant::now() >= deadline {
+                        return Err(e);
+                    }
+                    std::thread::sleep(std::time::Duration::from_millis(25));
+                }
+            }
+        }
+    }
+}
+
+/// Deny read access to `dir`, for tests that exercise the unreadable-directory
+/// guards.
+///
+/// Exposed (hidden) rather than duplicated per test module because
+/// `tests/full_index.rs` is a separate crate and needs it too. Windows uses
+/// `icacls`: a deny ACE binds even the owner until the paired
+/// [`restore_read`] rewrites it, and neither call needs elevation.
+#[doc(hidden)]
+pub fn deny_read(dir: &Path) -> std::io::Result<()> {
+    #[cfg(unix)]
+    {
+        use std::os::unix::fs::PermissionsExt;
+        std::fs::set_permissions(dir, std::fs::Permissions::from_mode(0o000))
+    }
+    #[cfg(windows)]
+    {
+        icacls(dir, &["/deny", &format!("{}:(OI)(CI)(RD)", current_user()?)])
+    }
+    #[cfg(not(any(unix, windows)))]
+    {
+        let _ = dir;
+        Err(std::io::Error::new(
+            std::io::ErrorKind::Unsupported,
+            "deny_read is not supported on this target",
+        ))
+    }
+}
+
+/// Undo [`deny_read`] so the directory can be cleaned up.
+#[doc(hidden)]
+pub fn restore_read(dir: &Path) -> std::io::Result<()> {
+    #[cfg(unix)]
+    {
+        use std::os::unix::fs::PermissionsExt;
+        std::fs::set_permissions(dir, std::fs::Permissions::from_mode(0o755))
+    }
+    #[cfg(windows)]
+    {
+        icacls(dir, &["/remove:d", &current_user()?])
+    }
+    #[cfg(not(any(unix, windows)))]
+    {
+        let _ = dir;
+        Ok(())
+    }
+}
+
+#[cfg(windows)]
+fn current_user() -> std::io::Result<String> {
+    match (std::env::var("USERDOMAIN"), std::env::var("USERNAME")) {
+        (Ok(domain), Ok(user)) => Ok(format!("{}\\{}", domain, user)),
+        (_, Ok(user)) => Ok(user),
+        _ => Err(std::io::Error::other("USERNAME is not set")),
+    }
+}
+
+#[cfg(windows)]
+fn icacls(dir: &Path, args: &[&str]) -> std::io::Result<()> {
+    let out = std::process::Command::new("icacls")
+        .arg(dir)
+        .args(args)
+        .output()?;
+    if out.status.success() {
+        Ok(())
+    } else {
+        Err(std::io::Error::other(format!(
+            "icacls {}: {}",
+            dir.display(),
+            String::from_utf8_lossy(&out.stderr).trim()
+        )))
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn unc_spellings() {
+        assert!(is_unc_string(r"\\server\share"));
+        assert!(is_unc_string(r"\\server\share\dir\file.txt"));
+        assert!(is_unc_string(r"\\?\UNC\server\share"));
+        // A verbatim *drive* path is local, not a share. This is the case the
+        // original `&&`/`||` precedence got wrong.
+        assert!(!is_unc_string(r"\\?\C:\Users\me"));
+        assert!(!is_unc_string(r"C:\Users\me"));
+        assert!(!is_unc_string("/home/me"));
+        assert!(!is_unc_string(""));
+    }
+
+    #[test]
+    fn collation_matches_like_case_folding() {
+        // LIKE folds ASCII case on every platform; the `=` half of a path
+        // filter has to agree with it, which is what this constant is for.
+        assert_eq!(PATH_COLLATION, if cfg!(windows) { "NOCASE" } else { "BINARY" });
+    }
+
+    #[test]
+    fn dotfiles_are_hidden_without_consulting_metadata() {
+        let mut called = false;
+        assert!(entry_is_hidden(".git", || {
+            called = true;
+            None
+        }));
+        assert!(!called, "a dot prefix must short-circuit before any stat");
+    }
+
+    #[test]
+    fn ordinary_names_are_not_hidden() {
+        assert!(!entry_is_hidden("Documents", || None));
+        assert!(!entry_is_hidden("report.txt", || None));
+    }
+
+    #[test]
+    fn hidden_components_are_measured_from_the_innermost_root() {
+        let root = PathBuf::from(format!("{}.config", sep_prefix()));
+        let roots = vec![root.clone()];
+
+        // The root itself is hidden, but it was chosen explicitly — the walk
+        // keeps it, so the watcher must too.
+        assert!(!path_has_hidden_component_under(&root, &roots));
+        assert!(!path_has_hidden_component_under(&root.join("app.conf"), &roots));
+
+        // A dot *below* the root still counts.
+        assert!(path_has_hidden_component_under(
+            &root.join(".secret").join("x"),
+            &roots
+        ));
+    }
+
+    #[test]
+    fn a_path_under_no_root_is_checked_in_full() {
+        let roots = vec![PathBuf::from(format!("{}srv", sep_prefix()))];
+        let stray = PathBuf::from(format!("{}home{}me{}.ssh", sep_prefix(), SEP, SEP));
+        assert!(path_has_hidden_component_under(&stray, &roots));
+    }
+
+    #[test]
+    fn sibling_roots_do_not_capture_each_other() {
+        // `/a/bc` does not live under `/a/b`, so the `.x` below it is judged,
+        // not exempted.
+        let roots = vec![PathBuf::from(format!("{}a{}b", sep_prefix(), SEP))];
+        let other = PathBuf::from(format!("{}a{}bc{}.x", sep_prefix(), SEP, SEP));
+        assert!(path_has_hidden_component_under(&other, &roots));
+    }
+
+    const SEP: char = std::path::MAIN_SEPARATOR;
+
+    /// An absolute-path prefix for the running platform, so these tests read
+    /// the same on both.
+    fn sep_prefix() -> String {
+        if cfg!(windows) {
+            r"C:\".to_string()
+        } else {
+            "/".to_string()
+        }
+    }
+}

crates/quicksearch-core/src/query/lexer.rs

@@ -32,7 +32,32 @@ impl std::fmt::Display for LexError {
 
 impl std::error::Error for LexError {}
 
+/// Whether the `:` at `colon` is the one in a drive letter rather than a
+/// property operator.
+///
+/// True only when the word so far is exactly one ASCII letter *and* a path
+/// separator follows, which is narrow enough to leave `12:30`, `a:b` and
+/// `type:Audio` tokenizing exactly as before.
+fn is_drive_letter_colon(bytes: &[u8], start: usize, colon: usize) -> bool {
+    colon == start + 1
+        && bytes[start].is_ascii_alphabetic()
+        && matches!(bytes.get(colon + 1), Some(b'\\') | Some(b'/'))
+}
+
 pub fn tokenize(input: &str) -> Result<Vec<Token>, LexError> {
+    let (tokens, err) = tokenize_spanned(input);
+    match err {
+        Some(e) => Err(e),
+        None => Ok(tokens.into_iter().map(|(t, _)| t).collect()),
+    }
+}
+
+/// [`tokenize`], but each token carries its byte range in `input`, and a
+/// trailing error (unterminated quote, invalid UTF-8) is returned alongside
+/// the tokens lexed before it instead of discarding them. `Quoted` spans
+/// include both quote characters. This is what the GUI's syntax highlighter
+/// runs on: it must color the intact prefix of a half-typed query.
+pub fn tokenize_spanned(input: &str) -> (Vec<(Token, std::ops::Range<usize>)>, Option<LexError>) {
     let bytes = input.as_bytes();
     let mut i = 0usize;
     let mut out = Vec::new();
@@ -45,62 +70,72 @@ pub fn tokenize(input: &str) -> Result<Vec<Token>, LexError> {
         }
         match b {
             b'(' => {
-                out.push(Token::LParen);
+                out.push((Token::LParen, i..i + 1));
                 i += 1;
             }
             b')' => {
-                out.push(Token::RParen);
+                out.push((Token::RParen, i..i + 1));
                 i += 1;
             }
             b':' => {
-                out.push(Token::Op(Op::Contains));
+                out.push((Token::Op(Op::Contains), i..i + 1));
                 i += 1;
             }
             b'=' => {
-                out.push(Token::Op(Op::Eq));
+                out.push((Token::Op(Op::Eq), i..i + 1));
                 i += 1;
             }
             b'<' => {
                 if bytes.get(i + 1) == Some(&b'=') {
-                    out.push(Token::Op(Op::Le));
+                    out.push((Token::Op(Op::Le), i..i + 2));
                     i += 2;
                 } else {
-                    out.push(Token::Op(Op::Lt));
+                    out.push((Token::Op(Op::Lt), i..i + 1));
                     i += 1;
                 }
             }
             b'>' => {
                 if bytes.get(i + 1) == Some(&b'=') {
-                    out.push(Token::Op(Op::Ge));
+                    out.push((Token::Op(Op::Ge), i..i + 2));
                     i += 2;
                 } else {
-                    out.push(Token::Op(Op::Gt));
+                    out.push((Token::Op(Op::Gt), i..i + 1));
                     i += 1;
                 }
             }
             b'"' => {
                 // Double-quoted phrase. Supports doubled-quote escape `""`.
+                //
+                // Copied as a UTF-8 slice, not byte by byte: `bytes[j] as char`
+                // decodes Latin-1, so `"José"` came back as `José` and matched
+                // nothing. Quoting is also how people write paths containing
+                // spaces, which makes this the more visible of the two.
                 let mut j = i + 1;
                 let mut buf = String::new();
+                let mut segment_start = j;
                 while j < bytes.len() {
                     if bytes[j] == b'"' {
+                        buf.push_str(&input[segment_start..j]);
                         if bytes.get(j + 1) == Some(&b'"') {
                             buf.push('"');
                             j += 2;
+                            segment_start = j;
                             continue;
                         }
                         break;
                     }
-                    buf.push(bytes[j] as char);
                     j += 1;
                 }
                 if j >= bytes.len() {
-                    return Err(LexError {
+                    return (
-                        message: "unterminated quoted phrase".into(),
+                        out,
-                        offset: i,
+                        Some(LexError {
-                    });
+                            message: "unterminated quoted phrase".into(),
+                            offset: i,
+                        }),
+                    );
                 }
-                out.push(Token::Quoted(buf));
+                out.push((Token::Quoted(buf), i..j + 1));
                 i = j + 1;
             }
             _ => {
@@ -108,6 +143,14 @@ pub fn tokenize(input: &str) -> Result<Vec<Token>, LexError> {
                 let start = i;
                 while i < bytes.len() {
                     let c = bytes[i];
+                    if c == b':' && is_drive_letter_colon(bytes, start, i) {
+                        // `C:\Users\me` is one word, not `C` `:` `\Users\me`.
+                        // Without this, `path:C:\Users\me` parses as the
+                        // filter `path` = `C` and silently matches nothing —
+                        // the first thing a Windows user types.
+                        i += 1;
+                        continue;
+                    }
                     if c.is_ascii_whitespace()
                         || matches!(c, b'(' | b')' | b':' | b'=' | b'<' | b'>' | b'"')
                     {
@@ -115,22 +158,29 @@ pub fn tokenize(input: &str) -> Result<Vec<Token>, LexError> {
                     }
                     i += 1;
                 }
-                let word = std::str::from_utf8(&bytes[start..i])
+                let word = match std::str::from_utf8(&bytes[start..i]) {
-                    .map_err(|e| LexError {
+                    Ok(w) => w.to_string(),
-                        message: format!("invalid UTF-8 in word: {}", e),
+                    Err(e) => {
-                        offset: start,
+                        return (
-                    })?
+                            out,
-                    .to_string();
+                            Some(LexError {
-                match word.as_str() {
+                                message: format!("invalid UTF-8 in word: {}", e),
-                    "AND" => out.push(Token::And),
+                                offset: start,
-                    "OR" => out.push(Token::Or),
+                            }),
-                    _ => out.push(Token::Word(word)),
+                        )
-                }
+                    }
+                };
+                let tok = match word.as_str() {
+                    "AND" => Token::And,
+                    "OR" => Token::Or,
+                    _ => Token::Word(word),
+                };
+                out.push((tok, start..i));
             }
         }
     }
 
-    Ok(out)
+    (out, None)
 }
 
 #[cfg(test)]
@@ -211,4 +261,153 @@ mod tests {
     fn unterminated_quote_is_error() {
         assert!(tokenize(r#""oops"#).is_err());
     }
+
+    #[test]
+    fn non_ascii_survives_a_quoted_phrase() {
+        // Byte-wise copying decoded this as Latin-1 (`José`), so the phrase
+        // never matched anything.
+        let t = tokenize(r#""C:\Users\José\docs""#).unwrap();
+        assert_eq!(t, vec![Token::Quoted(r"C:\Users\José\docs".into())]);
+
+        // ...including around a doubled-quote escape, which splits the copy.
+        let t = tokenize(r#""ü""ö""#).unwrap();
+        assert_eq!(t, vec![Token::Quoted(r#"ü"ö"#.into())]);
+    }
+
+    #[test]
+    fn a_drive_letter_colon_does_not_split_the_word() {
+        let t = tokenize(r"path:C:\Users\me\docs").unwrap();
+        assert_eq!(
+            t,
+            vec![
+                Token::Word("path".into()),
+                Token::Op(Op::Contains),
+                Token::Word(r"C:\Users\me\docs".into()),
+            ]
+        );
+
+        // Forward slashes are equally valid on Windows.
+        let t = tokenize("path:D:/data").unwrap();
+        assert_eq!(
+            t,
+            vec![
+                Token::Word("path".into()),
+                Token::Op(Op::Contains),
+                Token::Word("D:/data".into()),
+            ]
+        );
+    }
+
+    #[test]
+    fn spans_cover_every_token_shape() {
+        let input = r#"type:Audio "a b" (x) size<=5"#;
+        let (toks, err) = tokenize_spanned(input);
+        assert!(err.is_none());
+        let spanned: Vec<(&str, Token)> = toks
+            .iter()
+            .map(|(t, r)| (&input[r.clone()], t.clone()))
+            .collect();
+        assert_eq!(
+            spanned,
+            vec![
+                ("type", Token::Word("type".into())),
+                (":", Token::Op(Op::Contains)),
+                ("Audio", Token::Word("Audio".into())),
+                (r#""a b""#, Token::Quoted("a b".into())),
+                ("(", Token::LParen),
+                ("x", Token::Word("x".into())),
+                (")", Token::RParen),
+                ("size", Token::Word("size".into())),
+                ("<=", Token::Op(Op::Le)),
+                ("5", Token::Word("5".into())),
+            ]
+        );
+    }
+
+    #[test]
+    fn quoted_span_includes_quotes_and_escapes() {
+        let input = r#"x "a""b" y"#;
+        let (toks, err) = tokenize_spanned(input);
+        assert!(err.is_none());
+        assert_eq!(toks[1].0, Token::Quoted(r#"a"b"#.into()));
+        assert_eq!(&input[toks[1].1.clone()], r#""a""b""#);
+    }
+
+    #[test]
+    fn spans_are_byte_offsets_around_non_ascii() {
+        let input = r#"José "café" naïve"#;
+        let (toks, err) = tokenize_spanned(input);
+        assert!(err.is_none());
+        assert_eq!(&input[toks[0].1.clone()], "José");
+        assert_eq!(&input[toks[1].1.clone()], r#""café""#);
+        assert_eq!(&input[toks[2].1.clone()], "naïve");
+        assert_eq!(toks[2].1.end, input.len());
+    }
+
+    #[test]
+    fn unterminated_quote_keeps_prefix_tokens() {
+        let (toks, err) = tokenize_spanned(r#"type:Audio "oops"#);
+        let err = err.expect("should report the unterminated quote");
+        assert_eq!(err.offset, 11, "offset of the opening quote");
+        assert_eq!(
+            toks.iter().map(|(t, _)| t.clone()).collect::<Vec<_>>(),
+            vec![
+                Token::Word("type".into()),
+                Token::Op(Op::Contains),
+                Token::Word("Audio".into()),
+            ]
+        );
+    }
+
+    #[test]
+    fn stars_stay_inside_words() {
+        let (toks, err) = tokenize_spanned("foo*bar *");
+        assert!(err.is_none());
+        assert_eq!(
+            toks.iter().map(|(t, _)| t.clone()).collect::<Vec<_>>(),
+            vec![Token::Word("foo*bar".into()), Token::Word("*".into())]
+        );
+    }
+
+    #[test]
+    fn tokenize_matches_span_stripped_tokenize_spanned() {
+        for input in ["a AND (b:c)", r#""q" x>=2"#, "path:C:\\U foo*"] {
+            let plain = tokenize(input).unwrap();
+            let (spanned, err) = tokenize_spanned(input);
+            assert!(err.is_none());
+            let stripped: Vec<Token> = spanned.into_iter().map(|(t, _)| t).collect();
+            assert_eq!(plain, stripped, "input {:?}", input);
+        }
+    }
+
+    #[test]
+    fn the_drive_letter_rule_stays_narrow() {
+        // Two digits before the colon: still a time, not a drive.
+        assert_eq!(
+            tokenize("12:30").unwrap(),
+            vec![
+                Token::Word("12".into()),
+                Token::Op(Op::Contains),
+                Token::Word("30".into()),
+            ]
+        );
+        // One letter, but no separator after the colon.
+        assert_eq!(
+            tokenize("a:b").unwrap(),
+            vec![
+                Token::Word("a".into()),
+                Token::Op(Op::Contains),
+                Token::Word("b".into()),
+            ]
+        );
+        // A separator, but the key is longer than one character.
+        assert_eq!(
+            tokenize("type:/Audio").unwrap(),
+            vec![
+                Token::Word("type".into()),
+                Token::Op(Op::Contains),
+                Token::Word("/Audio".into()),
+            ]
+        );
+    }
 }

crates/quicksearch-core/src/query/mod.rs

@@ -22,7 +22,12 @@
 pub mod ast;
 pub mod lexer;
 pub mod parser;
+pub mod pattern;
+pub mod split;
 pub mod translator;
 
 pub use ast::{Op, Term};
-pub use translator::{parse_and_build, SqlQuery};
+pub use lexer::tokenize_spanned;
+pub use pattern::{RegexQuery, TermPattern};
+pub use split::{split_for_cascade, CascadeQuery};
+pub use translator::{parse_and_build, SqlQuery, TranslateError};

crates/quicksearch-core/src/query/pattern.rs

@@ -0,0 +1,533 @@
+//! Compiled matchers for the cascade term and the `regex:` keyword.
+//!
+//! A term with an unquoted `*` compiles to a small regex (every literal
+//! chunk escaped, stars joined with `.*`), so wildcards and `regex:` share
+//! one linear-time matching engine. Terms without stars stay on the
+//! [`Literal`](TermPattern::Literal) path, which reproduces the cascade's
+//! original `==`/`find`/`count_occurrences` semantics byte for byte.
+//!
+//! `.` never matches `\n`, so a star cannot span lines of extracted text —
+//! a `*` bridging a whole document would produce absurd match ranges and
+//! page-sized snippets. Names and paths contain no newlines, so the rule
+//! only shows up in content matching.
+
+use std::ops::Range;
+
+use regex::{Regex, RegexBuilder};
+
+use super::translator::TranslateError;
+use crate::snippet;
+
+/// Compile-time memory cap for user-supplied and derived regexes. Keeps a
+/// hostile pattern (`a{1000000}{1000}` and friends) from ballooning the
+/// compiled program; matching itself is linear-time by construction.
+const REGEX_SIZE_LIMIT: usize = 4 << 20;
+
+/// Occurrence counts saturate here, matching `count_frac` in the cascade.
+const COUNT_CAP: usize = 1000;
+
+/// The cascade term, compiled once at split time.
+#[derive(Debug, Clone, Default)]
+pub enum TermPattern {
+    /// No matchable content: an empty term, or only stars (`*`, `**`).
+    /// Matches nothing — a bare `*` must not become a scan of everything.
+    #[default]
+    Empty,
+    /// A star-free term. Kept as plain string operations, not a regex, so
+    /// existing queries keep their exact semantics (and speed).
+    Literal(LiteralPattern),
+    /// A term with at least one active wildcard.
+    Wildcard(WildcardPattern),
+}
+
+#[derive(Debug, Clone)]
+pub struct LiteralPattern {
+    text: String,
+    folded: String,
+}
+
+#[derive(Debug, Clone)]
+pub struct WildcardPattern {
+    /// Literal chunks between stars, in order. Never empty, and no chunk
+    /// is empty: edge stars are folded into the compiled regexes, doubled
+    /// stars collapse.
+    segments: Vec<String>,
+    /// Unanchored search regexes with non-greedy joins — leftmost-shortest
+    /// match, which is what a snippet window wants.
+    search_cs: Regex,
+    search_ci: Regex,
+    /// Anchored (`^…$`) regexes for whole-field matching (rank tiers 1/2).
+    anchored_cs: Regex,
+    anchored_ci: Regex,
+}
+
+/// One piece of the search phrase as split out of the token stream.
+/// `glob` is true only for plain unquoted words — quoted phrases and
+/// reassembled `key:value` text keep their stars literal.
+#[derive(Debug, Clone)]
+pub struct TermPart {
+    pub text: String,
+    pub glob: bool,
+}
+
+/// A chunk stream: literal text interleaved with active stars.
+enum Chunk {
+    Lit(String),
+    Star,
+}
+
+impl TermPattern {
+    /// Compile the joined term parts. Parts are joined with a single space,
+    /// exactly like the display term (`parts.join(" ")`).
+    pub fn build(parts: &[TermPart]) -> Result<TermPattern, TranslateError> {
+        let mut chunks: Vec<Chunk> = Vec::new();
+        let push_lit = |chunks: &mut Vec<Chunk>, s: &str| {
+            if s.is_empty() {
+                return;
+            }
+            if let Some(Chunk::Lit(prev)) = chunks.last_mut() {
+                prev.push_str(s);
+            } else {
+                chunks.push(Chunk::Lit(s.to_string()));
+            }
+        };
+        for (idx, part) in parts.iter().enumerate() {
+            if idx > 0 {
+                push_lit(&mut chunks, " ");
+            }
+            if part.glob {
+                let mut first = true;
+                for piece in part.text.split('*') {
+                    if !first && !matches!(chunks.last(), Some(Chunk::Star)) {
+                        chunks.push(Chunk::Star);
+                    }
+                    first = false;
+                    push_lit(&mut chunks, piece);
+                }
+            } else {
+                push_lit(&mut chunks, &part.text);
+            }
+        }
+
+        let leading = matches!(chunks.first(), Some(Chunk::Star));
+        let trailing = chunks.len() > 1 && matches!(chunks.last(), Some(Chunk::Star));
+        let has_star = chunks.iter().any(|c| matches!(c, Chunk::Star));
+        let segments: Vec<String> = chunks
+            .into_iter()
+            .filter_map(|c| match c {
+                Chunk::Lit(s) => Some(s),
+                Chunk::Star => None,
+            })
+            .collect();
+
+        if segments.is_empty() {
+            // "" or stars only.
+            return Ok(TermPattern::Empty);
+        }
+        if !has_star {
+            let text = segments.into_iter().next().unwrap();
+            let folded = text.to_ascii_lowercase();
+            return Ok(TermPattern::Literal(LiteralPattern { text, folded }));
+        }
+
+        let escaped: Vec<String> = segments.iter().map(|s| regex::escape(s)).collect();
+        let compile = |src: &str, ci: bool| -> Result<Regex, TranslateError> {
+            RegexBuilder::new(src)
+                .case_insensitive(ci)
+                .size_limit(REGEX_SIZE_LIMIT)
+                .build()
+                .map_err(|e| TranslateError::BadRegex(e.to_string()))
+        };
+        // Edge stars are dropped from the search form — under substring
+        // semantics a leading/trailing `.*?` adds nothing.
+        let search_src = escaped.join(".*?");
+        // The anchored form keeps them: `*foo` must whole-match "myfoo".
+        let anchored_src = format!(
+            "^{}{}{}$",
+            if leading { ".*" } else { "" },
+            escaped.join(".*"),
+            if trailing { ".*" } else { "" },
+        );
+        Ok(TermPattern::Wildcard(WildcardPattern {
+            search_cs: compile(&search_src, false)?,
+            search_ci: compile(&search_src, true)?,
+            anchored_cs: compile(&anchored_src, false)?,
+            anchored_ci: compile(&anchored_src, true)?,
+            segments,
+        }))
+    }
+
+    pub fn is_empty(&self) -> bool {
+        matches!(self, TermPattern::Empty)
+    }
+
+    pub fn is_wildcard(&self) -> bool {
+        matches!(self, TermPattern::Wildcard(_))
+    }
+
+    /// The literal text, when the term has no wildcard. SQL builders branch
+    /// on this to keep the original single-`LIKE`/phrase-`MATCH` shapes.
+    pub fn literal(&self) -> Option<&str> {
+        match self {
+            TermPattern::Literal(l) => Some(&l.text),
+            _ => None,
+        }
+    }
+
+    /// Literal chunks between wildcards (the whole term when literal).
+    pub fn segments(&self) -> &[String] {
+        match self {
+            TermPattern::Empty => &[],
+            TermPattern::Literal(l) => std::slice::from_ref(&l.text),
+            TermPattern::Wildcard(w) => &w.segments,
+        }
+    }
+
+    /// Characters of literal (non-star) content — the trigram floor and
+    /// path-tier switch count these.
+    pub fn literal_char_count(&self) -> usize {
+        self.segments().iter().map(|s| s.chars().count()).sum()
+    }
+
+    /// Does the pattern match the entire field?
+    pub fn whole_match(&self, text: &str, case_insensitive: bool) -> bool {
+        match self {
+            TermPattern::Empty => false,
+            TermPattern::Literal(l) => {
+                if case_insensitive {
+                    text.eq_ignore_ascii_case(&l.text)
+                } else {
+                    text == l.text
+                }
+            }
+            TermPattern::Wildcard(w) => {
+                let re = if case_insensitive {
+                    &w.anchored_ci
+                } else {
+                    &w.anchored_cs
+                };
+                re.is_match(text)
+            }
+        }
+    }
+
+    /// Leftmost match as a byte range. Literal folding is ASCII-only and
+    /// byte-length preserving, so folded offsets are valid in the original —
+    /// the same invariant the cascade has always relied on.
+    pub fn find_first(&self, text: &str, case_insensitive: bool) -> Option<Range<usize>> {
+        match self {
+            TermPattern::Empty => None,
+            TermPattern::Literal(l) => {
+                let pos = if case_insensitive {
+                    text.to_ascii_lowercase().find(&l.folded)?
+                } else {
+                    text.find(&l.text)?
+                };
+                Some(pos..pos + l.text.len())
+            }
+            TermPattern::Wildcard(w) => {
+                let re = if case_insensitive {
+                    &w.search_ci
+                } else {
+                    &w.search_cs
+                };
+                re.find(text).map(|m| m.range())
+            }
+        }
+    }
+
+    /// Non-overlapping occurrence count, capped at 1000 (the cascade's
+    /// `count_frac` saturates there anyway).
+    pub fn count(&self, text: &str, case_insensitive: bool) -> usize {
+        match self {
+            TermPattern::Empty => 0,
+            TermPattern::Literal(l) => {
+                snippet::count_occurrences(text, &l.text, !case_insensitive)
+            }
+            TermPattern::Wildcard(w) => {
+                let re = if case_insensitive {
+                    &w.search_ci
+                } else {
+                    &w.search_cs
+                };
+                re.find_iter(text).take(COUNT_CAP).count()
+            }
+        }
+    }
+}
+
+/// A compiled `regex:` query. Case-insensitive by default (override with an
+/// inline `(?-i:…)`); `multi_line` makes `^`/`$` per-line over extracted
+/// text, which is what they mean in a search box.
+#[derive(Debug, Clone)]
+pub struct RegexQuery {
+    pub source: String,
+    re: Regex,
+}
+
+impl RegexQuery {
+    pub fn new(source: &str) -> Result<RegexQuery, TranslateError> {
+        let re = RegexBuilder::new(source)
+            .case_insensitive(true)
+            .multi_line(true)
+            .size_limit(REGEX_SIZE_LIMIT)
+            .build()
+            .map_err(|e| TranslateError::BadRegex(e.to_string()))?;
+        // The regex analog of the bare-`*` rule, but loud: the user typed an
+        // explicit keyword, so tell them instead of matching every file.
+        if re.is_match("") {
+            return Err(TranslateError::BadRegex(format!(
+                "'{}' can match the empty string and would match every file",
+                source
+            )));
+        }
+        Ok(RegexQuery {
+            source: source.to_string(),
+            re,
+        })
+    }
+
+    pub fn is_match(&self, text: &str) -> bool {
+        self.re.is_match(text)
+    }
+
+    pub fn find_first(&self, text: &str) -> Option<Range<usize>> {
+        self.re.find(text).map(|m| m.range())
+    }
+
+    /// Non-overlapping occurrence count, capped at 1000.
+    pub fn count(&self, text: &str) -> usize {
+        self.re.find_iter(text).take(COUNT_CAP).count()
+    }
+}
+
+/// Cap a match range at `max_len` bytes (aligned back to a char boundary)
+/// before handing it to `snippet::window_around`. A greedy user regex can
+/// legitimately match megabytes of a minified file; the snippet window
+/// wants the start of that, not all of it.
+pub fn clamp_match_range(text: &str, range: Range<usize>, max_len: usize) -> Range<usize> {
+    let mut end = range.end.min(range.start + max_len);
+    while end > range.start && !text.is_char_boundary(end) {
+        end -= 1;
+    }
+    range.start..end
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn part(text: &str, glob: bool) -> TermPart {
+        TermPart {
+            text: text.into(),
+            glob,
+        }
+    }
+
+    fn wildcard(parts: &[TermPart]) -> WildcardPattern {
+        match TermPattern::build(parts).unwrap() {
+            TermPattern::Wildcard(w) => w,
+            other => panic!("expected wildcard, got {:?}", other),
+        }
+    }
+
+    #[test]
+    fn starless_parts_build_a_literal() {
+        let p = TermPattern::build(&[part("hello", false), part("world", true)]).unwrap();
+        assert_eq!(p.literal(), Some("hello world"));
+        assert!(!p.is_wildcard());
+    }
+
+    #[test]
+    fn empty_and_star_only_terms_match_nothing() {
+        for parts in [
+            vec![],
+            vec![part("", false)],
+            vec![part("*", true)],
+            vec![part("**", true)],
+        ] {
+            let p = TermPattern::build(&parts).unwrap();
+            assert!(p.is_empty(), "{:?}", parts);
+            assert!(!p.whole_match("anything", true));
+            assert!(p.find_first("anything", true).is_none());
+            assert_eq!(p.count("anything", true), 0);
+        }
+    }
+
+    #[test]
+    fn quoted_star_stays_literal() {
+        // A quoted "*" arrives with glob = false.
+        let p = TermPattern::build(&[part("a*b", false)]).unwrap();
+        assert_eq!(p.literal(), Some("a*b"));
+        assert!(p.find_first("xa*by", false).is_some());
+        assert!(p.find_first("aXb", false).is_none());
+    }
+
+    #[test]
+    fn segment_shapes() {
+        // Edge stars vanish into the anchors: `*foo` whole-matches any
+        // suffix `foo`, `foo*` any prefix.
+        let p = TermPattern::build(&[part("*foo", true)]).unwrap();
+        assert_eq!(p.segments(), ["foo"]);
+        assert!(p.whole_match("myfoo", false));
+        assert!(!p.whole_match("foomy", false));
+
+        let p = TermPattern::build(&[part("foo*", true)]).unwrap();
+        assert_eq!(p.segments(), ["foo"]);
+        assert!(p.whole_match("foomy", false));
+        assert!(!p.whole_match("myfoo", false));
+
+        let w = wildcard(&[part("f*o*o", true)]);
+        assert_eq!(w.segments, ["f", "o", "o"]);
+
+        // Doubled stars collapse.
+        let w = wildcard(&[part("f**o", true)]);
+        assert_eq!(w.segments, ["f", "o"]);
+
+        // The implicit joining space is literal content.
+        let w = wildcard(&[part("a*", true), part("b", false)]);
+        assert_eq!(w.segments, ["a", " b"]);
+
+        // `* *` — the joining space between two stars is interior literal
+        // content, so this is a real (if odd) pattern, not Empty.
+        let p = TermPattern::build(&[part("*", true), part("*", true)]).unwrap();
+        assert_eq!(p.segments(), [" "]);
+        assert!(p.whole_match("a b", false));
+        assert!(!p.whole_match("ab", false));
+
+        let p = TermPattern::build(&[part("*x", true), part("y*", true)]).unwrap();
+        assert_eq!(p.segments(), ["x y"]);
+        assert!(p.whole_match("ax yb", false));
+    }
+
+    #[test]
+    fn whole_match_uses_anchors() {
+        let p = TermPattern::build(&[part("*.txt", true)]).unwrap();
+        assert!(p.whole_match("notes.txt", false));
+        assert!(p.whole_match("NOTES.TXT", true));
+        assert!(!p.whole_match("NOTES.TXT", false));
+        assert!(!p.whole_match("notes.txt.bak", false));
+
+        let p = TermPattern::build(&[part("rep*rt", true)]).unwrap();
+        assert!(p.whole_match("report", false));
+        assert!(!p.whole_match("report2024", false));
+    }
+
+    #[test]
+    fn find_first_is_leftmost_shortest() {
+        let p = TermPattern::build(&[part("a*b", true)]).unwrap();
+        // Leftmost-first with a lazy join: starts at 0, ends at the first b.
+        assert_eq!(p.find_first("aXXbYYb", false), Some(0..4));
+        // Case-insensitive variant.
+        assert_eq!(p.find_first("AXXB", true), Some(0..4));
+        assert_eq!(p.find_first("AXXB", false), None);
+    }
+
+    #[test]
+    fn star_does_not_cross_newlines() {
+        let p = TermPattern::build(&[part("foo*bar", true)]).unwrap();
+        assert!(p.find_first("foo bar", false).is_some());
+        assert!(p.find_first("foo\nbar", false).is_none());
+    }
+
+    #[test]
+    fn utf8_boundaries_in_segments_and_haystacks() {
+        let p = TermPattern::build(&[part("café*menu", true)]).unwrap();
+        let hay = "le café du menu";
+        let r = p.find_first(hay, false).unwrap();
+        assert_eq!(&hay[r], "café du menu");
+        // Case-insensitive over non-ASCII haystack: offsets stay valid.
+        let hay = "LE CAFÉ DU MENU";
+        let r = p.find_first(hay, true).unwrap();
+        assert!(hay.is_char_boundary(r.start) && hay.is_char_boundary(r.end));
+    }
+
+    #[test]
+    fn count_is_nonoverlapping_and_capped() {
+        let p = TermPattern::build(&[part("a*b", true)]).unwrap();
+        assert_eq!(p.count("ab ab ab", false), 3);
+        let many = "ab ".repeat(2000);
+        assert_eq!(p.count(&many, false), 1000);
+    }
+
+    #[test]
+    fn literal_parity_with_string_ops() {
+        let p = TermPattern::build(&[part("Report", false)]).unwrap();
+        assert!(p.whole_match("Report", false));
+        assert!(!p.whole_match("report", false));
+        assert!(p.whole_match("report", true));
+        assert_eq!(p.find_first("my Report.pdf", false), Some(3..9));
+        assert_eq!(p.find_first("my report.pdf", true), Some(3..9));
+        assert_eq!(p.find_first("my report.pdf", false), None);
+        assert_eq!(p.count("report Report", false), 1);
+        assert_eq!(p.count("report Report", true), 2);
+    }
+
+    #[test]
+    fn regex_defaults_case_insensitive_with_optout() {
+        let r = RegexQuery::new("foo\\d+").unwrap();
+        assert!(r.is_match("FOO123"));
+        let r = RegexQuery::new("(?-i:FOO)\\d+").unwrap();
+        assert!(r.is_match("FOO1"));
+        assert!(!r.is_match("foo1"));
+    }
+
+    #[test]
+    fn regex_multiline_anchors() {
+        let r = RegexQuery::new("^total:").unwrap();
+        assert!(r.is_match("line one\ntotal: 5"));
+    }
+
+    #[test]
+    fn invalid_regex_is_an_error_not_a_panic() {
+        for src in ["[", "(", "a{2,1}", "(?P<)"] {
+            assert!(
+                matches!(RegexQuery::new(src), Err(TranslateError::BadRegex(_))),
+                "{:?}",
+                src
+            );
+        }
+    }
+
+    #[test]
+    fn empty_matchable_regexes_are_rejected() {
+        for src in ["", ".*", "a*", "x|", "()", "(a+)*"] {
+            assert!(
+                matches!(RegexQuery::new(src), Err(TranslateError::BadRegex(_))),
+                "{:?} should be rejected",
+                src
+            );
+        }
+    }
+
+    #[test]
+    fn hostile_regexes_fail_fast_or_run_linear() {
+        // Deep nesting: rejected cleanly by the parser's nest limit.
+        assert!(RegexQuery::new(&"(".repeat(2000)).is_err());
+        // Huge counted repetition: rejected by size_limit, not compiled.
+        assert!(RegexQuery::new("a{1000000}{1000}").is_err());
+        // Classic backtracking bomb: the linear engine answers immediately
+        // (a backtracker would take exponential time here).
+        let r = RegexQuery::new("(a+)+$").unwrap();
+        let hay = format!("{}b", "a".repeat(10_000));
+        assert!(!r.is_match(&hay));
+    }
+
+    #[test]
+    fn find_and_count_on_regex() {
+        let r = RegexQuery::new("b[aeiou]d").unwrap();
+        let hay = "bad bed bodkin";
+        assert_eq!(r.find_first(hay), Some(0..3));
+        assert_eq!(r.count(hay), 3);
+    }
+
+    #[test]
+    fn clamp_respects_char_boundaries() {
+        let text = "aééééb";
+        let r = clamp_match_range(text, 0..text.len(), 4);
+        assert!(text.is_char_boundary(r.end));
+        assert!(r.end <= 4);
+        // No-op when already short enough.
+        assert_eq!(clamp_match_range(text, 1..3, 100), 1..3);
+    }
+}

crates/quicksearch-core/src/query/split.rs

@@ -0,0 +1,457 @@
+//! Split raw search-box input into (cascade term, structured filters).
+//!
+//! The ranked search cascade has no boolean logic by design: everything
+//! that isn't a recognized `key:value` filter joins the *term* — the single
+//! phrase the cascade matches through its filename/full-text/fuzzy stages.
+//! Recognized filters (`type:`, `modified:`, `path:`, `mime:`, `name:`)
+//! become parameterized SQL fragments ANDed onto every cascade stage.
+//!
+//! Robustness rules for search-as-you-type:
+//! - A lex error (e.g. a half-typed quote) degrades to "whole input is the
+//!   term" — incremental typing must never surface an error.
+//! - An *unrecognized* `key:value` (like `12:30`) is reassembled verbatim
+//!   into the term.
+//! - A recognized key whose value doesn't translate (bad date, unknown type
+//!   name) is a real [`TranslateError`] — the caller shows it inline.
+//! - `AND`/`OR`/parens are not operators here; the words pass through into
+//!   the term, parens are dropped.
+
+use super::ast::Op;
+use super::lexer::{tokenize, Token};
+use super::pattern::{RegexQuery, TermPart, TermPattern};
+use super::translator::{build_filter, is_filter_key, TranslateError};
+
+/// The cascade's parsed input: one term string plus composable filter SQL.
+#[derive(Debug, Clone, Default)]
+pub struct CascadeQuery {
+    /// The ranked search phrase; may be empty when the input was
+    /// filter-only or regex-only.
+    pub term: String,
+    /// `term` compiled for matching: literal, wildcard, or empty.
+    pub pattern: TermPattern,
+    /// A `regex:` filter, matched in Rust against name, path and content —
+    /// never part of the SQL.
+    pub regex: Option<RegexQuery>,
+    /// Zero or more ` AND (...)` fragments with anonymous `?` placeholders
+    /// over alias `f`; appended verbatim to every stage's WHERE clause.
+    pub filter_sql: String,
+    pub filter_params: Vec<rusqlite::types::Value>,
+}
+
+impl CascadeQuery {
+    /// Nothing to rank on: no term pattern and no regex. (Filters alone
+    /// don't drive a search.)
+    pub fn is_empty(&self) -> bool {
+        self.pattern.is_empty() && self.regex.is_none()
+    }
+
+    fn term_only(term: &str) -> CascadeQuery {
+        let term = term.trim().to_string();
+        // Un-lexable input is searched verbatim — stars are not wildcards
+        // here, mirroring the "whole input is the term" degrade rule.
+        let pattern = TermPattern::build(&[TermPart {
+            text: term.clone(),
+            glob: false,
+        }])
+        .expect("literal patterns always compile");
+        CascadeQuery {
+            term,
+            pattern,
+            ..CascadeQuery::default()
+        }
+    }
+}
+
+fn op_str(op: Op) -> &'static str {
+    match op {
+        Op::Contains => ":",
+        Op::Eq => "=",
+        Op::Lt => "<",
+        Op::Le => "<=",
+        Op::Gt => ">",
+        Op::Ge => ">=",
+    }
+}
+
+pub fn split_for_cascade(input: &str) -> Result<CascadeQuery, TranslateError> {
+    // NUL bytes can't occur in filenames or extracted text, but they do
+    // break SQLite text binding and the FTS5 query parser — strip them.
+    let input = input.replace('\0', "");
+    let input = input.as_str();
+    let tokens = match tokenize(input) {
+        Ok(t) => t,
+        // Half-typed input (unterminated quote, invalid word): the whole
+        // raw string is the term. Never an error mid-keystroke.
+        Err(_) => return Ok(CascadeQuery::term_only(input)),
+    };
+
+    let mut out = CascadeQuery::default();
+    let mut term_parts: Vec<TermPart> = Vec::new();
+    let mut i = 0usize;
+    // Only plain unquoted words are wildcard-eligible; everything else
+    // (quoted phrases, demoted AND/OR, reassembled key:value glue) is
+    // searched verbatim.
+    let literal = |text: &str| TermPart {
+        text: text.to_string(),
+        glob: false,
+    };
+
+    while i < tokens.len() {
+        match &tokens[i] {
+            Token::Word(word) => {
+                // Candidate filter: Word(key) Op [Op] (Word|Quoted).
+                // The lexer emits `modified:>=x` as Word Op(:) Op(>=) Word.
+                if let Some(Token::Op(op1)) = tokens.get(i + 1) {
+                    let (op, value_idx) = match tokens.get(i + 2) {
+                        Some(Token::Op(op2)) => (*op2, i + 3),
+                        _ => (*op1, i + 2),
+                    };
+                    let value = match tokens.get(value_idx) {
+                        Some(Token::Word(v)) | Some(Token::Quoted(v)) => Some(v.clone()),
+                        _ => None,
+                    };
+                    if let Some(value) = value {
+                        // Quoted values keep `*` literal; only a bare word's
+                        // stars act as wildcards (`name:` honors this too).
+                        let value_is_word =
+                            matches!(tokens.get(value_idx), Some(Token::Word(_)));
+                        if word.eq_ignore_ascii_case("regex") {
+                            // Not a SQL filter: compiled here, matched in
+                            // Rust against name, path and content.
+                            if op != Op::Contains {
+                                return Err(TranslateError::UnsupportedOp {
+                                    key: word.clone(),
+                                    op,
+                                });
+                            }
+                            if out.regex.is_some() {
+                                return Err(TranslateError::BadRegex(
+                                    "only one regex: per query".into(),
+                                ));
+                            }
+                            out.regex = Some(RegexQuery::new(&value)?);
+                            i = value_idx + 1;
+                            continue;
+                        }
+                        if is_filter_key(word) {
+                            let frag = build_filter(word, op, &value, value_is_word)?;
+                            out.filter_sql.push_str(" AND (");
+                            out.filter_sql.push_str(&frag.sql);
+                            out.filter_sql.push(')');
+                            out.filter_params.extend(frag.params);
+                            i = value_idx + 1;
+                            continue;
+                        }
+                        // Unrecognized key — reassemble verbatim (`12:30`,
+                        // `foo:bar`), gluing any further `:value` chains
+                        // (`foo:bar:baz`).
+                        let mut glued = format!("{}{}{}", word, op_str(op), value);
+                        i = value_idx + 1;
+                        while let Some(Token::Op(next_op)) = tokens.get(i) {
+                            glued.push_str(op_str(*next_op));
+                            i += 1;
+                            if let Some(Token::Word(v)) | Some(Token::Quoted(v)) = tokens.get(i)
+                            {
+                                glued.push_str(v);
+                                i += 1;
+                            }
+                        }
+                        term_parts.push(literal(&glued));
+                        continue;
+                    }
+                    // Key + op with no value yet (mid-typing "type:"):
+                    // pass through as literal text.
+                    term_parts.push(literal(&format!("{}{}", word, op_str(*op1))));
+                    i += 2;
+                    continue;
+                }
+                term_parts.push(TermPart {
+                    text: word.clone(),
+                    glob: word.contains('*'),
+                });
+            }
+            Token::Quoted(q) => term_parts.push(literal(q)),
+            // Not operators in the cascade grammar — plain words.
+            Token::And => term_parts.push(literal("AND")),
+            Token::Or => term_parts.push(literal("OR")),
+            // Grouping has no meaning without boolean logic.
+            Token::LParen | Token::RParen => {}
+            // Dangling operator (e.g. "a > b" typed literally).
+            Token::Op(op) => term_parts.push(literal(op_str(*op))),
+        }
+        i += 1;
+    }
+
+    out.term = term_parts
+        .iter()
+        .map(|p| p.text.as_str())
+        .collect::<Vec<_>>()
+        .join(" ");
+    out.pattern = TermPattern::build(&term_parts)?;
+    Ok(out)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use rusqlite::types::Value;
+
+    #[test]
+    fn plain_words_join_in_order() {
+        let q = split_for_cascade("hello brave world").unwrap();
+        assert_eq!(q.term, "hello brave world");
+        assert!(q.filter_sql.is_empty());
+        assert!(q.filter_params.is_empty());
+    }
+
+    #[test]
+    fn empty_input() {
+        let q = split_for_cascade("").unwrap();
+        assert_eq!(q.term, "");
+        assert!(q.filter_sql.is_empty());
+    }
+
+    #[test]
+    fn each_recognized_filter_key_extracts() {
+        for input in [
+            "type:Audio",
+            "modified:>=2024-01-01",
+            "mtime:<2023-12-01",
+            "path:/home/me",
+            "folder:/home/me",
+            "includefolder:/home/me",
+            "name:report",
+            "filename:report",
+            "mime:application/pdf",
+        ] {
+            let q = split_for_cascade(input).unwrap();
+            assert_eq!(q.term, "", "input {:?} should be pure filter", input);
+            assert!(
+                q.filter_sql.starts_with(" AND ("),
+                "input {:?} → {:?}",
+                input,
+                q.filter_sql
+            );
+            assert!(!q.filter_params.is_empty(), "input {:?}", input);
+        }
+    }
+
+    #[test]
+    fn filters_and_term_mix() {
+        let q = split_for_cascade("type:Document budget report modified:>=2024-01-01").unwrap();
+        assert_eq!(q.term, "budget report");
+        assert_eq!(q.filter_sql.matches(" AND (").count(), 2);
+        assert_eq!(q.filter_params.len(), 2); // type bits + mtime bound
+    }
+
+    #[test]
+    fn unknown_key_stays_literal() {
+        let q = split_for_cascade("meeting 12:30 notes").unwrap();
+        assert_eq!(q.term, "meeting 12:30 notes");
+        assert!(q.filter_sql.is_empty());
+    }
+
+    #[test]
+    fn unknown_key_chain_reassembles() {
+        let q = split_for_cascade("foo:bar:baz").unwrap();
+        assert_eq!(q.term, "foo:bar:baz");
+    }
+
+    #[test]
+    fn half_typed_quote_is_whole_term() {
+        let q = split_for_cascade("\"unclosed phrase").unwrap();
+        assert_eq!(q.term, "\"unclosed phrase");
+        assert!(q.filter_sql.is_empty());
+    }
+
+    #[test]
+    fn half_typed_filter_key_is_literal() {
+        let q = split_for_cascade("type:").unwrap();
+        assert_eq!(q.term, "type:");
+        assert!(q.filter_sql.is_empty());
+    }
+
+    #[test]
+    fn recognized_key_bad_value_errors() {
+        assert!(matches!(
+            split_for_cascade("modified:>=not-a-date"),
+            Err(TranslateError::BadDate(_))
+        ));
+        assert!(matches!(
+            split_for_cascade("type:NotAThing"),
+            Err(TranslateError::UnknownProperty(_))
+        ));
+    }
+
+    #[test]
+    fn and_or_parens_are_plain_text() {
+        let q = split_for_cascade("(alpha AND beta) OR gamma").unwrap();
+        assert_eq!(q.term, "alpha AND beta OR gamma");
+    }
+
+    /// The end-to-end shape of the bug: before the lexer fix this produced
+    /// the filter `parent = "C"` plus a junk term, and returned nothing.
+    #[test]
+    fn a_windows_drive_path_reaches_the_filter_intact() {
+        let q = split_for_cascade(r"path:C:\Users\me\docs").unwrap();
+        assert_eq!(q.term, "", "the whole input is a filter");
+        assert!(matches!(
+            &q.filter_params[0],
+            Value::Text(t) if t == r"C:\Users\me\docs"
+        ), "{:?}", q.filter_params);
+    }
+
+    #[test]
+    fn quoted_value_for_filter() {
+        let q = split_for_cascade("path:\"/home/me/My Documents\"").unwrap();
+        assert_eq!(q.term, "");
+        assert_eq!(
+            q.filter_params[0],
+            Value::Text("/home/me/My Documents".into())
+        );
+    }
+
+    #[test]
+    fn quoted_phrase_joins_term() {
+        let q = split_for_cascade("\"exact phrase\" extra").unwrap();
+        assert_eq!(q.term, "exact phrase extra");
+    }
+
+    #[test]
+    fn injection_shapes_stay_bound() {
+        // Everything lands either in the term (never interpolated into
+        // SQL by the cascade — bound as parameters there too) or in
+        // filter_params. filter_sql must never contain user text.
+        let q = split_for_cascade("mime:application/x-foo'; DROP TABLE files; --").unwrap();
+        assert!(!q.filter_sql.contains("DROP"), "{}", q.filter_sql);
+        // The value went into params (term got the trailing junk words).
+        assert!(matches!(&q.filter_params[0], Value::Text(t) if t.contains("x-foo'")));
+
+        let q = split_for_cascade("name:%_\\").unwrap();
+        // LIKE-escaped inside the bound param, not the SQL.
+        assert!(matches!(&q.filter_params[0], Value::Text(t) if t == "%\\%\\_\\\\%"));
+    }
+
+    #[test]
+    fn unquoted_star_builds_a_wildcard_pattern() {
+        let q = split_for_cascade("foo*").unwrap();
+        assert_eq!(q.term, "foo*");
+        assert!(q.pattern.is_wildcard());
+        assert!(q.pattern.whole_match("foobar", false));
+    }
+
+    #[test]
+    fn quoted_star_stays_literal() {
+        let q = split_for_cascade("\"foo*\"").unwrap();
+        assert_eq!(q.pattern.literal(), Some("foo*"));
+    }
+
+    #[test]
+    fn bare_star_matches_nothing() {
+        for input in ["*", "**", "* *"] {
+            let q = split_for_cascade(input).unwrap();
+            assert!(
+                q.pattern.find_first("anything", true).is_none(),
+                "{:?}",
+                input
+            );
+        }
+        // "* *" has an interior literal space; plain stars are Empty.
+        assert!(split_for_cascade("*").unwrap().pattern.is_empty());
+        assert!(split_for_cascade("*").unwrap().is_empty());
+    }
+
+    #[test]
+    fn glued_unknown_keys_keep_stars_literal() {
+        let q = split_for_cascade("foo:ba*r").unwrap();
+        assert_eq!(q.pattern.literal(), Some("foo:ba*r"));
+    }
+
+    #[test]
+    fn lex_error_degrade_keeps_stars_literal() {
+        let q = split_for_cascade("re*port \"unclosed").unwrap();
+        assert_eq!(q.pattern.literal(), Some("re*port \"unclosed"));
+    }
+
+    #[test]
+    fn regex_keyword_compiles_out_of_band() {
+        let q = split_for_cascade("regex:foo\\d+").unwrap();
+        assert_eq!(q.term, "");
+        assert!(q.pattern.is_empty());
+        assert!(!q.is_empty(), "a regex-only query still searches");
+        let re = q.regex.unwrap();
+        assert!(re.is_match("FOO12"));
+        assert!(q.filter_sql.is_empty(), "regex is not a SQL filter");
+    }
+
+    #[test]
+    fn regex_value_may_be_quoted_and_key_is_case_insensitive() {
+        let q = split_for_cascade("REGEX:\"foo (bar|baz)\"").unwrap();
+        assert!(q.regex.unwrap().is_match("foo bar"));
+    }
+
+    #[test]
+    fn regex_mixes_with_filters_and_term() {
+        let q = split_for_cascade("regex:\\d+ type:Text budget").unwrap();
+        assert_eq!(q.term, "budget");
+        assert!(q.regex.is_some());
+        assert_eq!(q.filter_sql.matches(" AND (").count(), 1);
+    }
+
+    #[test]
+    fn regex_error_shapes() {
+        assert!(matches!(
+            split_for_cascade("regex:["),
+            Err(TranslateError::BadRegex(_))
+        ));
+        assert!(matches!(
+            split_for_cascade("regex=x"),
+            Err(TranslateError::UnsupportedOp { .. })
+        ));
+        assert!(matches!(
+            split_for_cascade("regex:a regex:b"),
+            Err(TranslateError::BadRegex(_))
+        ));
+        // Empty-matchable patterns are rejected loudly.
+        assert!(matches!(
+            split_for_cascade("regex:.*"),
+            Err(TranslateError::BadRegex(_))
+        ));
+    }
+
+    #[test]
+    fn dangling_regex_key_is_literal_text() {
+        let q = split_for_cascade("regex:").unwrap();
+        assert_eq!(q.term, "regex:");
+        assert!(q.regex.is_none());
+    }
+
+    #[test]
+    fn name_filter_star_becomes_like_wildcard() {
+        let q = split_for_cascade("name:foo*bar").unwrap();
+        assert!(matches!(&q.filter_params[0], Value::Text(t) if t == "%foo%bar%"));
+
+        // Quoted value: star stays a literal character.
+        let q = split_for_cascade("name:\"fo*o\"").unwrap();
+        assert!(matches!(&q.filter_params[0], Value::Text(t) if t == "%fo*o%"));
+
+        // User LIKE metacharacters stay escaped even in glob values.
+        let q = split_for_cascade("name:%*_").unwrap();
+        assert!(matches!(&q.filter_params[0], Value::Text(t) if t == "%\\%%\\_%"));
+
+        // path: values never glob.
+        let q = split_for_cascade("path:/da*ta").unwrap();
+        assert!(matches!(&q.filter_params[0], Value::Text(t) if t == "/da*ta"));
+    }
+
+    #[test]
+    fn nul_bytes_stripped_and_long_terms_pass_through() {
+        // NULs would break SQLite binding / FTS5 parsing downstream.
+        let q = split_for_cascade("abc\0def").unwrap();
+        assert_eq!(q.term, "abcdef");
+
+        let long = "x".repeat(10_240);
+        let q = split_for_cascade(&long).unwrap();
+        assert_eq!(q.term.len(), 10_240);
+    }
+}

crates/quicksearch-core/src/query/translator.rs

@@ -35,6 +35,7 @@ pub enum TranslateError {
     Parse(ParseError),
     UnknownProperty(String),
     BadDate(String),
+    BadRegex(String),
     UnsupportedOp {
         key: String,
         op: Op,
@@ -47,6 +48,7 @@ impl std::fmt::Display for TranslateError {
             TranslateError::Parse(e) => write!(f, "{}", e),
             TranslateError::UnknownProperty(k) => write!(f, "unknown property '{}'", k),
             TranslateError::BadDate(s) => write!(f, "bad date '{}'", s),
+            TranslateError::BadRegex(s) => write!(f, "regex error: {}", s),
             TranslateError::UnsupportedOp { key, op } => {
                 write!(f, "operator {:?} is not supported for property '{}'", op, key)
             }
@@ -172,8 +174,10 @@ fn fts_expr(parts: &[FtsFragment]) -> String {
 }
 
 /// Escape a phrase for FTS5 MATCH. FTS5 itself uses doubled quotes for
-/// literal quotes inside a quoted phrase.
+/// literal quotes inside a quoted phrase; wrapping in quotes renders all
-fn quote_phrase(s: &str) -> String {
+/// other MATCH metacharacters (`( ) * :` etc.) inert. Injection-safe by
+/// construction.
+pub fn quote_phrase(s: &str) -> String {
     let mut buf = String::with_capacity(s.len() + 2);
     buf.push('"');
     for c in s.chars() {
@@ -243,119 +247,214 @@ impl Builder {
         }
     }
 
+    /// Delegate to the shared [`build_filter`] fragment builder, then
+    /// convert its anonymous `?` placeholders to this builder's numbered
+    /// scheme (params[0] is reserved for the FTS MATCH when one exists;
+    /// `build` shifts numbers afterwards).
     fn translate_property(
         &mut self,
         key: &str,
         op: Op,
         value: &str,
     ) -> Result<String, TranslateError> {
-        let lower_key = key.to_ascii_lowercase();
+        let frag = build_filter(key, op, value, false)?;
-        match lower_key.as_str() {
+        let mut params = frag.params.into_iter();
-            "type" => self.prop_type(op, value, key),
+        let mut out = String::with_capacity(frag.sql.len() + 8);
-            "modified" | "mtime" => self.prop_mtime(op, value, key),
+        for c in frag.sql.chars() {
-            "path" | "folder" | "includefolder" => self.prop_path(op, value, key),
+            if c == '?' {
-            "name" | "filename" => self.prop_name(op, value, key),
+                let v = params
-            "mime" => self.prop_mime(op, value, key),
+                    .next()
-            _ => Err(TranslateError::UnknownProperty(key.to_string())),
+                    .expect("FilterFragment placeholder/param counts match");
+                self.all_params.push(v);
+                out.push('?');
+                out.push_str(&self.all_params.len().to_string());
+            } else {
+                out.push(c);
+            }
         }
+        Ok(out)
     }
+}
 
-    fn prop_type(&mut self, op: Op, value: &str, key: &str) -> Result<String, TranslateError> {
+/// A structured-filter fragment over table alias `f`: SQL with anonymous
+/// `?` placeholders plus the values they bind. Anonymous placeholders
+/// compose by simple appending — the search cascade tacks fragments onto
+/// every stage's WHERE clause with `AND (...)`.
+#[derive(Debug, Clone)]
+pub struct FilterFragment {
+    pub sql: String,
+    pub params: Vec<rusqlite::types::Value>,
+}
+
+/// Whether `key` is a recognized structured-filter property.
+pub fn is_filter_key(key: &str) -> bool {
+    matches!(
+        key.to_ascii_lowercase().as_str(),
+        "type" | "modified" | "mtime" | "path" | "folder" | "includefolder" | "name"
+            | "filename" | "mime"
+    )
+}
+
+/// Translate one `key op value` filter into a [`FilterFragment`]. The
+/// single source of filter semantics, shared by the legacy numbered
+/// [`build`] path and the cascade's [`super::split`].
+///
+/// `glob` marks a value whose unquoted `*` should act as a wildcard — only
+/// `name:`/`filename:` honor it; every other key treats the star literally.
+/// The caller decides, because only the tokenizer knows whether the value
+/// was quoted (quoted stars are always literal).
+pub fn build_filter(
+    key: &str,
+    op: Op,
+    value: &str,
+    glob: bool,
+) -> Result<FilterFragment, TranslateError> {
+    use rusqlite::types::Value;
+    let frag = |sql: &str, params: Vec<Value>| FilterFragment {
+        sql: sql.to_string(),
+        params,
+    };
+    let eq_like_only = |op: Op| -> Result<(), TranslateError> {
         if op != Op::Contains && op != Op::Eq {
             return Err(TranslateError::UnsupportedOp {
                 key: key.into(),
                 op,
             });
         }
-        let bits = FileType::from_name(value).bits() as i64;
+        Ok(())
-        if bits == 0 {
+    };
-            return Err(TranslateError::UnknownProperty(format!(
+
-                "type name '{}'",
+    match key.to_ascii_lowercase().as_str() {
+        "type" => {
+            eq_like_only(op)?;
+            let bits = FileType::from_name(value).bits() as i64;
+            if bits == 0 {
+                return Err(TranslateError::UnknownProperty(format!(
+                    "type name '{}'",
+                    value
+                )));
+            }
+            Ok(frag("(f.type & ?) != 0", vec![Value::Integer(bits)]))
+        }
+        "modified" | "mtime" => {
+            let unix =
+                parse_date_to_unix(value).ok_or_else(|| TranslateError::BadDate(value.into()))?;
+            // `modified:=2024-01-01` matches the whole day, not the second.
+            if op == Op::Eq || op == Op::Contains {
+                return Ok(frag(
+                    "(f.mtime >= ? AND f.mtime < ?)",
+                    vec![Value::Integer(unix), Value::Integer(unix + 86_400)],
+                ));
+            }
+            let sql_op = match op {
+                Op::Lt => "<",
+                Op::Le => "<=",
+                Op::Gt => ">",
+                Op::Ge => ">=",
+                Op::Contains | Op::Eq => unreachable!(),
+            };
+            Ok(frag(
+                &format!("f.mtime {} ?", sql_op),
+                vec![Value::Integer(unix)],
+            ))
+        }
+        "path" | "folder" | "includefolder" => {
+            eq_like_only(op)?;
+            let base = normalize_folder_value(value);
+            if base.is_empty() {
+                // "everything". On Unix the old `parent = '/' OR parent LIKE
+                // '/%'` happened to match every absolute path; Windows has no
+                // single root, so say it directly rather than by accident.
+                return Ok(frag("1=1", Vec::new()));
+            }
+            // The `=` half needs the collation spelled out: `LIKE` folds ASCII
+            // case on its own, so without this the two halves of the same
+            // filter disagree about `C:\Users` versus `c:\users`.
+            Ok(frag(
+                &format!(
+                    "(f.parent = ? COLLATE {} OR f.parent LIKE ? ESCAPE '\\')",
+                    crate::platform::PATH_COLLATION
+                ),
+                vec![
+                    Value::Text(base.clone()),
+                    Value::Text(like_subtree_pattern(&base)),
+                ],
+            ))
+        }
+        "name" | "filename" => {
+            if op != Op::Contains {
+                return Err(TranslateError::UnsupportedOp {
+                    key: key.into(),
+                    op,
+                });
+            }
+            // With `glob`, each `*` becomes an unescaped `%`; the pieces
+            // around it still get `%`/`_`/`\` escaped so user metacharacters
+            // stay literal either way.
+            let pattern = if glob && value.contains('*') {
                 value
-            )));
+                    .split('*')
+                    .map(escape_like)
+                    .collect::<Vec<_>>()
+                    .join("%")
+            } else {
+                escape_like(value)
+            };
+            Ok(frag(
+                "f.name LIKE ? ESCAPE '\\'",
+                vec![Value::Text(format!("%{}%", pattern))],
+            ))
         }
-        self.all_params
+        "mime" => {
-            .push(rusqlite::types::Value::Integer(bits));
+            eq_like_only(op)?;
-        Ok(format!("(f.type & ?{}) != 0", self.param_placeholder_idx()))
+            Ok(frag("f.mime = ?", vec![Value::Text(value.into())]))
-    }
-
-    fn prop_mtime(&mut self, op: Op, value: &str, key: &str) -> Result<String, TranslateError> {
-        let unix = parse_date_to_unix(value).ok_or_else(|| TranslateError::BadDate(value.into()))?;
-        let col = "f.mtime";
-        let sql_op = match op {
-            Op::Contains | Op::Eq => "=",
-            Op::Lt => "<",
-            Op::Le => "<=",
-            Op::Gt => ">",
-            Op::Ge => ">=",
-        };
-        // `modified:=2024-01-01` should match the whole day, not the second.
-        if op == Op::Eq || op == Op::Contains {
-            let start = unix;
-            let end = unix + 86_400;
-            self.all_params.push(rusqlite::types::Value::Integer(start));
-            let i = self.param_placeholder_idx();
-            self.all_params.push(rusqlite::types::Value::Integer(end));
-            let j = self.param_placeholder_idx();
-            return Ok(format!("({} >= ?{} AND {} < ?{})", col, i, col, j));
         }
-        self.all_params.push(rusqlite::types::Value::Integer(unix));
+        _ => Err(TranslateError::UnknownProperty(key.to_string())),
-        let i = self.param_placeholder_idx();
-        let _ = key;
-        Ok(format!("{} {} ?{}", col, sql_op, i))
     }
+}
 
-    fn prop_path(&mut self, op: Op, value: &str, key: &str) -> Result<String, TranslateError> {
+/// Escape `%`, `_` and `\` for use inside a `LIKE ... ESCAPE '\'` pattern.
-        if op != Op::Contains && op != Op::Eq {
+pub fn escape_like(s: &str) -> String {
-            return Err(TranslateError::UnsupportedOp {
+    let mut out = String::with_capacity(s.len());
-                key: key.into(),
+    for c in s.chars() {
-                op,
+        if matches!(c, '%' | '_' | '\\') {
-            });
+            out.push('\\');
         }
-        self.all_params
+        out.push(c);
-            .push(rusqlite::types::Value::Text(value.into()));
-        let i = self.param_placeholder_idx();
-        self.all_params
-            .push(rusqlite::types::Value::Text(format!("{}/%", value.trim_end_matches('/'))));
-        let j = self.param_placeholder_idx();
-        Ok(format!("(f.parent = ?{} OR f.parent LIKE ?{})", i, j))
     }
+    out
+}
 
-    fn prop_name(&mut self, op: Op, value: &str, key: &str) -> Result<String, TranslateError> {
+/// Tidy a user-supplied folder value into the spelling `files.parent` stores.
-        if op != Op::Contains {
+///
-            return Err(TranslateError::UnsupportedOp {
+/// Trailing separators are how people naturally write directories, and either
-                key: key.into(),
+/// separator may show up on Windows. A bare drive (`C:`) is *not* a path — the
-                op,
+/// stored parent is `C:\` — so the separator goes back on.
-            });
+fn normalize_folder_value(value: &str) -> String {
-        }
+    let base = value.trim().trim_end_matches(['/', '\\']);
-        self.all_params
+    if base.len() == 2 && base.ends_with(':') && base.starts_with(|c: char| c.is_ascii_alphabetic())
-            .push(rusqlite::types::Value::Text(format!("%{}%", value)));
+    {
-        let i = self.param_placeholder_idx();
+        return format!("{}{}", base, std::path::MAIN_SEPARATOR);
-        Ok(format!("f.name LIKE ?{}", i))
     }
+    base.to_string()
+}
 
-    fn prop_mime(&mut self, op: Op, value: &str, key: &str) -> Result<String, TranslateError> {
+/// A `LIKE ... ESCAPE '\'` pattern matching every path strictly beneath `dir`.
-        if op != Op::Contains && op != Op::Eq {
+///
-            return Err(TranslateError::UnsupportedOp {
+/// The separator is escaped along with the base, because on Windows the
-                key: key.into(),
+/// separator *is* the escape character — a hand-written `format!("{}/%", dir)`
-                op,
+/// is wrong twice over there: wrong separator, and the one it emits would be
-            });
+/// swallowed as an escape.
-        }
+///
-        self.all_params
+/// SQLite's `patternCompare` takes the character after the escape literally
-            .push(rusqlite::types::Value::Text(value.into()));
+/// whatever it is, so a doubled `\` is well defined here; the folklore that an
-        let i = self.param_placeholder_idx();
+/// escape must be followed by `%`, `_` or itself does not apply.
-        Ok(format!("f.mime = ?{}", i))
+pub fn like_subtree_pattern(dir: &str) -> String {
-    }
+    format!(
-
+        "{}{}%",
-    fn param_placeholder_idx(&self) -> usize {
+        escape_like(dir.trim_end_matches(['/', '\\'])),
-        // params[0] is reserved for the FTS MATCH if one is built; structured
+        escape_like(std::path::MAIN_SEPARATOR_STR)
-        // params start at index 2 in that case (1-based). We track it by
+    )
-        // calling this *after* pushing the value; result is `len` so the SQL
-        // says `?<len>` which matches the 1-based positional binding rusqlite
-        // uses for `?N` placeholders. When an FTS match is prepended at
-        // `build`, each index shifts by 1 implicitly.
-        self.all_params.len()
-    }
 }
 
 fn join_with(sep: &str, pieces: &[String]) -> String {
@@ -497,6 +596,101 @@ mod tests {
         let q = build_q("path:/home/me/docs");
         assert!(q.sql.contains("f.parent = ?"));
         assert!(q.sql.contains("f.parent LIKE ?"));
+        // The LIKE half must be escaped and declare its escape character;
+        // without the clause a Windows separator would be eaten as an escape.
+        assert!(q.sql.contains("ESCAPE '\\'"), "{}", q.sql);
+    }
+
+    /// The subtree pattern is the one place the separator and the LIKE escape
+    /// character collide (on Windows they are the same byte), so it is checked
+    /// against real SQLite rather than by string comparison.
+    #[test]
+    fn like_subtree_pattern_matches_only_the_subtree() {
+        use std::path::MAIN_SEPARATOR as SEP;
+
+        let conn = rusqlite::Connection::open_in_memory().unwrap();
+        conn.execute("CREATE TABLE files (parent TEXT NOT NULL)", [])
+            .unwrap();
+
+        let base = format!("{}a{}b", root_prefix(), SEP);
+        let rows = [
+            format!("{}{}sub", base, SEP),          // inside
+            format!("{}{}sub{}deep", base, SEP, SEP), // deeper
+            base.clone(),                            // the folder itself
+            format!("{}a{}bc", root_prefix(), SEP),  // prefix sibling: outside
+            format!("{}a", root_prefix()),           // parent: outside
+        ];
+        for r in &rows {
+            conn.execute("INSERT INTO files (parent) VALUES (?1)", [r])
+                .unwrap();
+        }
+
+        let matched: i64 = conn
+            .query_row(
+                "SELECT COUNT(*) FROM files WHERE parent LIKE ?1 ESCAPE '\\'",
+                [like_subtree_pattern(&base)],
+                |r| r.get(0),
+            )
+            .unwrap();
+        assert_eq!(matched, 2, "only the two rows strictly beneath {}", base);
+    }
+
+    #[test]
+    fn like_subtree_pattern_escapes_metacharacters() {
+        use std::path::MAIN_SEPARATOR as SEP;
+
+        let conn = rusqlite::Connection::open_in_memory().unwrap();
+        conn.execute("CREATE TABLE files (parent TEXT NOT NULL)", [])
+            .unwrap();
+
+        let base = format!("{}a_b", root_prefix());
+        for r in [
+            format!("{}{}inside", base, SEP),         // real child
+            format!("{}axb{}bait", root_prefix(), SEP), // `_` must not glob to `x`
+            format!("{}100%_done{}x", root_prefix(), SEP),
+        ] {
+            conn.execute("INSERT INTO files (parent) VALUES (?1)", [&r])
+                .unwrap();
+        }
+
+        let matched: i64 = conn
+            .query_row(
+                "SELECT COUNT(*) FROM files WHERE parent LIKE ?1 ESCAPE '\\'",
+                [like_subtree_pattern(&base)],
+                |r| r.get(0),
+            )
+            .unwrap();
+        assert_eq!(matched, 1, "`_` and `%` are literals, not wildcards");
+    }
+
+    #[test]
+    fn folder_value_normalization() {
+        use std::path::MAIN_SEPARATOR as SEP;
+        // Trailing separators of either flavour are stripped.
+        assert_eq!(normalize_folder_value("/home/me/"), "/home/me");
+        assert_eq!(normalize_folder_value(r"C:\Users\me\"), r"C:\Users\me");
+        // A bare drive is not a path; the stored parent is `C:\`.
+        assert_eq!(normalize_folder_value("C:"), format!("C:{}", SEP));
+        assert_eq!(normalize_folder_value(r"C:\"), format!("C:{}", SEP));
+        // Empty means "everywhere".
+        assert_eq!(normalize_folder_value("/"), "");
+        assert_eq!(normalize_folder_value("  "), "");
+    }
+
+    #[test]
+    fn empty_folder_value_matches_everything() {
+        let frag = build_filter("path", Op::Contains, "/", false).unwrap();
+        assert_eq!(frag.sql, "1=1");
+        assert!(frag.params.is_empty(), "no placeholders to renumber");
+    }
+
+    /// An absolute-path prefix for the running platform.
+    fn root_prefix() -> String {
+        if cfg!(windows) {
+            r"C:\".to_string()
+        } else {
+            "/".to_string()
+        }
     }
 
     #[test]

crates/quicksearch-core/src/search/cascade.rs

@@ -0,0 +1,937 @@
+//! The ranked search cascade.
+//!
+//! One term, four table scans, eleven ranks. Rank base = stage number, so
+//! later stages only ever append to a rank-sorted result list:
+//!
+//! | rank | meaning                          | scan |
+//! |-----:|----------------------------------|------|
+//! |  1.x | exact filename, exact case       | A    |
+//! |  2.x | exact filename, any case         | A    |
+//! |  3.x | filename substring, exact case   | A    |
+//! |  4.x | filename substring, any case     | A    |
+//! |  5.x | full text occurrence, exact case | B    |
+//! |  6.x | full text occurrence, any case   | B    |
+//! |  7.x | fuzzy filename                   | C    |
+//! |  8.x | fuzzy full text                  | D    |
+//! |  9.x | full path substring, exact case  | A    |
+//! | 10.x | full path substring, any case    | A    |
+//! | 11.x | fuzzy full path                  | C    |
+//!
+//! Pass A is a single `files` scan (`LIKE`, the ASCII-nocase superset of
+//! its ranks) classified per-row in Rust — no index needed, the substring
+//! stage visits every row anyway. Because a path always ends in its own
+//! name, `path LIKE` is a superset of `name LIKE`, so that one scan covers
+//! the filename *and* the path tiers. Pass B is one FTS phrase MATCH
+//! verified against the decompressed text. Passes C/D (opt-in) iterate the
+//! whole table with a bitap matcher, C covering both the name and the path.
+//!
+//! Wildcard terms (`rep*rt`) rank through the same tiers, with 1/2 meaning
+//! the whole name matches the pattern; they skip the fuzzy passes (bitap is
+//! a literal matcher). A regex-only query (`regex:…` with no term) runs two
+//! dedicated scans that reuse tiers 4 (name), 6 (content) and 10 (path), so
+//! downstream stage handling is unchanged. When `regex:` accompanies a
+//! term, it is an accept-predicate on every pass, not a rank source.
+//!
+//! The path tiers rank below everything else, so pass A and pass C buffer
+//! them instead of emitting them — stages E and F flush those buffers at
+//! the end, dropping files an earlier stage already emitted. Path matching
+//! needs a term of at least three characters, the same floor pass B has.
+//!
+//! Full-text ranks order equal-based hits by occurrence count via a
+//! decimal fraction: `base + (1000 - min(count, 1000)) / 1000` — more
+//! occurrences sorts earlier, 1000+ occurrences adds zero. Fuzzy ranks add
+//! `0.1 × edit_distance` instead.
+//!
+//! Every scan appends the caller's structured-filter SQL (anonymous
+//! placeholders over alias `f`) and checks the generation counter as it
+//! streams; a bumped generation aborts mid-statement.
+
+use std::collections::HashSet;
+use std::sync::atomic::{AtomicU64, Ordering};
+
+use rusqlite::Connection;
+
+use rusqlite::OptionalExtension;
+
+use crate::config::IgnoreSet;
+use crate::query::pattern::clamp_match_range;
+use crate::query::split::CascadeQuery;
+use crate::query::translator::{escape_like, quote_phrase};
+use crate::snippet;
+
+use super::fuzzy::{edit_budget, Bitap};
+use super::{SearchHit, SearchOptions};
+
+/// Cancellation is checked every this many scanned rows in row-cheap
+/// passes; decompression-heavy passes check every row.
+const CANCEL_CHECK_ROWS: usize = 256;
+
+/// Snippet window budget. Generous on purpose: the GUI trims the cell
+/// text down to its column width around the match, and the mouseover
+/// shows the rest of this window as extended context.
+const SNIPPET_WINDOW_CHARS: usize = 600;
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub struct Outcome {
+    pub total: usize,
+    pub limited: bool,
+}
+
+/// Run the cascade, streaming rank-ordered batches into `sink`.
+/// `Ok(None)` means the search was cancelled (generation moved on) — the
+/// caller sends no completion. SQL errors are returned as strings *unless*
+/// the search was already cancelled (an interrupted statement is normal
+/// cancellation, not an error).
+pub fn run(
+    conn: &Connection,
+    query: &CascadeQuery,
+    options: &SearchOptions,
+    generation: u64,
+    latest_gen: &AtomicU64,
+    sink: &mut dyn FnMut(Vec<SearchHit>),
+) -> Result<Option<Outcome>, String> {
+    if query.is_empty() {
+        return Ok(Some(Outcome {
+            total: 0,
+            limited: false,
+        }));
+    }
+    let ignore = IgnoreSet::compile(&options.session_ignores)
+        .map_err(|e| format!("session ignore filter: {}", e))?;
+
+    let mut cx = Cx {
+        conn,
+        query,
+        options,
+        generation,
+        latest_gen,
+        ignore,
+        emitted: HashSet::new(),
+        deferred_path: Deferred::default(),
+        deferred_fuzzy_path: Deferred::default(),
+        total: 0,
+        limited: false,
+        sink,
+    };
+
+    // With no term at all the regex drives its own scans; `Path` still
+    // flushes the deferred rank-10 buffer the name pass sets aside.
+    let passes: &[Pass] = if query.pattern.is_empty() {
+        &[Pass::RegexName, Pass::RegexContent, Pass::Path]
+    } else {
+        &[
+            Pass::Filename,
+            Pass::FullText,
+            Pass::FuzzyFilename,
+            Pass::FuzzyFullText,
+            Pass::Path,
+            Pass::FuzzyPath,
+        ]
+    };
+    for pass in passes {
+        if cx.cancelled() {
+            return Ok(None);
+        }
+        if cx.remaining() == 0 {
+            cx.limited = true;
+            break;
+        }
+        let run_pass = match pass {
+            Pass::Filename => cx.pass_filename(),
+            Pass::FullText => cx.pass_fulltext(),
+            Pass::FuzzyFilename => cx.pass_fuzzy_filename(),
+            Pass::FuzzyFullText => cx.pass_fuzzy_fulltext(),
+            Pass::RegexName => cx.pass_regex_name(),
+            Pass::RegexContent => cx.pass_regex_content(),
+            Pass::Path => {
+                let d = std::mem::take(&mut cx.deferred_path);
+                cx.flush_deferred(d)
+            }
+            Pass::FuzzyPath => {
+                let d = std::mem::take(&mut cx.deferred_fuzzy_path);
+                cx.flush_deferred(d)
+            }
+        };
+        match run_pass {
+            Ok(true) => {}
+            Ok(false) => return Ok(None), // cancelled mid-pass
+            Err(e) => {
+                if cx.cancelled() {
+                    return Ok(None); // interrupt() killed the statement
+                }
+                return Err(e);
+            }
+        }
+    }
+
+    Ok(Some(Outcome {
+        total: cx.total,
+        limited: cx.limited,
+    }))
+}
+
+enum Pass {
+    Filename,
+    FullText,
+    FuzzyFilename,
+    FuzzyFullText,
+    /// Regex-only: name hits at rank 4 now, path hits deferred to rank 10.
+    RegexName,
+    /// Regex-only: content hits at rank 6.
+    RegexContent,
+    /// Flush of the rank 9–10 hits pass A set aside.
+    Path,
+    /// Flush of the rank 11 hits pass C set aside.
+    FuzzyPath,
+}
+
+/// Occurrence-count fraction: more occurrences → smaller fraction → sorts
+/// earlier within a rank base; 1000+ adds zero.
+fn count_frac(count: usize) -> f64 {
+    (1000usize.saturating_sub(count.min(1000))) as f64 / 1000.0
+}
+
+/// The path tiers only make sense with enough term to be specific — same
+/// floor the trigram full-text pass uses. Wildcards count only their
+/// literal content (`a*b` is two characters of specificity, not three).
+fn path_tiers_enabled(pattern: &crate::query::pattern::TermPattern) -> bool {
+    pattern.literal_char_count() >= 3
+}
+
+/// Hits collected by one scan but ranked below later scans, so held back
+/// until every better stage has emitted.
+#[derive(Default)]
+struct Deferred {
+    hits: Vec<SearchHit>,
+    overflowed: bool,
+}
+
+struct Cx<'a> {
+    conn: &'a Connection,
+    query: &'a CascadeQuery,
+    options: &'a SearchOptions,
+    generation: u64,
+    latest_gen: &'a AtomicU64,
+    ignore: IgnoreSet,
+    emitted: HashSet<i64>,
+    /// Ranks 9–10, filled by pass A.
+    deferred_path: Deferred,
+    /// Rank 11, filled by pass C.
+    deferred_fuzzy_path: Deferred,
+    total: usize,
+    limited: bool,
+    sink: &'a mut dyn FnMut(Vec<SearchHit>),
+}
+
+impl<'a> Cx<'a> {
+    fn cancelled(&self) -> bool {
+        self.generation != self.latest_gen.load(Ordering::Relaxed)
+    }
+
+    fn remaining(&self) -> usize {
+        self.options.limit.saturating_sub(self.total)
+    }
+
+    /// Buffer cap for scan passes: enough headroom that sorting keeps the
+    /// best candidates, without unbounded growth on huge hit sets.
+    fn buffer_cap(&self) -> usize {
+        4096.max(2 * self.remaining())
+    }
+
+    fn params_with_filters(
+        &self,
+        leading: Vec<rusqlite::types::Value>,
+    ) -> Vec<rusqlite::types::Value> {
+        let mut p = leading;
+        p.extend(self.query.filter_params.iter().cloned());
+        p
+    }
+
+    /// Skip rows already emitted at a better rank or hidden by session
+    /// ignore chips.
+    fn skip(&self, file_id: i64, path: &str) -> bool {
+        self.emitted.contains(&file_id) || self.ignore.matches_path(std::path::Path::new(path))
+    }
+
+    /// The `regex:` accept-predicate applied to every candidate row when a
+    /// regex accompanies a term. The path contains the name, so one path
+    /// check covers both; content is fetched (and decompressed) only for
+    /// rows whose path missed — bounded by the pass's hit count, not its
+    /// scan count. Pass `text` when the pass already has the content.
+    fn regex_accepts(
+        &self,
+        file_id: i64,
+        path: &str,
+        text: Option<&str>,
+    ) -> Result<bool, String> {
+        let Some(re) = &self.query.regex else {
+            return Ok(true);
+        };
+        if re.is_match(path) {
+            return Ok(true);
+        }
+        if let Some(text) = text {
+            return Ok(re.is_match(text));
+        }
+        let blob: Option<Vec<u8>> = self
+            .conn
+            .query_row(
+                "SELECT text_zstd FROM documents_text WHERE file_id = ?1",
+                [file_id],
+                |r| r.get(0),
+            )
+            .optional()
+            .map_err(|e| e.to_string())?;
+        let Some(raw) = blob.and_then(|b| zstd::decode_all(b.as_slice()).ok()) else {
+            return Ok(false);
+        };
+        Ok(re.is_match(&String::from_utf8_lossy(&raw)))
+    }
+
+    /// Sort a finished pass buffer, truncate to what's left of the display
+    /// limit, and stream it out in `options.batch`-sized events.
+    fn flush_pass(&mut self, mut buf: Vec<SearchHit>, overflowed: bool) {
+        buf.sort_by(|a, b| {
+            a.rank
+                .partial_cmp(&b.rank)
+                .unwrap_or(std::cmp::Ordering::Equal)
+                .then_with(|| a.name.cmp(&b.name))
+                .then_with(|| a.path.cmp(&b.path))
+        });
+        let room = self.remaining();
+        if buf.len() > room {
+            buf.truncate(room);
+            self.limited = true;
+        }
+        if overflowed {
+            self.limited = true;
+        }
+        self.total += buf.len();
+        for hit in &buf {
+            self.emitted.insert(hit.file_id);
+        }
+        let batch = self.options.batch.max(1);
+        let mut buf = buf.into_iter().peekable();
+        while buf.peek().is_some() {
+            // A cancelled search stops emitting immediately — the newer
+            // generation owns the UI.
+            if self.cancelled() {
+                return;
+            }
+            let chunk: Vec<SearchHit> = buf.by_ref().take(batch).collect();
+            (self.sink)(chunk);
+        }
+    }
+
+    /// Emit a buffer held back from an earlier scan. Anything a better
+    /// stage already emitted drops out here — `emitted` was still empty (or
+    /// smaller) when these hits were collected.
+    fn flush_deferred(&mut self, mut deferred: Deferred) -> Result<bool, String> {
+        deferred.hits.retain(|h| !self.emitted.contains(&h.file_id));
+        self.flush_pass(deferred.hits, deferred.overflowed);
+        Ok(true)
+    }
+
+    /// Keep a scan buffer bounded: sort + cut back to the display-limit
+    /// room once it doubles past it. Returns whether anything was dropped.
+    fn enforce_cap(&self, buf: &mut Vec<SearchHit>) -> bool {
+        if buf.len() <= self.buffer_cap() {
+            return false;
+        }
+        buf.sort_by(|a, b| {
+            a.rank
+                .partial_cmp(&b.rank)
+                .unwrap_or(std::cmp::Ordering::Equal)
+                .then_with(|| a.name.cmp(&b.name))
+                .then_with(|| a.path.cmp(&b.path))
+        });
+        buf.truncate(self.remaining());
+        true
+    }
+
+    /// Pass A — ranks 1–4 now, ranks 9–10 deferred, from one `files` scan.
+    /// Returns Ok(false) on cancellation.
+    fn pass_filename(&mut self) -> Result<bool, String> {
+        let query = self.query;
+        let pattern = &query.pattern;
+        let with_paths = path_tiers_enabled(pattern);
+        // A path always ends in its own name, so `path LIKE` is the
+        // superset that feeds both the name and the path tiers.
+        let sql = format!(
+            "SELECT f.id, f.name, f.path, f.size, f.mtime FROM files f \
+             WHERE {} LIKE ? ESCAPE '\\'{}",
+            if with_paths { "f.path" } else { "f.name" },
+            query.filter_sql
+        );
+        // Wildcard patterns turn each star into an unescaped `%`; the
+        // substring wrap absorbs leading/trailing stars. User `%`/`_`
+        // remain escaped literals either way.
+        let like = pattern
+            .segments()
+            .iter()
+            .map(|s| escape_like(s))
+            .collect::<Vec<_>>()
+            .join("%");
+        let params = self.params_with_filters(vec![rusqlite::types::Value::Text(format!(
+            "%{}%",
+            like
+        ))]);
+
+        let mut stmt = self.conn.prepare(&sql).map_err(|e| e.to_string())?;
+        let mut rows = stmt
+            .query(rusqlite::params_from_iter(params))
+            .map_err(|e| e.to_string())?;
+
+        let mut buf: Vec<SearchHit> = Vec::new();
+        let mut path_buf: Vec<SearchHit> = Vec::new();
+        let mut overflowed = false;
+        let mut path_overflowed = false;
+        let mut scanned = 0usize;
+        while let Some(row) = rows.next().map_err(|e| e.to_string())? {
+            scanned += 1;
+            if scanned % CANCEL_CHECK_ROWS == 0 && self.cancelled() {
+                return Ok(false);
+            }
+            let file_id: i64 = row.get(0).map_err(|e| e.to_string())?;
+            let name: String = row.get(1).map_err(|e| e.to_string())?;
+            let path: String = row.get(2).map_err(|e| e.to_string())?;
+            if self.skip(file_id, &path) {
+                continue;
+            }
+            // For a literal pattern these are exactly the old `==` /
+            // `eq_ignore_ascii_case` / `find` / folded-`find` operations
+            // (folding is byte-length preserving, so folded offsets are
+            // valid in the original). Wildcards run the same ladder through
+            // their compiled matcher — tiers 1/2 mean "whole name matches
+            // the pattern", which is what `*.txt` should do.
+            let (rank, match_range) = if pattern.whole_match(&name, false) {
+                (1.0, (0, name.len()))
+            } else if pattern.whole_match(&name, true) {
+                (2.0, (0, name.len()))
+            } else if let Some(r) = pattern.find_first(&name, false) {
+                (3.0, (r.start, r.end))
+            } else if let Some(r) = pattern.find_first(&name, true) {
+                (4.0, (r.start, r.end))
+            } else if !with_paths {
+                continue;
+            } else if let Some(r) = pattern.find_first(&path, false) {
+                (9.0, (r.start, r.end))
+            } else if let Some(r) = pattern.find_first(&path, true) {
+                (10.0, (r.start, r.end))
+            } else {
+                // LIKE folds ASCII case only; a row that matched it but
+                // neither field is a non-ASCII near-miss. Drop it.
+                continue;
+            };
+            if !self.regex_accepts(file_id, &path, None)? {
+                continue;
+            }
+            let is_path_tier = rank >= 9.0;
+            // The "snippet" of a name or path hit is that field itself with
+            // the matched span marked — the GUI renders it as [the field].
+            let snip = snippet::Snippet {
+                ranges: vec![match_range],
+                window: if is_path_tier { path.clone() } else { name.clone() },
+                truncated_start: false,
+                truncated_end: false,
+            };
+            let hit = SearchHit {
+                file_id,
+                name,
+                path,
+                size: row.get::<_, i64>(3).map_err(|e| e.to_string())?.max(0) as u64,
+                mtime: row.get(4).map_err(|e| e.to_string())?,
+                rank,
+                stage: rank as u8,
+                snippet: Some(snip),
+            };
+            if is_path_tier {
+                path_buf.push(hit);
+                path_overflowed |= self.enforce_cap(&mut path_buf);
+            } else {
+                buf.push(hit);
+                overflowed |= self.enforce_cap(&mut buf);
+            }
+        }
+        drop(rows);
+        if self.cancelled() {
+            return Ok(false);
+        }
+        self.deferred_path = Deferred {
+            hits: path_buf,
+            overflowed: path_overflowed,
+        };
+        self.flush_pass(buf, overflowed);
+        Ok(true)
+    }
+
+    /// Pass B — ranks 5–6 from one FTS MATCH, verified/counted in the
+    /// decompressed text.
+    fn pass_fulltext(&mut self) -> Result<bool, String> {
+        let query = self.query;
+        let pattern = &query.pattern;
+        if pattern.literal_char_count() < 3 {
+            // Below the trigram floor the MATCH can't return anything.
+            return Ok(true);
+        }
+        // Column filter goes inside the MATCH expression (`text: "..."`)
+        // so only document bodies match — filenames get ranks 1-4 from the
+        // filename pass instead. A literal term is one quoted phrase; a
+        // wildcard term narrows with an AND of its trigram-sized segments.
+        // With no segment of 3+ chars (`ab*cd`) FTS can't narrow at all, so
+        // fall back to scanning every stored document — every row is
+        // pattern-verified either way.
+        let match_expr: Option<String> = match pattern.literal() {
+            Some(term) => Some(format!("text: {}", quote_phrase(term))),
+            None => {
+                let usable: Vec<String> = pattern
+                    .segments()
+                    .iter()
+                    .filter(|s| s.chars().count() >= 3)
+                    .map(|s| format!("text: {}", quote_phrase(s)))
+                    .collect();
+                if usable.is_empty() {
+                    None
+                } else {
+                    Some(usable.join(" AND "))
+                }
+            }
+        };
+        let narrowed = match_expr.is_some();
+        let (sql, params) = match match_expr {
+            Some(expr) => (
+                format!(
+                    "SELECT f.id, f.name, f.path, f.size, f.mtime, dt.text_zstd \
+                     FROM searchabletext \
+                     JOIN files f ON f.id = searchabletext.rowid \
+                     LEFT JOIN documents_text dt ON dt.file_id = f.id \
+                     WHERE searchabletext MATCH ?{}",
+                    query.filter_sql
+                ),
+                self.params_with_filters(vec![rusqlite::types::Value::Text(expr)]),
+            ),
+            None => (
+                format!(
+                    "SELECT f.id, f.name, f.path, f.size, f.mtime, dt.text_zstd \
+                     FROM documents_text dt \
+                     JOIN files f ON f.id = dt.file_id WHERE 1=1{}",
+                    query.filter_sql
+                ),
+                self.params_with_filters(Vec::new()),
+            ),
+        };
+
+        let mut stmt = self.conn.prepare(&sql).map_err(|e| e.to_string())?;
+        let mut rows = stmt
+            .query(rusqlite::params_from_iter(params))
+            .map_err(|e| e.to_string())?;
+
+        let snippet_opts = snippet::Options { approx_chars: SNIPPET_WINDOW_CHARS };
+        let mut buf: Vec<SearchHit> = Vec::new();
+        let mut overflowed = false;
+        while let Some(row) = rows.next().map_err(|e| e.to_string())? {
+            // Decompression dominates: check every row.
+            if self.cancelled() {
+                return Ok(false);
+            }
+            let file_id: i64 = row.get(0).map_err(|e| e.to_string())?;
+            let path: String = row.get(2).map_err(|e| e.to_string())?;
+            if self.skip(file_id, &path) {
+                continue;
+            }
+            let blob: Option<Vec<u8>> = row.get(5).map_err(|e| e.to_string())?;
+            let text = blob.and_then(|b| zstd::decode_all(b.as_slice()).ok()).map(
+                |raw| String::from_utf8_lossy(&raw).into_owned(),
+            );
+
+            let (rank, stage, snip) = match &text {
+                Some(text) => {
+                    let (count, stage, ci) = {
+                        let count_cs = pattern.count(text, false);
+                        if count_cs > 0 {
+                            (count_cs, 5, false)
+                        } else {
+                            let count_ci = pattern.count(text, true);
+                            if count_ci > 0 {
+                                (count_ci, 6, true)
+                            } else {
+                                // Folded/unordered FTS candidate: the
+                                // pattern never occurs — drop it.
+                                continue;
+                            }
+                        }
+                    };
+                    // Literal terms keep the richer multi-occurrence
+                    // extract; a wildcard match marks its own first range.
+                    let snip = match pattern.literal() {
+                        Some(term) => Some(snippet::extract(text, &[term], &snippet_opts)),
+                        None => pattern.find_first(text, ci).map(|r| {
+                            let r = clamp_match_range(text, r, SNIPPET_WINDOW_CHARS);
+                            snippet::window_around(text, (r.start, r.end), &snippet_opts)
+                        }),
+                    };
+                    (stage as f64 + count_frac(count), stage as u8, snip)
+                }
+                // No stored text (store_text_for_snippets = false or empty
+                // body): can't case-verify or count. On the FTS-narrowed
+                // path accept at the bottom of rank 6 as count-unknown (for
+                // wildcards the AND-of-segments guarantee is weaker —
+                // unordered co-occurrence — accepted for recall). On the
+                // full-scan fallback there is no FTS evidence at all, so an
+                // unverifiable row is just skipped.
+                None => {
+                    if !narrowed {
+                        continue;
+                    }
+                    (6.0 + count_frac(1), 6, None)
+                }
+            };
+            if !self.regex_accepts(file_id, &path, text.as_deref())? {
+                continue;
+            }
+
+            buf.push(SearchHit {
+                file_id,
+                name: row.get(1).map_err(|e| e.to_string())?,
+                path,
+                size: row.get::<_, i64>(3).map_err(|e| e.to_string())?.max(0) as u64,
+                mtime: row.get(4).map_err(|e| e.to_string())?,
+                rank,
+                stage,
+                snippet: snip,
+            });
+            overflowed |= self.enforce_cap(&mut buf);
+        }
+        drop(rows);
+        if self.cancelled() {
+            return Ok(false);
+        }
+        self.flush_pass(buf, overflowed);
+        Ok(true)
+    }
+
+    /// Pass C — rank 7 now, rank 11 deferred: one bitap sweep over every
+    /// filename, falling back to the full path where the name misses.
+    fn pass_fuzzy_filename(&mut self) -> Result<bool, String> {
+        if !self.options.fuzzy {
+            return Ok(true);
+        }
+        // Bitap is a literal matcher; wildcard terms don't fuzz.
+        if self.query.pattern.is_wildcard() {
+            return Ok(true);
+        }
+        let folded_term = self.query.term.to_ascii_lowercase();
+        let Some(k) = edit_budget(folded_term.len(), self.options.fuzzy_max_edits) else {
+            return Ok(true);
+        };
+        let Some(bitap) = Bitap::new(folded_term.as_bytes(), k) else {
+            return Ok(true);
+        };
+        let with_paths = path_tiers_enabled(&self.query.pattern);
+
+        let sql = format!(
+            "SELECT f.id, f.name, f.path, f.size, f.mtime FROM files f WHERE 1=1{}",
+            self.query.filter_sql
+        );
+        let params = self.params_with_filters(Vec::new());
+        let mut stmt = self.conn.prepare(&sql).map_err(|e| e.to_string())?;
+        let mut rows = stmt
+            .query(rusqlite::params_from_iter(params))
+            .map_err(|e| e.to_string())?;
+
+        let mut buf: Vec<SearchHit> = Vec::new();
+        let mut path_buf: Vec<SearchHit> = Vec::new();
+        let mut overflowed = false;
+        let mut path_overflowed = false;
+        let mut scanned = 0usize;
+        while let Some(row) = rows.next().map_err(|e| e.to_string())? {
+            scanned += 1;
+            if scanned % 1024 == 0 && self.cancelled() {
+                return Ok(false);
+            }
+            let file_id: i64 = row.get(0).map_err(|e| e.to_string())?;
+            let name: String = row.get(1).map_err(|e| e.to_string())?;
+            let path: String = row.get(2).map_err(|e| e.to_string())?;
+            if self.skip(file_id, &path) {
+                continue;
+            }
+            // The name is the better match when both fire, so it wins and
+            // only a name miss falls through to the path tier.
+            let folded_name = name.to_ascii_lowercase();
+            let (rank, field, folded_field) = match bitap.best_distance(folded_name.as_bytes()) {
+                Some(distance) => (7.0 + 0.1 * distance as f64, &name, folded_name),
+                None if with_paths => {
+                    let folded_path = path.to_ascii_lowercase();
+                    match bitap.best_distance(folded_path.as_bytes()) {
+                        Some(distance) => (11.0 + 0.1 * distance as f64, &path, folded_path),
+                        None => continue,
+                    }
+                }
+                None => continue,
+            };
+            if !self.regex_accepts(file_id, &path, None)? {
+                continue;
+            }
+            // Mark the approximate matched span in the matched field for
+            // the GUI's [matched field] rendering. window_around clamps
+            // and aligns.
+            let snip = bitap.count_and_first(folded_field.as_bytes()).1.map(|range| {
+                snippet::window_around(
+                    field,
+                    range,
+                    &snippet::Options {
+                        approx_chars: field.len().saturating_mul(2).max(8),
+                    },
+                )
+            });
+            let is_path_tier = rank >= 11.0;
+            let hit = SearchHit {
+                file_id,
+                name,
+                path,
+                size: row.get::<_, i64>(3).map_err(|e| e.to_string())?.max(0) as u64,
+                mtime: row.get(4).map_err(|e| e.to_string())?,
+                rank,
+                stage: rank as u8,
+                snippet: snip,
+            };
+            if is_path_tier {
+                path_buf.push(hit);
+                path_overflowed |= self.enforce_cap(&mut path_buf);
+            } else {
+                buf.push(hit);
+                overflowed |= self.enforce_cap(&mut buf);
+            }
+        }
+        drop(rows);
+        if self.cancelled() {
+            return Ok(false);
+        }
+        self.deferred_fuzzy_path = Deferred {
+            hits: path_buf,
+            overflowed: path_overflowed,
+        };
+        self.flush_pass(buf, overflowed);
+        Ok(true)
+    }
+
+    /// Pass D — rank 8, bitap over every stored document text.
+    fn pass_fuzzy_fulltext(&mut self) -> Result<bool, String> {
+        if !self.options.fuzzy {
+            return Ok(true);
+        }
+        // Bitap is a literal matcher; wildcard terms don't fuzz.
+        if self.query.pattern.is_wildcard() {
+            return Ok(true);
+        }
+        let folded_term = self.query.term.to_ascii_lowercase();
+        let Some(k) = edit_budget(folded_term.len(), self.options.fuzzy_max_edits) else {
+            return Ok(true);
+        };
+        let Some(bitap) = Bitap::new(folded_term.as_bytes(), k) else {
+            return Ok(true);
+        };
+
+        let sql = format!(
+            "SELECT f.id, f.name, f.path, f.size, f.mtime, dt.text_zstd \
+             FROM documents_text dt JOIN files f ON f.id = dt.file_id WHERE 1=1{}",
+            self.query.filter_sql
+        );
+        let params = self.params_with_filters(Vec::new());
+        let mut stmt = self.conn.prepare(&sql).map_err(|e| e.to_string())?;
+        let mut rows = stmt
+            .query(rusqlite::params_from_iter(params))
+            .map_err(|e| e.to_string())?;
+
+        let snippet_opts = snippet::Options { approx_chars: SNIPPET_WINDOW_CHARS };
+        let mut buf: Vec<SearchHit> = Vec::new();
+        let mut overflowed = false;
+        while let Some(row) = rows.next().map_err(|e| e.to_string())? {
+            if self.cancelled() {
+                return Ok(false);
+            }
+            let file_id: i64 = row.get(0).map_err(|e| e.to_string())?;
+            let path: String = row.get(2).map_err(|e| e.to_string())?;
+            if self.skip(file_id, &path) {
+                continue;
+            }
+            let blob: Option<Vec<u8>> = row.get(5).map_err(|e| e.to_string())?;
+            let Some(blob) = blob else {
+                continue;
+            };
+            let Ok(raw) = zstd::decode_all(blob.as_slice()) else {
+                continue;
+            };
+            let text = String::from_utf8_lossy(&raw).into_owned();
+            // ASCII folding is byte-length preserving, so ranges found in
+            // the folded buffer are valid in the original.
+            let folded = text.to_ascii_lowercase();
+            let (count, first) = bitap.count_and_first(folded.as_bytes());
+            if count == 0 {
+                continue;
+            }
+            if !self.regex_accepts(file_id, &path, Some(&text))? {
+                continue;
+            }
+            let snip = first.map(|range| snippet::window_around(&text, range, &snippet_opts));
+            buf.push(SearchHit {
+                file_id,
+                name: row.get(1).map_err(|e| e.to_string())?,
+                path,
+                size: row.get::<_, i64>(3).map_err(|e| e.to_string())?.max(0) as u64,
+                mtime: row.get(4).map_err(|e| e.to_string())?,
+                rank: 8.0 + count_frac(count),
+                stage: 8,
+                snippet: snip,
+            });
+            overflowed |= self.enforce_cap(&mut buf);
+        }
+        drop(rows);
+        if self.cancelled() {
+            return Ok(false);
+        }
+        self.flush_pass(buf, overflowed);
+        Ok(true)
+    }
+
+    /// Regex-only pass over `files`: the regex bypasses the FTS trigram
+    /// entirely and runs on every name, falling back to the full path.
+    /// Name hits reuse rank 4, path hits defer to rank 10, so the GUI's
+    /// stage-based rendering needs no new cases.
+    fn pass_regex_name(&mut self) -> Result<bool, String> {
+        let query = self.query;
+        let re = query.regex.as_ref().expect("regex-only pass list");
+        let sql = format!(
+            "SELECT f.id, f.name, f.path, f.size, f.mtime FROM files f WHERE 1=1{}",
+            query.filter_sql
+        );
+        let params = self.params_with_filters(Vec::new());
+        let mut stmt = self.conn.prepare(&sql).map_err(|e| e.to_string())?;
+        let mut rows = stmt
+            .query(rusqlite::params_from_iter(params))
+            .map_err(|e| e.to_string())?;
+
+        let mut buf: Vec<SearchHit> = Vec::new();
+        let mut path_buf: Vec<SearchHit> = Vec::new();
+        let mut overflowed = false;
+        let mut path_overflowed = false;
+        let mut scanned = 0usize;
+        while let Some(row) = rows.next().map_err(|e| e.to_string())? {
+            scanned += 1;
+            if scanned % 1024 == 0 && self.cancelled() {
+                return Ok(false);
+            }
+            let file_id: i64 = row.get(0).map_err(|e| e.to_string())?;
+            let name: String = row.get(1).map_err(|e| e.to_string())?;
+            let path: String = row.get(2).map_err(|e| e.to_string())?;
+            if self.skip(file_id, &path) {
+                continue;
+            }
+            // The name is the better hit; only a name miss falls through
+            // to the path tier — mirroring pass A.
+            let (rank, match_range, is_path_tier) = match re.find_first(&name) {
+                Some(r) => (4.0, (r.start, r.end), false),
+                None => match re.find_first(&path) {
+                    Some(r) => (10.0, (r.start, r.end), true),
+                    None => continue,
+                },
+            };
+            let snip = snippet::Snippet {
+                ranges: vec![match_range],
+                window: if is_path_tier { path.clone() } else { name.clone() },
+                truncated_start: false,
+                truncated_end: false,
+            };
+            let hit = SearchHit {
+                file_id,
+                name,
+                path,
+                size: row.get::<_, i64>(3).map_err(|e| e.to_string())?.max(0) as u64,
+                mtime: row.get(4).map_err(|e| e.to_string())?,
+                rank,
+                stage: rank as u8,
+                snippet: Some(snip),
+            };
+            if is_path_tier {
+                path_buf.push(hit);
+                path_overflowed |= self.enforce_cap(&mut path_buf);
+            } else {
+                buf.push(hit);
+                overflowed |= self.enforce_cap(&mut buf);
+            }
+        }
+        drop(rows);
+        if self.cancelled() {
+            return Ok(false);
+        }
+        self.deferred_path = Deferred {
+            hits: path_buf,
+            overflowed: path_overflowed,
+        };
+        self.flush_pass(buf, overflowed);
+        Ok(true)
+    }
+
+    /// Regex-only pass over every stored document text, reusing rank 6.
+    fn pass_regex_content(&mut self) -> Result<bool, String> {
+        let query = self.query;
+        let re = query.regex.as_ref().expect("regex-only pass list");
+        let sql = format!(
+            "SELECT f.id, f.name, f.path, f.size, f.mtime, dt.text_zstd \
+             FROM documents_text dt JOIN files f ON f.id = dt.file_id WHERE 1=1{}",
+            query.filter_sql
+        );
+        let params = self.params_with_filters(Vec::new());
+        let mut stmt = self.conn.prepare(&sql).map_err(|e| e.to_string())?;
+        let mut rows = stmt
+            .query(rusqlite::params_from_iter(params))
+            .map_err(|e| e.to_string())?;
+
+        let snippet_opts = snippet::Options { approx_chars: SNIPPET_WINDOW_CHARS };
+        let mut buf: Vec<SearchHit> = Vec::new();
+        let mut overflowed = false;
+        while let Some(row) = rows.next().map_err(|e| e.to_string())? {
+            // Decompression dominates: check every row.
+            if self.cancelled() {
+                return Ok(false);
+            }
+            let file_id: i64 = row.get(0).map_err(|e| e.to_string())?;
+            let path: String = row.get(2).map_err(|e| e.to_string())?;
+            if self.skip(file_id, &path) {
+                continue;
+            }
+            let blob: Option<Vec<u8>> = row.get(5).map_err(|e| e.to_string())?;
+            let Some(raw) = blob.and_then(|b| zstd::decode_all(b.as_slice()).ok()) else {
+                continue;
+            };
+            let text = String::from_utf8_lossy(&raw).into_owned();
+            let count = re.count(&text);
+            if count == 0 {
+                continue;
+            }
+            // A greedy user regex can match megabytes; clamp the range
+            // before the snippet window is cut.
+            let snip = re.find_first(&text).map(|r| {
+                let r = clamp_match_range(&text, r, SNIPPET_WINDOW_CHARS);
+                snippet::window_around(&text, (r.start, r.end), &snippet_opts)
+            });
+            buf.push(SearchHit {
+                file_id,
+                name: row.get(1).map_err(|e| e.to_string())?,
+                path,
+                size: row.get::<_, i64>(3).map_err(|e| e.to_string())?.max(0) as u64,
+                mtime: row.get(4).map_err(|e| e.to_string())?,
+                rank: 6.0 + count_frac(count),
+                stage: 6,
+                snippet: snip,
+            });
+            overflowed |= self.enforce_cap(&mut buf);
+        }
+        drop(rows);
+        if self.cancelled() {
+            return Ok(false);
+        }
+        self.flush_pass(buf, overflowed);
+        Ok(true)
+    }
+}

crates/quicksearch-core/src/search/duplicates.rs

@@ -0,0 +1,175 @@
+//! Duplicate-file listing, grouped by content hash. Backs the GUI's
+//! Duplicates tab; standalone and synchronous — callers run it on their
+//! own worker thread.
+
+use rusqlite::params;
+
+use crate::db;
+
+#[derive(Debug, Clone, PartialEq)]
+pub struct DuplicateGroup {
+    pub hash: Vec<u8>,
+    pub count: i64,
+    pub total_size: i64,
+    /// Bytes reclaimable by deduplicating: `size × (count - 1)` — the
+    /// group's sort key.
+    pub redundant_size: i64,
+    /// `(file_id, name, path, size, mtime)` per member, path-ordered.
+    pub members: Vec<(i64, String, String, u64, i64)>,
+}
+
+/// Page through hash groups having more than one member, ordered by
+/// reclaimable bytes (largest first). Rows with a NULL hash (never
+/// hashed) and zero-size files (all trivially identical) are excluded.
+pub fn find_duplicate_groups(
+    db_path: &str,
+    limit: u32,
+    offset: u32,
+) -> Result<Vec<DuplicateGroup>, String> {
+    let conn = db::open_existing(db_path, false)?;
+    let mut groups: Vec<DuplicateGroup> = Vec::new();
+    {
+        // SUM(size) - MAX(size) == size × (count - 1); members of a group
+        // share a size because the hash covers it.
+        let mut stmt = conn
+            .prepare(
+                "SELECT hash, COUNT(*) AS cnt, SUM(size), SUM(size) - MAX(size) AS redundant \
+                 FROM files \
+                 WHERE hash IS NOT NULL AND size > 0 \
+                 GROUP BY hash HAVING cnt > 1 \
+                 ORDER BY redundant DESC, hash \
+                 LIMIT ?1 OFFSET ?2",
+            )
+            .map_err(|e| e.to_string())?;
+        let rows = stmt
+            .query_map(params![limit, offset], |r| {
+                Ok(DuplicateGroup {
+                    hash: r.get(0)?,
+                    count: r.get(1)?,
+                    total_size: r.get(2)?,
+                    redundant_size: r.get(3)?,
+                    members: Vec::new(),
+                })
+            })
+            .map_err(|e| e.to_string())?;
+        for g in rows {
+            groups.push(g.map_err(|e| e.to_string())?);
+        }
+    }
+
+    let mut member_stmt = conn
+        .prepare(
+            "SELECT id, name, path, size, mtime FROM files WHERE hash = ?1 ORDER BY path",
+        )
+        .map_err(|e| e.to_string())?;
+    for group in &mut groups {
+        let rows = member_stmt
+            .query_map(params![group.hash], |r| {
+                Ok((
+                    r.get::<_, i64>(0)?,
+                    r.get::<_, String>(1)?,
+                    r.get::<_, String>(2)?,
+                    r.get::<_, i64>(3)?.max(0) as u64,
+                    r.get::<_, i64>(4)?,
+                ))
+            })
+            .map_err(|e| e.to_string())?;
+        for m in rows {
+            group.members.push(m.map_err(|e| e.to_string())?);
+        }
+    }
+
+    Ok(groups)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::db::open_or_recreate;
+    use crate::db::repo::{insert_file, NewFile};
+    use crate::mime::FileType;
+
+    fn seed_db() -> std::path::PathBuf {
+        let mut p = std::env::temp_dir();
+        p.push(format!(
+            "qs-dups-{}-{}.sqlite",
+            std::process::id(),
+            std::time::SystemTime::now()
+                .duration_since(std::time::UNIX_EPOCH)
+                .unwrap()
+                .as_nanos()
+        ));
+        let mut conn = open_or_recreate(p.to_str().unwrap(), "trigram").unwrap();
+        let tx = conn.transaction().unwrap();
+        let add = |name: &str, path: &str, size: u64, hash: Option<&[u8]>| {
+            insert_file(
+                &tx,
+                &NewFile {
+                    name,
+                    path,
+                    parent: "/d",
+                    size,
+                    mtime: 1_700_000_000,
+                    inode: None,
+                    device_id: None,
+                    mime: None,
+                    ftype: FileType::TEXT,
+                    hash,
+                },
+            )
+            .unwrap()
+            .expect("unique path");
+        };
+        // Triple group of small files: redundant = 10 × 2 = 20.
+        add("a1.txt", "/d/a1.txt", 10, Some(b"AAA"));
+        add("a2.txt", "/d/a2.txt", 10, Some(b"AAA"));
+        add("a3.txt", "/d/a3.txt", 10, Some(b"AAA"));
+        // Pair of large files: redundant = 100 × 1 = 100 — sorts first
+        // despite the smaller member count.
+        add("b1.txt", "/d/b1.txt", 100, Some(b"BBB"));
+        add("b2.txt", "/d/b2.txt", 100, Some(b"BBB"));
+        // Singletons and NULL hashes never appear.
+        add("c.txt", "/d/c.txt", 30, Some(b"CCC"));
+        add("n1.txt", "/d/n1.txt", 40, None);
+        add("n2.txt", "/d/n2.txt", 40, None);
+        // Zero-size files are trivially identical — excluded outright.
+        add("z1.txt", "/d/z1.txt", 0, Some(b"ZZZ"));
+        add("z2.txt", "/d/z2.txt", 0, Some(b"ZZZ"));
+        tx.commit().unwrap();
+        drop(conn);
+        p
+    }
+
+    #[test]
+    fn groups_ordered_by_redundant_size_zero_size_excluded() {
+        let p = seed_db();
+        let groups = find_duplicate_groups(p.to_str().unwrap(), 10, 0).unwrap();
+        assert_eq!(
+            groups.len(),
+            2,
+            "singletons, NULL hashes, and zero-size groups excluded"
+        );
+        // Reclaimable bytes beat member count for ordering.
+        assert_eq!(groups[0].count, 2);
+        assert_eq!(groups[0].total_size, 200);
+        assert_eq!(groups[0].redundant_size, 100);
+        assert_eq!(groups[1].count, 3);
+        assert_eq!(groups[1].redundant_size, 20);
+        assert_eq!(groups[1].members.len(), 3);
+        assert_eq!(groups[1].members[0].1, "a1.txt", "members path-ordered");
+        std::fs::remove_file(&p).ok();
+    }
+
+    #[test]
+    fn pagination() {
+        let p = seed_db();
+        let page1 = find_duplicate_groups(p.to_str().unwrap(), 1, 0).unwrap();
+        let page2 = find_duplicate_groups(p.to_str().unwrap(), 1, 1).unwrap();
+        assert_eq!(page1.len(), 1);
+        assert_eq!(page2.len(), 1);
+        assert_ne!(page1[0].hash, page2[0].hash);
+        let page3 = find_duplicate_groups(p.to_str().unwrap(), 1, 2).unwrap();
+        assert!(page3.is_empty());
+        std::fs::remove_file(&p).ok();
+    }
+}

crates/quicksearch-core/src/search/fuzzy.rs

@@ -0,0 +1,330 @@
+//! Approximate substring matching for the fuzzy cascade stages.
+//!
+//! Bitap (shift-and with errors, Wu–Manber): finds occurrences of a
+//! pattern *within* a haystack with at most `k` Levenshtein edits
+//! (insertion / deletion / substitution). Substring semantics are the
+//! right fit for search-as-you-type — users type fragments, not whole
+//! filenames — and the u64 bit-parallel update costs O(k) word ops per
+//! haystack byte with zero allocations.
+//!
+//! Callers fold both sides to ASCII lowercase first (the pipeline-wide
+//! convention). Patterns are limited to 64 bytes by the machine word; the
+//! cascade skips fuzzy stages for longer terms.
+
+pub struct Bitap {
+    /// `masks[c]` has bit `i` set iff `pattern[i] == c`.
+    masks: [u64; 256],
+    /// Pattern length in bytes (1..=64).
+    len: usize,
+    /// Maximum edit distance.
+    k: usize,
+}
+
+impl Bitap {
+    /// `None` when the pattern is empty, longer than 64 bytes, or the edit
+    /// budget reaches the word size (`initial_registers` shifts by `k`).
+    pub fn new(pattern: &[u8], k: usize) -> Option<Bitap> {
+        if pattern.is_empty() || pattern.len() > 64 || k >= 64 {
+            return None;
+        }
+        let mut masks = [0u64; 256];
+        for (i, &b) in pattern.iter().enumerate() {
+            masks[b as usize] |= 1u64 << i;
+        }
+        Some(Bitap {
+            masks,
+            len: pattern.len(),
+            k,
+        })
+    }
+
+    /// Fresh per-distance state registers. Bit `i` of `r[d]` set means "a
+    /// match of pattern[..=i] with ≤ d errors ends at the current text
+    /// position". With d errors the first d pattern bytes can be deleted
+    /// before any text is read, hence the pre-set low bits.
+    fn initial_registers(&self) -> Vec<u64> {
+        (0..=self.k)
+            .map(|d| if d == 0 { 0 } else { (1u64 << d) - 1 })
+            .collect()
+    }
+
+    /// Advance all registers by one haystack byte. Returns the smallest
+    /// error count d for which the full pattern just matched, if any.
+    #[inline]
+    fn step(&self, r: &mut [u64], byte: u8) -> Option<usize> {
+        let mask = self.masks[byte as usize];
+        let done = 1u64 << (self.len - 1);
+        let mut hit = None;
+        let mut prev_old = r[0]; // R_old[d-1] for the d-th iteration
+        // d = 0: exact prefix extension only.
+        r[0] = ((r[0] << 1) | 1) & mask;
+        if r[0] & done != 0 {
+            hit = Some(0);
+        }
+        for d in 1..=self.k {
+            let old = r[d];
+            r[d] = (((old << 1) | 1) & mask)      // extend a ≤d-error state
+                | prev_old                          // insertion in text
+                | (prev_old << 1)                   // substitution
+                | ((r[d - 1] << 1) | 1);            // deletion (pattern byte skipped)
+            prev_old = old;
+            if hit.is_none() && r[d] & done != 0 {
+                hit = Some(d);
+            }
+        }
+        hit
+    }
+
+    /// Minimum edit distance (≤ k) of any occurrence of the pattern in
+    /// `hay`, or `None` if nothing matches within k edits.
+    pub fn best_distance(&self, hay: &[u8]) -> Option<usize> {
+        let mut r = self.initial_registers();
+        let mut best: Option<usize> = None;
+        for &b in hay {
+            if let Some(d) = self.step(&mut r, b) {
+                if d == 0 {
+                    return Some(0);
+                }
+                if best.map_or(true, |cur| d < cur) {
+                    best = Some(d);
+                }
+            }
+        }
+        best
+    }
+
+    /// Count non-overlapping occurrences (at ≤ k edits) and report the
+    /// first match's approximate byte range in `hay`. After each hit the
+    /// automaton resets, so an exact match followed by trailing bytes
+    /// counts once, and overlapping suffix matches don't inflate counts.
+    /// The reported range assumes pattern-length matches — edits can shift
+    /// the true start by up to k bytes, which is fine for snippet windows.
+    pub fn count_and_first(&self, hay: &[u8]) -> (usize, Option<(usize, usize)>) {
+        let mut r = self.initial_registers();
+        let mut count = 0usize;
+        let mut first: Option<(usize, usize)> = None;
+        for (i, &b) in hay.iter().enumerate() {
+            if self.step(&mut r, b).is_some() {
+                count += 1;
+                if first.is_none() {
+                    let end = i + 1;
+                    first = Some((end.saturating_sub(self.len), end));
+                }
+                for (d, reg) in r.iter_mut().enumerate() {
+                    *reg = if d == 0 { 0 } else { (1u64 << d) - 1 };
+                }
+            }
+        }
+        (count, first)
+    }
+}
+
+/// The cascade's edit-distance budget for a folded term: one edit per
+/// three characters, capped by `[search].fuzzy_max_edits`. Terms outside
+/// 3..=64 bytes skip the fuzzy stages entirely (< 3 is noise, > 64 exceeds
+/// the word size), and a cap of 0 disables them everywhere.
+///
+/// At the default cap of 2 this is the historic ladder: 3–5 bytes get one
+/// edit, 6–64 get two.
+pub fn edit_budget(term_len: usize, max_edits: usize) -> Option<usize> {
+    if !(3..=64).contains(&term_len) || max_edits == 0 {
+        return None;
+    }
+    Some((term_len / 3).min(max_edits))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn best(pattern: &str, hay: &str, k: usize) -> Option<usize> {
+        Bitap::new(pattern.as_bytes(), k)
+            .unwrap()
+            .best_distance(hay.as_bytes())
+    }
+
+    #[test]
+    fn exact_substring_is_distance_zero() {
+        assert_eq!(best("hello", "say hello world", 2), Some(0));
+        assert_eq!(best("hello", "hello", 0), Some(0));
+    }
+
+    #[test]
+    fn single_edits_are_distance_one() {
+        assert_eq!(best("hello", "xx hxllo xx", 2), Some(1), "substitution");
+        assert_eq!(best("hello", "xx helo xx", 2), Some(1), "deletion");
+        assert_eq!(best("hello", "xx heXllo xx", 2), Some(1), "insertion");
+    }
+
+    #[test]
+    fn two_edits() {
+        assert_eq!(best("hello", "xx hxlo xx", 2), Some(2));
+        assert_eq!(best("hello", "xx ho xx", 2), None, "3 edits > k");
+    }
+
+    #[test]
+    fn no_match_within_budget() {
+        assert_eq!(best("hello", "completely different", 1), None);
+        assert_eq!(best("abc", "", 1), None);
+    }
+
+    #[test]
+    fn k_zero_is_exact_search() {
+        assert_eq!(best("abc", "xxabcxx", 0), Some(0));
+        assert_eq!(best("abc", "xxabxcx", 0), None);
+    }
+
+    #[test]
+    fn pattern_length_limits() {
+        assert!(Bitap::new(b"", 1).is_none());
+        assert!(Bitap::new(&[b'a'; 65], 1).is_none());
+        assert!(Bitap::new(&[b'a'; 64], 1).is_some());
+    }
+
+    #[test]
+    fn oversized_k_is_rejected_not_shifted() {
+        // `initial_registers` shifts by k; k >= 64 would overflow u64.
+        assert!(Bitap::new(b"abc", 64).is_none());
+        assert!(Bitap::new(b"abc", usize::MAX).is_none());
+        assert!(Bitap::new(b"abc", 63).is_some());
+    }
+
+    #[test]
+    fn count_non_overlapping() {
+        let b = Bitap::new(b"ab", 0).unwrap();
+        let (count, first) = b.count_and_first(b"ab ab ab");
+        assert_eq!(count, 3);
+        assert_eq!(first, Some((0, 2)));
+
+        // "aaaa" contains "aaa" once non-overlapping.
+        let b = Bitap::new(b"aaa", 0).unwrap();
+        let (count, _) = b.count_and_first(b"aaaa");
+        assert_eq!(count, 1);
+    }
+
+    #[test]
+    fn count_fuzzy_and_range_sane() {
+        let b = Bitap::new(b"hello", 1).unwrap();
+        let hay = b"say helo and hxllo again";
+        let (count, first) = b.count_and_first(hay);
+        assert_eq!(count, 2);
+        let (s, e) = first.unwrap();
+        assert!(s < e && e <= hay.len());
+        let window = &hay[s..e];
+        assert!(
+            std::str::from_utf8(window).unwrap().contains("hel"),
+            "first range should cover the first hit, got {:?}",
+            std::str::from_utf8(window)
+        );
+    }
+
+    #[test]
+    fn edit_budget_default_cap_is_the_historic_ladder() {
+        assert_eq!(edit_budget(0, 2), None);
+        assert_eq!(edit_budget(2, 2), None);
+        assert_eq!(edit_budget(3, 2), Some(1));
+        assert_eq!(edit_budget(5, 2), Some(1));
+        assert_eq!(edit_budget(6, 2), Some(2));
+        assert_eq!(edit_budget(64, 2), Some(2));
+        assert_eq!(edit_budget(65, 2), None);
+        assert_eq!(edit_budget(usize::MAX, 2), None);
+    }
+
+    #[test]
+    fn edit_budget_scales_with_length_up_to_the_cap() {
+        assert_eq!(edit_budget(3, 4), Some(1));
+        assert_eq!(edit_budget(6, 4), Some(2));
+        assert_eq!(edit_budget(9, 4), Some(3));
+        assert_eq!(edit_budget(12, 4), Some(4));
+        assert_eq!(edit_budget(64, 4), Some(4), "cap wins over length");
+    }
+
+    #[test]
+    fn edit_budget_cap_of_one_stays_strict() {
+        for len in 3..=64 {
+            assert_eq!(edit_budget(len, 1), Some(1));
+        }
+    }
+
+    #[test]
+    fn edit_budget_zero_disables_fuzzy() {
+        for len in 0..=70 {
+            assert_eq!(edit_budget(len, 0), None);
+        }
+    }
+
+    /// Even a hostile config value can't produce a k the bitap rejects:
+    /// the length ladder caps it at 21 for the longest legal term.
+    #[test]
+    fn edit_budget_stays_within_the_bitap_word_size() {
+        for len in 3..=64 {
+            let k = edit_budget(len, usize::MAX).unwrap();
+            assert!(k <= 21, "len={} gave k={}", len, k);
+            assert!(Bitap::new(&vec![b'a'; len], k).is_some());
+        }
+    }
+
+    /// Brute-force oracle: minimum Levenshtein distance between `pattern`
+    /// and any substring of `hay`, capped at k.
+    fn oracle(pattern: &[u8], hay: &[u8], k: usize) -> Option<usize> {
+        // An occurrence must end at some text position; empty text has
+        // none. Without this, k >= pattern-length "matches" empty text by
+        // deleting every pattern byte — a degenerate non-occurrence the
+        // automaton rightly never reports. (Production keeps k < len via
+        // the len/3 budget, so only the oracle ever saw this edge.)
+        if hay.is_empty() {
+            return None;
+        }
+        // Standard DP where row 0 is all zeros (match can start anywhere).
+        let m = pattern.len();
+        let mut prev: Vec<usize> = vec![0; hay.len() + 1];
+        let mut cur = vec![0; hay.len() + 1];
+        let mut best = usize::MAX;
+        // dp[i][j] = min edits to match pattern[..i] ending at hay[..j]
+        for i in 1..=m {
+            cur[0] = i;
+            for j in 1..=hay.len() {
+                let cost = if pattern[i - 1] == hay[j - 1] { 0 } else { 1 };
+                cur[j] = (prev[j - 1] + cost).min(prev[j] + 1).min(cur[j - 1] + 1);
+            }
+            std::mem::swap(&mut prev, &mut cur);
+        }
+        for j in 0..=hay.len() {
+            best = best.min(prev[j]);
+        }
+        if best <= k {
+            Some(best)
+        } else {
+            None
+        }
+    }
+
+    #[test]
+    fn matches_brute_force_oracle() {
+        // Deterministic LCG so the test is reproducible.
+        let mut seed: u64 = 0x2545F4914F6CDD1D;
+        let mut rng = move || {
+            seed = seed.wrapping_mul(6364136223846793005).wrapping_add(1442695040888963407);
+            (seed >> 33) as usize
+        };
+        let alphabet = b"abcx";
+        for _ in 0..500 {
+            let plen = 3 + rng() % 6;
+            let hlen = rng() % 20;
+            let pattern: Vec<u8> = (0..plen).map(|_| alphabet[rng() % 4]).collect();
+            let hay: Vec<u8> = (0..hlen).map(|_| alphabet[rng() % 4]).collect();
+            for k in 0..=4 {
+                let got = Bitap::new(&pattern, k).unwrap().best_distance(&hay);
+                let want = oracle(&pattern, &hay, k);
+                assert_eq!(
+                    got,
+                    want,
+                    "pattern={:?} hay={:?} k={}",
+                    std::str::from_utf8(&pattern),
+                    std::str::from_utf8(&hay),
+                    k
+                );
+            }
+        }
+    }
+}

crates/quicksearch-core/src/search/mod.rs

@@ -0,0 +1,307 @@
+//! Interruptible, streaming search service.
+//!
+//! One dedicated worker thread owns the cascade. The GUI (or any caller)
+//! sends queries via [`SearchService::search`]; results stream back over
+//! an mpsc receiver as [`SearchUpdate`] events tagged with a generation
+//! number. Starting a new search bumps the generation and interrupts the
+//! in-flight SQLite statement, so a keystroke never waits on the previous
+//! query.
+//!
+//! Cancellation is two-layer:
+//! - **cooperative** — the cascade compares its generation against the
+//!   latest every few hundred rows and stops silently when stale;
+//! - **interrupt** — [`rusqlite::InterruptHandle::interrupt`] kills the
+//!   statement currently executing (covering the "no rows produced yet"
+//!   phases like FTS candidate gathering). An interrupted stale search is
+//!   normal cancellation, not an error.
+//!
+//! Consumers that want a plain blocking search (the CLI mode) skip the
+//! service entirely and call [`cascade::run`] with a collecting sink.
+
+pub mod cascade;
+pub mod duplicates;
+pub mod fuzzy;
+
+use std::path::PathBuf;
+use std::sync::atomic::{AtomicU64, Ordering};
+use std::sync::{mpsc, Arc, Mutex};
+use std::thread::JoinHandle;
+
+use crate::db;
+use crate::query::split::split_for_cascade;
+use crate::snippet::Snippet;
+
+pub use cascade::Outcome;
+pub use duplicates::{find_duplicate_groups, DuplicateGroup};
+
+/// One search result. `rank` is the sort key (lower = better): integer
+/// part = cascade stage (1–11), fraction = occurrence-count or
+/// edit-distance tiebreak. Batches arrive already rank-ordered and later
+/// batches only append, so a rank-sorted view never reshuffles.
+#[derive(Debug, Clone)]
+pub struct SearchHit {
+    pub file_id: i64,
+    pub name: String,
+    pub path: String,
+    pub size: u64,
+    pub mtime: i64,
+    pub rank: f64,
+    pub stage: u8,
+    /// The matched span in context: the filename for name stages, the full
+    /// path for path stages, a window of the body for full-text stages
+    /// (absent there when document text isn't stored).
+    pub snippet: Option<Snippet>,
+}
+
+#[derive(Debug, Clone)]
+pub enum SearchUpdate {
+    Started { generation: u64 },
+    Hits { generation: u64, hits: Vec<SearchHit> },
+    Completed { generation: u64, total: usize, limited: bool },
+    Error { generation: u64, message: String },
+}
+
+impl SearchUpdate {
+    pub fn generation(&self) -> u64 {
+        match self {
+            SearchUpdate::Started { generation }
+            | SearchUpdate::Hits { generation, .. }
+            | SearchUpdate::Completed { generation, .. }
+            | SearchUpdate::Error { generation, .. } => *generation,
+        }
+    }
+}
+
+#[derive(Debug, Clone)]
+pub struct SearchOptions {
+    /// Enable the fuzzy stages (ranks 7, 8 and 11).
+    pub fuzzy: bool,
+    /// Ceiling on the fuzzy edit budget (`[search].fuzzy_max_edits`); see
+    /// [`fuzzy::edit_budget`]. 0 disables the fuzzy stages.
+    pub fuzzy_max_edits: usize,
+    /// Hard cap on total hits per search (`[search].display_limit`).
+    pub limit: usize,
+    /// Streaming batch size (`[search].results_per_page`).
+    pub batch: usize,
+    /// Session-scoped ignore patterns (GUI chips), same glob semantics as
+    /// the config's `ignore_patterns`. Applied before the display cap.
+    pub session_ignores: Vec<String>,
+}
+
+impl Default for SearchOptions {
+    fn default() -> Self {
+        SearchOptions {
+            fuzzy: false,
+            fuzzy_max_edits: 2,
+            limit: 1000,
+            batch: 100,
+            session_ignores: Vec::new(),
+        }
+    }
+}
+
+struct SearchRequest {
+    generation: u64,
+    input: String,
+    options: SearchOptions,
+}
+
+pub struct SearchService {
+    req_tx: mpsc::Sender<SearchRequest>,
+    latest_gen: Arc<AtomicU64>,
+    interrupt: Arc<Mutex<Option<rusqlite::InterruptHandle>>>,
+    db_path: Arc<Mutex<PathBuf>>,
+    handle: Option<JoinHandle<()>>,
+}
+
+impl SearchService {
+    /// Spawn the worker. `notify` is invoked after every update event so
+    /// an egui frontend can `request_repaint` (pass a no-op for headless
+    /// use). Returns the service handle plus the update receiver, which
+    /// the caller drains non-blockingly.
+    pub fn new(
+        db_path: PathBuf,
+        notify: Arc<dyn Fn() + Send + Sync>,
+    ) -> (SearchService, mpsc::Receiver<SearchUpdate>) {
+        let (req_tx, req_rx) = mpsc::channel::<SearchRequest>();
+        let (update_tx, update_rx) = mpsc::channel::<SearchUpdate>();
+        let latest_gen = Arc::new(AtomicU64::new(0));
+        let interrupt = Arc::new(Mutex::new(None));
+        let db_path = Arc::new(Mutex::new(db_path));
+
+        let worker = Worker {
+            req_rx,
+            update_tx,
+            notify,
+            latest_gen: latest_gen.clone(),
+            interrupt: interrupt.clone(),
+            db_path: db_path.clone(),
+        };
+        let handle = std::thread::Builder::new()
+            .name("qs-search".into())
+            .spawn(move || worker.run())
+            .expect("spawn search worker");
+
+        (
+            SearchService {
+                req_tx,
+                latest_gen,
+                interrupt,
+                db_path,
+                handle: Some(handle),
+            },
+            update_rx,
+        )
+    }
+
+    /// Start a new search, cancelling any in-flight one. Returns the
+    /// generation whose events to keep.
+    pub fn search(&self, input: &str, options: SearchOptions) -> u64 {
+        let generation = self.latest_gen.fetch_add(1, Ordering::SeqCst) + 1;
+        let _ = self.req_tx.send(SearchRequest {
+            generation,
+            input: input.to_string(),
+            options,
+        });
+        // The new request can't be running yet (the worker hasn't dequeued
+        // it), so this only ever kills a stale generation's statement.
+        self.interrupt_current();
+        generation
+    }
+
+    /// Cancel without starting anything new.
+    pub fn cancel(&self) {
+        self.latest_gen.fetch_add(1, Ordering::SeqCst);
+        self.interrupt_current();
+    }
+
+    /// Point subsequent searches at a different index file.
+    pub fn set_db_path(&self, path: PathBuf) {
+        *self.db_path.lock().unwrap() = path;
+        self.cancel();
+    }
+
+    fn interrupt_current(&self) {
+        if let Ok(guard) = self.interrupt.lock() {
+            if let Some(handle) = guard.as_ref() {
+                handle.interrupt();
+            }
+        }
+    }
+
+    /// Cancel, close the request channel, and join the worker.
+    pub fn shutdown(self) {
+        self.cancel();
+        let SearchService { req_tx, handle, .. } = self;
+        drop(req_tx);
+        if let Some(handle) = handle {
+            let _ = handle.join();
+        }
+    }
+}
+
+/// Map SQLite-level errors to the tagged strings frontends key off.
+/// `DATABASE_CORRUPTED:` drives the GUI's recovery dialog.
+pub fn classify_sql_err(error_msg: &str) -> String {
+    if error_msg.contains("malformed")
+        || error_msg.contains("corrupt")
+        || error_msg.contains("database disk image is malformed")
+    {
+        format!("DATABASE_CORRUPTED: {}", error_msg)
+    } else if error_msg.contains("fts5: syntax error") {
+        "Search syntax error: the search term contains characters that cannot be processed."
+            .to_string()
+    } else {
+        format!("Search failed: {}", error_msg)
+    }
+}
+
+struct Worker {
+    req_rx: mpsc::Receiver<SearchRequest>,
+    update_tx: mpsc::Sender<SearchUpdate>,
+    notify: Arc<dyn Fn() + Send + Sync>,
+    latest_gen: Arc<AtomicU64>,
+    interrupt: Arc<Mutex<Option<rusqlite::InterruptHandle>>>,
+    db_path: Arc<Mutex<PathBuf>>,
+}
+
+impl Worker {
+    fn run(self) {
+        while let Ok(first) = self.req_rx.recv() {
+            // A fast typist queues several requests; only the newest one
+            // matters.
+            let mut req = first;
+            while let Ok(newer) = self.req_rx.try_recv() {
+                req = newer;
+            }
+            if req.generation != self.latest_gen.load(Ordering::SeqCst) {
+                continue;
+            }
+            self.handle(req);
+        }
+    }
+
+    fn send(&self, update: SearchUpdate) {
+        let _ = self.update_tx.send(update);
+        (self.notify)();
+    }
+
+    fn handle(&self, req: SearchRequest) {
+        let generation = req.generation;
+        self.send(SearchUpdate::Started { generation });
+
+        let split = match split_for_cascade(&req.input) {
+            Ok(s) => s,
+            Err(e) => {
+                self.send(SearchUpdate::Error {
+                    generation,
+                    message: e.to_string(),
+                });
+                return;
+            }
+        };
+
+        let db_path = self.db_path.lock().unwrap().clone();
+        // Per-request open: microseconds, and always sees a freshly
+        // rebuilt index file rather than pinning a deleted inode.
+        let conn = match db::open_existing(&db_path.to_string_lossy(), false) {
+            Ok(c) => c,
+            Err(e) => {
+                self.send(SearchUpdate::Error {
+                    generation,
+                    message: classify_sql_err(&e),
+                });
+                return;
+            }
+        };
+        *self.interrupt.lock().unwrap() = Some(conn.get_interrupt_handle());
+
+        let mut sink = |hits: Vec<SearchHit>| {
+            self.send(SearchUpdate::Hits { generation, hits });
+        };
+        let outcome = cascade::run(
+            &conn,
+            &split,
+            &req.options,
+            generation,
+            &self.latest_gen,
+            &mut sink,
+        );
+
+        *self.interrupt.lock().unwrap() = None;
+
+        match outcome {
+            Ok(Some(Outcome { total, limited })) => self.send(SearchUpdate::Completed {
+                generation,
+                total,
+                limited,
+            }),
+            // Cancelled — the newer generation owns the UI now.
+            Ok(None) => {}
+            Err(e) => self.send(SearchUpdate::Error {
+                generation,
+                message: classify_sql_err(&e),
+            }),
+        }
+    }
+}

crates/quicksearch-core/src/search_sql.rs

@@ -1,396 +0,0 @@
-//! SQL builders for the GUI's three legacy search modes (full-text,
-//! filename, duplicate). Lives in core so it has unit-test coverage; the
-//! GUI just composes these into per-page queries.
-//!
-//! For the structured Baloo-subset query language (`type:`, `modified:`,
-//! …), see [`crate::query`].
-
-/// All inputs needed to run (and re-run) one of the three search modes.
-/// Cached after a fresh search so paging buttons don't have to rebuild
-/// from form state.
-#[derive(Debug, Clone, PartialEq)]
-pub struct SearchArgs {
-    pub search_type: String,
-    pub term: String,
-    pub fulltext_exact: bool,
-    pub fulltext_case_sensitive: bool,
-}
-
-/// SQL that counts every row matching `args`, ignoring pagination. Used to
-/// drive the "page X of Y" UI. For very large FTS hit sets this can take a
-/// noticeable fraction of the per-page query time, but it only runs on a
-/// fresh search — page navigation reuses the cached total.
-pub fn build_count(args: &SearchArgs) -> Result<String, String> {
-    match args.search_type.as_str() {
-        "fulltext" => {
-            let where_clause = build_fulltext_where(args)?;
-            Ok(format!(
-                "SELECT COUNT(*) FROM searchabletext AS st WHERE {}",
-                where_clause
-            ))
-        }
-        "filename" => {
-            if args.term.trim().is_empty() {
-                return Err("Please enter a filename pattern".into());
-            }
-            Ok(format!(
-                "SELECT COUNT(*) FROM files WHERE name LIKE '%{}%'",
-                args.term.replace('\'', "''")
-            ))
-        }
-        "duplicates" => Ok(
-            "SELECT COUNT(*) FROM (SELECT 1 FROM files WHERE hash IS NOT NULL \
-             GROUP BY hash HAVING count(*) > 1)"
-                .into(),
-        ),
-        _ => Err("Unknown search type".into()),
-    }
-}
-
-/// SQL that returns one page of results. Columns emitted by the `fulltext`
-/// branch are `(name, path, file_id, text_zstd)` — the snippet is rendered
-/// in Rust from the zstd-compressed `documents_text` row (FTS5 is
-/// contentless, so SQLite's `snippet()` doesn't work on it). The GUI
-/// should use [`crate::indexing::IndexingService::execute_fulltext_search`]
-/// which stitches the decompress + snippet step on top of this SQL.
-pub fn build_select(args: &SearchArgs, limit: u32, offset: u32) -> Result<String, String> {
-    match args.search_type.as_str() {
-        "fulltext" => {
-            let where_clause = build_fulltext_where(args)?;
-            Ok(format!(
-                "SELECT f.name, f.path, f.id, dt.text_zstd \
-                 FROM searchabletext AS st \
-                 JOIN files f ON f.id = st.rowid \
-                 LEFT JOIN documents_text dt ON dt.file_id = f.id \
-                 WHERE {} ORDER BY rank LIMIT {} OFFSET {}",
-                where_clause, limit, offset
-            ))
-        }
-        "filename" => {
-            if args.term.trim().is_empty() {
-                return Err("Please enter a filename pattern".into());
-            }
-            Ok(format!(
-                "SELECT name, path FROM files WHERE name LIKE '%{}%' ORDER BY name LIMIT {} OFFSET {}",
-                args.term.replace('\'', "''"),
-                limit,
-                offset
-            ))
-        }
-        "duplicates" => Ok(format!(
-            "SELECT name, count(*) as cnt, path FROM files WHERE hash IS NOT NULL \
-             GROUP BY hash HAVING cnt > 1 ORDER BY cnt DESC LIMIT {} OFFSET {}",
-            limit, offset
-        )),
-        _ => Err("Unknown search type".into()),
-    }
-}
-
-/// Translate the user-typed term into the FTS5 `MATCH` expression and any
-/// supplemental case-sensitivity filters. Shared by count and select so
-/// pagination doesn't accidentally diverge from the totals.
-fn build_fulltext_where(args: &SearchArgs) -> Result<String, String> {
-    let trimmed = args.term.trim();
-    if trimmed.is_empty() {
-        return Err("Please enter a search term".into());
-    }
-
-    // Strip FTS5 control characters that confuse the parser. Replace with
-    // spaces so word boundaries survive.
-    let sanitized: String = trimmed
-        .chars()
-        .map(|c| {
-            if matches!(
-                c,
-                ':' | ';' | '(' | ')' | '[' | ']' | '{' | '}' | '^' | '~' | '"'
-            ) {
-                ' '
-            } else {
-                c
-            }
-        })
-        .collect();
-
-    let tokens: Vec<&str> = sanitized.split_whitespace().collect();
-    if tokens.is_empty() {
-        return Err("Please enter a valid search term".into());
-    }
-
-    // Trigram tokenizer needs each word to be at least 3 characters. Exact
-    // phrase mode skips this filter — a quoted phrase of short words still
-    // matches because the trigrams overlap the spaces between words.
-    let words: Vec<&str> = if args.fulltext_exact {
-        tokens
-    } else {
-        let filtered: Vec<&str> = tokens
-            .into_iter()
-            .filter(|w| w.chars().count() >= 3)
-            .collect();
-        if filtered.is_empty() {
-            return Err(
-                "Trigram index needs each word to be at least 3 characters unless you use exact phrase search."
-                    .into(),
-            );
-        }
-        filtered
-    };
-
-    let sql_quote = |s: &str| s.replace('\'', "''");
-
-    let fts_match = if args.fulltext_exact {
-        let phrase = words.join(" ");
-        format!("\"{}\"", phrase.replace('"', "\"\""))
-    } else {
-        words.join(" AND ")
-    };
-
-    // Contentless FTS5 doesn't store column text, so case-sensitive
-    // filtering can't live in SQL anymore. It's re-applied in
-    // `IndexingService::execute_fulltext_search` by checking the
-    // decompressed body text for literal-case matches before returning
-    // the row. The MATCH itself stays case-insensitive (tokenizer folds),
-    // which is the correct candidate-set for a post-filter.
-    let where_clause = format!("st.text MATCH '{}'", sql_quote(&fts_match));
-    Ok(where_clause)
-}
-
-/// Pull out the raw words the user typed so the snippet renderer and the
-/// case-sensitive post-filter can see the same tokens `build_fulltext_where`
-/// fed into FTS5. Returns empty when the user's term is empty or contains
-/// only too-short words under the non-exact path.
-pub fn fulltext_terms(args: &SearchArgs) -> Vec<String> {
-    let trimmed = args.term.trim();
-    if trimmed.is_empty() {
-        return Vec::new();
-    }
-    let sanitized: String = trimmed
-        .chars()
-        .map(|c| {
-            if matches!(
-                c,
-                ':' | ';' | '(' | ')' | '[' | ']' | '{' | '}' | '^' | '~' | '"'
-            ) {
-                ' '
-            } else {
-                c
-            }
-        })
-        .collect();
-    let tokens: Vec<String> = sanitized
-        .split_whitespace()
-        .map(|s| s.to_string())
-        .collect();
-    if tokens.is_empty() {
-        return Vec::new();
-    }
-    if args.fulltext_exact {
-        // One composite phrase. Snippet rendering wants to highlight the
-        // whole phrase contiguously; the renderer supports multiple terms
-        // already so we collapse to the joined form.
-        vec![tokens.join(" ")]
-    } else {
-        tokens
-            .into_iter()
-            .filter(|w| w.chars().count() >= 3)
-            .collect()
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-    use crate::db::{open_or_recreate, repo::{insert_file, set_content_done, NewFile}};
-    use crate::mime::FileType;
-
-    fn args(search_type: &str, term: &str) -> SearchArgs {
-        SearchArgs {
-            search_type: search_type.into(),
-            term: term.into(),
-            fulltext_exact: false,
-            fulltext_case_sensitive: false,
-        }
-    }
-
-    #[test]
-    fn fulltext_select_has_limit_and_offset() {
-        let sql = build_select(&args("fulltext", "hello world"), 50, 100).unwrap();
-        assert!(sql.contains("LIMIT 50"));
-        assert!(sql.contains("OFFSET 100"));
-        assert!(sql.contains("ORDER BY rank"));
-        // Snippet rendering moved to Rust; SQL returns the compressed blob
-        // so the post-processor can decompress + highlight.
-        assert!(sql.contains("dt.text_zstd"), "got {sql}");
-    }
-
-    #[test]
-    fn fulltext_case_sensitive_no_longer_in_sql() {
-        let mut a = args("fulltext", "Hello");
-        a.fulltext_case_sensitive = true;
-        let sql = build_select(&a, 50, 0).unwrap();
-        // Post-filter lives in Rust now — no instr() or st.text reference.
-        assert!(!sql.contains("instr("), "got {sql}");
-    }
-
-    #[test]
-    fn fulltext_terms_extract_non_exact() {
-        let a = args("fulltext", "the quick brown");
-        let t = fulltext_terms(&a);
-        // Short words like "the" are dropped (trigram min length 3 applies
-        // in non-exact mode — matches the SQL build rules).
-        assert!(t.iter().any(|s| s == "quick"));
-        assert!(t.iter().any(|s| s == "brown"));
-    }
-
-    #[test]
-    fn fulltext_terms_exact_mode_returns_joined_phrase() {
-        let mut a = args("fulltext", "hello world");
-        a.fulltext_exact = true;
-        let t = fulltext_terms(&a);
-        assert_eq!(t, vec!["hello world".to_string()]);
-    }
-
-    #[test]
-    fn fulltext_count_lacks_pagination_and_join() {
-        let sql = build_count(&args("fulltext", "hello world")).unwrap();
-        assert!(sql.starts_with("SELECT COUNT(*)"));
-        assert!(!sql.contains("LIMIT"));
-        assert!(!sql.contains("OFFSET"));
-        assert!(!sql.contains("JOIN files"));
-    }
-
-    #[test]
-    fn fulltext_short_words_filtered_unless_exact() {
-        let err = build_select(&args("fulltext", "a b"), 50, 0).unwrap_err();
-        assert!(err.contains("3 characters"));
-        let mut a = args("fulltext", "a b");
-        a.fulltext_exact = true;
-        assert!(build_select(&a, 50, 0).is_ok());
-    }
-
-    #[test]
-    fn fulltext_quotes_are_escaped() {
-        let sql = build_select(&args("fulltext", "it's working"), 50, 0).unwrap();
-        // SQL literals double single quotes.
-        assert!(sql.contains("it''s"));
-    }
-
-    #[test]
-    fn filename_select_has_limit_offset_and_order() {
-        let sql = build_select(&args("filename", "report"), 50, 0).unwrap();
-        assert!(sql.contains("LIMIT 50"));
-        assert!(sql.contains("OFFSET 0"));
-        assert!(sql.contains("ORDER BY name"));
-        assert!(sql.contains("name LIKE '%report%'"));
-    }
-
-    #[test]
-    fn filename_empty_term_errors() {
-        assert!(build_select(&args("filename", "   "), 50, 0).is_err());
-        assert!(build_count(&args("filename", "")).is_err());
-    }
-
-    #[test]
-    fn duplicates_select_has_limit_offset() {
-        let sql = build_select(&args("duplicates", ""), 50, 100).unwrap();
-        assert!(sql.contains("LIMIT 50"));
-        assert!(sql.contains("OFFSET 100"));
-        assert!(sql.contains("GROUP BY hash"));
-    }
-
-    #[test]
-    fn unknown_search_type_errors() {
-        assert!(build_select(&args("nope", ""), 50, 0).is_err());
-        assert!(build_count(&args("nope", "")).is_err());
-    }
-
-    fn tmp_path() -> std::path::PathBuf {
-        let mut p = std::env::temp_dir();
-        p.push(format!(
-            "qs-search-sql-{}-{}.sqlite",
-            std::process::id(),
-            std::time::SystemTime::now()
-                .duration_since(std::time::UNIX_EPOCH)
-                .unwrap()
-                .as_nanos()
-        ));
-        p
-    }
-
-    /// End-to-end: seed three rows, run count + paged select, verify
-    /// pagination boundaries actually behave on a real DB.
-    #[test]
-    fn end_to_end_pagination_smoke() {
-        let p = tmp_path();
-        let mut conn = open_or_recreate(p.to_str().unwrap(), "trigram").unwrap();
-        {
-            let tx = conn.transaction().unwrap();
-            for i in 0..7 {
-                let path = format!("/tmp/file_{}.txt", i);
-                let id = insert_file(
-                    &tx,
-                    &NewFile {
-                        name: &format!("file_{}.txt", i),
-                        path: &path,
-                        parent: "/tmp",
-                        size: 1,
-                        mtime: 1,
-                        inode: None,
-                        device_id: None,
-                        mime: Some("text/plain"),
-                        ftype: FileType::TEXT,
-                        hash: None,
-                    },
-                )
-                .unwrap()
-                .expect("unique path");
-                set_content_done(&tx, id, &format!("file_{}.txt", i), "shared body content", &[], true).unwrap();
-            }
-            tx.commit().unwrap();
-        }
-
-        // Count: all 7 rows match "shared".
-        let count_sql = build_count(&args("fulltext", "shared body content")).unwrap();
-        let n: i64 = conn.query_row(&count_sql, [], |r| r.get(0)).unwrap();
-        assert_eq!(n, 7);
-
-        // Page 1, page_size 3 → 3 rows.
-        let sel1 = build_select(&args("fulltext", "shared body content"), 3, 0).unwrap();
-        let rows1: Vec<String> = conn
-            .prepare(&sel1)
-            .unwrap()
-            .query_map([], |r| r.get::<_, String>(1))
-            .unwrap()
-            .map(|r| r.unwrap())
-            .collect();
-        assert_eq!(rows1.len(), 3);
-
-        // Page 3, page_size 3 → 1 row (offset 6, 7 total).
-        let sel3 = build_select(&args("fulltext", "shared body content"), 3, 6).unwrap();
-        let rows3: Vec<String> = conn
-            .prepare(&sel3)
-            .unwrap()
-            .query_map([], |r| r.get::<_, String>(1))
-            .unwrap()
-            .map(|r| r.unwrap())
-            .collect();
-        assert_eq!(rows3.len(), 1);
-
-        // Filename pagination on the same fixture.
-        let fn_count = build_count(&args("filename", "file_")).unwrap();
-        let n2: i64 = conn.query_row(&fn_count, [], |r| r.get(0)).unwrap();
-        assert_eq!(n2, 7);
-
-        let fn_sel = build_select(&args("filename", "file_"), 5, 0).unwrap();
-        let rows: Vec<String> = conn
-            .prepare(&fn_sel)
-            .unwrap()
-            .query_map([], |r| r.get::<_, String>(1))
-            .unwrap()
-            .map(|r| r.unwrap())
-            .collect();
-        assert_eq!(rows.len(), 5);
-
-        drop(conn);
-        std::fs::remove_file(&p).ok();
-    }
-}

crates/quicksearch-core/src/shutdown.rs

@@ -1,26 +1,23 @@
 //! Process-level shutdown helpers.
 //!
-//! Wires Ctrl-C (and on Unix, SIGTERM) to a graceful shutdown that flushes
+//! Wires Ctrl-C (and on Unix, SIGTERM) to a graceful shutdown that stops
-//! the indexing DB and exits. Replaces the ad-hoc `ctrlc::set_handler` +
+//! the watcher, aborts any running index pass, flushes the WAL, and
-//! `OnceLock` dance the GUI used to carry. Call
+//! exits. Call [`install_signal_handler`] once from a binary's `main`
-//! [`install_signal_handler`] once from a binary's `main` with a cloned
+//! with a cloned [`IndexCoordinator`] handle.
-//! [`IndexingService`] handle.
 
 use std::sync::Arc;
 
-use crate::indexing::IndexingService;
+use crate::coordinator::IndexCoordinator;
 
 /// Install a Ctrl-C (and, where supported, SIGTERM) handler that calls
-/// [`IndexingService::graceful_shutdown`] and then exits with status 0.
+/// [`IndexCoordinator::shutdown`] and then exits with status 0.
 ///
 /// Returns an error only if a handler was already installed elsewhere in
 /// this process (ctrlc::set_handler is one-shot).
-pub fn install_signal_handler(service: Arc<IndexingService>) -> Result<(), String> {
+pub fn install_signal_handler(coordinator: Arc<IndexCoordinator>) -> Result<(), String> {
     ctrlc::set_handler(move || {
-        eprintln!("Received Ctrl-C, shutting down gracefully...");
+        crate::log_info!("Received Ctrl-C, shutting down gracefully...");
-        if let Err(e) = service.graceful_shutdown() {
+        coordinator.shutdown();
-            eprintln!("Error during graceful shutdown: {}", e);
-        }
         std::process::exit(0);
     })
     .map_err(|e| format!("install signal handler: {}", e))

crates/quicksearch-core/src/snippet.rs

@@ -1,60 +1,70 @@
-//! Snippet / highlight rendering for search results.
+//! Snippet extraction for search results.
 //!
 //! We store extracted text in the `documents_text` sidecar (zstd-compressed)
 //! rather than in FTS5, so SQLite's built-in `snippet()` / `highlight()`
 //! auxiliary functions aren't available (contentless FTS5 doesn't support
-//! them). This module reproduces the parts we actually need in Rust: find
+//! them). This module reproduces the parts we actually need in Rust: find a
-//! a window of context around the first match, bold every query-term
+//! window of context around the first match and report every query-term
-//! occurrence inside that window, trim with ellipsis markers.
+//! occurrence inside that window.
 //!
-//! Matching is ASCII-case-insensitive on the *rendering* side. That aligns
+//! Output is *structural* — the window text plus byte ranges of the matches
-//! with the search path which is already case-insensitive via the trigram
+//! within it — so any frontend can render highlights natively (egui builds
-//! tokenizer; exact-case-only snippets aren't a feature users expect here.
+//! a `LayoutJob`, the CLI emits ANSI bold). Nothing here produces markup.
-//! Unicode accent folding isn't applied at the rendering layer — a query
-//! for `cafe` will still *find* a file containing `café` (because the FTS
-//! tokenizer strips diacritics) but the snippet won't highlight the
-//! accented occurrence. The surrounding text is still returned verbatim.
 //!
-//! The API is intentionally small: one `render` function plus an `Options`
+//! Matching is ASCII-case-insensitive. That aligns with the search path,
-//! struct. Callers that want different pre/post tags, ellipsis, or window
+//! which is already case-insensitive via the trigram tokenizer; Unicode
-//! size pass them in; there are sensible defaults for the GUI case.
+//! accent folding isn't applied at this layer — a query for `cafe` will
+//! still *find* a file containing `café` (the FTS tokenizer strips
+//! diacritics) but the snippet won't mark the accented occurrence. The
+//! window text is returned verbatim either way.
 
-/// Options controlling snippet rendering. The defaults mirror the old
+/// Options controlling snippet extraction.
-/// `snippet(searchabletext, 1, '<b>', '</b>', '<b>...</b>', 64)` call that
-/// the GUI used to run directly as SQL.
 #[derive(Debug, Clone)]
-pub struct Options<'a> {
+pub struct Options {
-    pub pre: &'a str,
+    /// Approximate byte budget for the returned window. Matches expand the
-    pub post: &'a str,
+    /// window if needed so a hit is never cut off; the budget is a soft
-    pub ellipsis: &'a str,
+    /// target, not a hard cap.
-    /// Approximate character budget for the returned snippet. Matches
-    /// expand the window if needed to keep their tags on; the budget is a
-    /// soft target, not a hard cap.
     pub approx_chars: usize,
 }
 
-impl<'a> Default for Options<'a> {
+impl Default for Options {
     fn default() -> Self {
-        Options {
+        Options { approx_chars: 200 }
-            pre: "<b>",
+    }
-            post: "</b>",
+}
-            ellipsis: "<b>...</b>",
+
-            approx_chars: 200,
+/// A context window from a document plus the match positions inside it.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct Snippet {
+    /// The excerpt, sliced verbatim from the source text on UTF-8 char
+    /// boundaries.
+    pub window: String,
+    /// Byte ranges *into `window`*, sorted, non-overlapping (overlapping
+    /// term hits are coalesced), always on char boundaries.
+    pub ranges: Vec<(usize, usize)>,
+    /// Content exists before/after the window — frontends render their own
+    /// ellipsis.
+    pub truncated_start: bool,
+    pub truncated_end: bool,
+}
+
+impl Snippet {
+    fn empty() -> Snippet {
+        Snippet {
+            window: String::new(),
+            ranges: Vec::new(),
+            truncated_start: false,
+            truncated_end: false,
         }
     }
 }
 
-/// Render a snippet from `text` highlighting every occurrence of any term
+/// Extract a snippet from `text` marking every occurrence of any term in
-/// in `terms`. Returns a string with `pre`/`post` wrapping each match, and
+/// `terms` (ASCII-case-insensitive). With no terms or no matches, returns
-/// `ellipsis` prepended/appended when the returned window doesn't reach
+/// the head of the text as the window with no ranges.
-/// the text's edges.
+pub fn extract(text: &str, terms: &[&str], opts: &Options) -> Snippet {
-///
-/// If no term matches, returns the first `approx_chars` of `text` (char-
-/// aligned), suffixed with `ellipsis` when truncated.
-pub fn render(text: &str, terms: &[&str], opts: &Options<'_>) -> String {
-    // Short-circuit trivial inputs.
     if text.is_empty() {
-        return String::new();
+        return Snippet::empty();
     }
     let effective_terms: Vec<&str> = terms
         .iter()
@@ -62,32 +72,24 @@ pub fn render(text: &str, terms: &[&str], opts: &Options<'_>) -> String {
         .filter(|t| !t.is_empty())
         .collect();
     if effective_terms.is_empty() {
-        return truncate_head(text, opts.approx_chars, opts.ellipsis);
+        return head_window(text, opts.approx_chars);
     }
 
-    // Case-fold once; we do all positioning on the folded buffer and emit
+    // Case-fold once; we do all positioning on the folded buffer and slice
-    // slices from the original. Both buffers have identical byte layout
+    // from the original. Both buffers have identical byte layout because
-    // because `to_ascii_lowercase` is a byte-for-byte map that preserves
+    // `to_ascii_lowercase` only touches ASCII letters.
-    // multi-byte UTF-8 sequences unchanged (it only touches ASCII letters).
     let folded = text.to_ascii_lowercase();
     let folded_bytes = folded.as_bytes();
 
     let mut matches: Vec<(usize, usize)> = Vec::new();
     for term in &effective_terms {
-        let pattern: String = term.to_ascii_lowercase();
+        let pattern = term.to_ascii_lowercase();
         let pbytes = pattern.as_bytes();
-        if pbytes.is_empty() {
-            continue;
-        }
         let mut start = 0;
         while start + pbytes.len() <= folded_bytes.len() {
             if let Some(rel) = memfind(&folded_bytes[start..], pbytes) {
                 let at = start + rel;
                 matches.push((at, at + pbytes.len()));
-                // Advance past this match to avoid zero-width loops on
-                // empty patterns (already guarded above) and to allow
-                // overlapping matches of *different* terms in the next
-                // outer-loop iteration.
                 start = at + pbytes.len();
             } else {
                 break;
@@ -96,24 +98,18 @@ pub fn render(text: &str, terms: &[&str], opts: &Options<'_>) -> String {
     }
 
     if matches.is_empty() {
-        return truncate_head(text, opts.approx_chars, opts.ellipsis);
+        return head_window(text, opts.approx_chars);
     }
 
-    // Dedupe + sort so overlapping matches from different terms (e.g.
-    // "rust" and "rustc") don't produce nested tags.
     matches.sort_by_key(|(a, _)| *a);
-    matches = coalesce_overlapping(matches);
+    let matches = coalesce_overlapping(matches);
 
     // Pick the window. Start a third of the budget before the first match
-    // so the hit isn't pinned to the left edge. Round both ends to char
+    // so the hit isn't pinned to the left edge; round both ends to char
     // boundaries so we never slice a multi-byte UTF-8 sequence.
     let pre_pad = opts.approx_chars / 3;
-    let first_match_start = matches[0].0;
+    let mut win_start = matches[0].0.saturating_sub(pre_pad);
-    let mut win_start = first_match_start.saturating_sub(pre_pad);
+    let mut win_end = (win_start + opts.approx_chars).min(text.len());
-    let mut win_end = win_start + opts.approx_chars;
-    if win_end > text.len() {
-        win_end = text.len();
-    }
     while win_start > 0 && !text.is_char_boundary(win_start) {
         win_start -= 1;
     }
@@ -121,11 +117,9 @@ pub fn render(text: &str, terms: &[&str], opts: &Options<'_>) -> String {
         win_end += 1;
     }
 
-    // Expand the window to include the full end of any match that would
+    // Expand the window so a match straddling the right edge is fully
-    // otherwise be cut off mid-tag. Keeps rendering sane when a long term
+    // included rather than cut mid-hit.
-    // sits at the right edge of the budget.
+    if let Some((_, end)) = matches.iter().rfind(|(s, _)| *s < win_end) {
-    let last_match_in_window = matches.iter().rfind(|(s, _)| *s < win_end);
-    if let Some((_, end)) = last_match_in_window {
         if *end > win_end {
             win_end = *end;
             while win_end < text.len() && !text.is_char_boundary(win_end) {
@@ -134,55 +128,114 @@ pub fn render(text: &str, terms: &[&str], opts: &Options<'_>) -> String {
         }
     }
 
-    // Render: walk matches that fall inside the window, splicing pre/post
+    let ranges = matches
-    // around each. Prepend/append ellipsis when we've chopped off content.
+        .iter()
-    let mut out = String::with_capacity(win_end - win_start + 32);
+        .filter(|(s, e)| *e > win_start && *s < win_end)
-    if win_start > 0 {
+        .map(|(s, e)| ((*s).max(win_start) - win_start, (*e).min(win_end) - win_start))
-        out.push_str(opts.ellipsis);
+        .collect();
+
+    Snippet {
+        window: text[win_start..win_end].to_string(),
+        ranges,
+        truncated_start: win_start > 0,
+        truncated_end: win_end < text.len(),
     }
-    let mut cursor = win_start;
-    for (ms, me) in matches.iter() {
-        if *me <= win_start || *ms >= win_end {
-            continue;
-        }
-        // Clamp to the window.
-        let ms = (*ms).max(win_start);
-        let me = (*me).min(win_end);
-        if ms > cursor {
-            out.push_str(&text[cursor..ms]);
-        }
-        out.push_str(opts.pre);
-        out.push_str(&text[ms..me]);
-        out.push_str(opts.post);
-        cursor = me;
-    }
-    if cursor < win_end {
-        out.push_str(&text[cursor..win_end]);
-    }
-    if win_end < text.len() {
-        out.push_str(opts.ellipsis);
-    }
-    out
 }
 
-/// Return the first `n` characters of `text`, suffixed with `ellipsis` if
+/// Build a snippet window around one known match range in `text` (byte
-/// truncation actually happened. Respects UTF-8 char boundaries.
+/// offsets into `text`). Used by fuzzy full-text search, where the match
-fn truncate_head(text: &str, n: usize, ellipsis: &str) -> String {
+/// was located by the fuzzy matcher rather than exact term search. The
+/// range is clamped and char-boundary-aligned defensively.
+pub fn window_around(text: &str, range: (usize, usize), opts: &Options) -> Snippet {
+    if text.is_empty() {
+        return Snippet::empty();
+    }
+    let (mut ms, mut me) = range;
+    ms = ms.min(text.len());
+    me = me.clamp(ms, text.len());
+    while ms > 0 && !text.is_char_boundary(ms) {
+        ms -= 1;
+    }
+    while me < text.len() && !text.is_char_boundary(me) {
+        me += 1;
+    }
+
+    let pre_pad = opts.approx_chars / 3;
+    let mut win_start = ms.saturating_sub(pre_pad);
+    let mut win_end = (win_start + opts.approx_chars).max(me).min(text.len());
+    while win_start > 0 && !text.is_char_boundary(win_start) {
+        win_start -= 1;
+    }
+    while win_end < text.len() && !text.is_char_boundary(win_end) {
+        win_end += 1;
+    }
+
+    let ranges = if me > ms {
+        vec![(ms - win_start, me - win_start)]
+    } else {
+        Vec::new()
+    };
+    Snippet {
+        window: text[win_start..win_end].to_string(),
+        ranges,
+        truncated_start: win_start > 0,
+        truncated_end: win_end < text.len(),
+    }
+}
+
+/// Count non-overlapping occurrences of `term` in `text`. Empty terms count
+/// zero. Case-insensitive counting folds ASCII only, matching the rest of
+/// the search pipeline.
+pub fn count_occurrences(text: &str, term: &str, case_sensitive: bool) -> usize {
+    if term.is_empty() || text.len() < term.len() {
+        return 0;
+    }
+    let (hay, needle);
+    let (hay_ref, needle_ref): (&[u8], &[u8]) = if case_sensitive {
+        (text.as_bytes(), term.as_bytes())
+    } else {
+        hay = text.to_ascii_lowercase();
+        needle = term.to_ascii_lowercase();
+        (hay.as_bytes(), needle.as_bytes())
+    };
+    let mut count = 0;
+    let mut start = 0;
+    while start + needle_ref.len() <= hay_ref.len() {
+        match memfind(&hay_ref[start..], needle_ref) {
+            Some(rel) => {
+                count += 1;
+                start += rel + needle_ref.len();
+            }
+            None => break,
+        }
+    }
+    count
+}
+
+/// The first `n` bytes of `text` (char-aligned) as a match-less window.
+fn head_window(text: &str, n: usize) -> Snippet {
     if text.len() <= n {
-        return text.to_string();
+        return Snippet {
+            window: text.to_string(),
+            ranges: Vec::new(),
+            truncated_start: false,
+            truncated_end: false,
+        };
     }
     let mut cut = n;
     while cut > 0 && !text.is_char_boundary(cut) {
         cut -= 1;
     }
-    let mut out = String::with_capacity(cut + ellipsis.len());
+    Snippet {
-    out.push_str(&text[..cut]);
+        window: text[..cut].to_string(),
-    out.push_str(ellipsis);
+        ranges: Vec::new(),
-    out
+        truncated_start: false,
+        truncated_end: true,
+    }
 }
 
-/// Merge adjacent / overlapping (start, end) ranges in place. Input must be
+/// Merge adjacent / overlapping (start, end) ranges. Input must be sorted
-/// sorted by start.
+/// by start.
 fn coalesce_overlapping(mut v: Vec<(usize, usize)>) -> Vec<(usize, usize)> {
     if v.len() < 2 {
         return v;
@@ -202,7 +255,7 @@ fn coalesce_overlapping(mut v: Vec<(usize, usize)>) -> Vec<(usize, usize)> {
 }
 
 /// Locate the first occurrence of `needle` in `hay`. A byte-level search;
-/// callers have already lowercased both sides so case is normalized.
+/// callers have already normalized case where needed.
 fn memfind(hay: &[u8], needle: &[u8]) -> Option<usize> {
     if needle.is_empty() || needle.len() > hay.len() {
         return None;
@@ -222,115 +275,170 @@ fn memfind(hay: &[u8], needle: &[u8]) -> Option<usize> {
 mod tests {
     use super::*;
 
-    fn opts_small() -> Options<'static> {
+    fn opts_small() -> Options {
-        Options {
+        Options { approx_chars: 40 }
-            pre: "<b>",
+    }
-            post: "</b>",
+
-            ellipsis: "…",
+    /// Every range must be in-bounds, ordered, non-overlapping, and sit on
-            approx_chars: 40,
+    /// char boundaries — the contract egui's LayoutJob sections rely on.
+    fn assert_ranges_valid(s: &Snippet) {
+        let mut prev_end = 0;
+        for &(a, b) in &s.ranges {
+            assert!(a < b, "empty/inverted range {:?}", (a, b));
+            assert!(b <= s.window.len(), "range {:?} beyond window", (a, b));
+            assert!(a >= prev_end, "overlapping ranges");
+            assert!(s.window.is_char_boundary(a) && s.window.is_char_boundary(b));
+            prev_end = b;
         }
     }
 
+    fn marked(s: &Snippet) -> Vec<&str> {
+        s.ranges.iter().map(|&(a, b)| &s.window[a..b]).collect()
+    }
+
     #[test]
     fn empty_text_returns_empty() {
-        let s = render("", &["foo"], &Options::default());
+        let s = extract("", &["foo"], &Options::default());
-        assert_eq!(s, "");
+        assert_eq!(s, Snippet::empty());
     }
 
     #[test]
-    fn no_terms_returns_head_with_ellipsis_when_truncated() {
+    fn no_terms_returns_head_marked_truncated() {
         let long = "abcdefghijklmnop".repeat(10);
-        let s = render(&long, &[], &opts_small());
+        let s = extract(&long, &[], &opts_small());
-        assert!(s.ends_with("…"));
+        assert!(s.truncated_end);
-        assert!(s.len() < long.len() + 4);
+        assert!(!s.truncated_start);
+        assert!(s.ranges.is_empty());
+        assert!(s.window.len() <= 40);
     }
 
     #[test]
-    fn no_terms_untruncated_has_no_ellipsis() {
+    fn no_terms_untruncated() {
-        let s = render("short text", &[], &opts_small());
+        let s = extract("short text", &[], &opts_small());
-        assert_eq!(s, "short text");
+        assert_eq!(s.window, "short text");
+        assert!(!s.truncated_end && !s.truncated_start);
     }
 
     #[test]
-    fn simple_highlight_wraps_matches() {
+    fn simple_match_range() {
-        let s = render("the quick brown fox", &["quick"], &opts_small());
+        let s = extract("the quick brown fox", &["quick"], &opts_small());
-        assert!(s.contains("<b>quick</b>"));
+        assert_eq!(marked(&s), vec!["quick"]);
+        assert_ranges_valid(&s);
     }
 
     #[test]
-    fn case_insensitive_match() {
+    fn case_insensitive_match_reports_original_case() {
-        let s = render("The QUICK brown fox", &["quick"], &opts_small());
+        let s = extract("The QUICK brown fox", &["quick"], &opts_small());
-        assert!(s.contains("<b>QUICK</b>"), "got {s}");
+        assert_eq!(marked(&s), vec!["QUICK"]);
     }
 
     #[test]
-    fn multiple_terms_both_highlighted() {
+    fn multiple_terms_both_marked() {
-        let s = render(
+        let s = extract(
             "the quick brown fox jumps over the lazy dog",
             &["quick", "lazy"],
-            &Options {
+            &Options::default(),
-                approx_chars: 200,
-                ..Options::default()
-            },
         );
-        assert!(s.contains("<b>quick</b>"), "got {s}");
+        assert_eq!(marked(&s), vec!["quick", "lazy"]);
-        assert!(s.contains("<b>lazy</b>"), "got {s}");
+        assert_ranges_valid(&s);
     }
 
     #[test]
-    fn window_trims_with_ellipsis_on_both_sides() {
+    fn window_truncation_flags_on_both_sides() {
-        let text =
+        let text = "prefix ".repeat(20) + "MATCH in middle " + &"suffix ".repeat(20);
-            "prefix ".repeat(20) + "MATCH in middle " + &"suffix ".repeat(20);
+        let s = extract(&text, &["MATCH"], &opts_small());
-        let s = render(&text, &["MATCH"], &opts_small());
+        assert!(s.truncated_start);
-        assert!(s.starts_with("…"), "got {s}");
+        assert!(s.truncated_end);
-        assert!(s.ends_with("…"), "got {s}");
+        assert_eq!(marked(&s), vec!["MATCH"]);
-        assert!(s.contains("<b>MATCH</b>"), "got {s}");
     }
 
     #[test]
-    fn match_at_start_has_no_leading_ellipsis() {
+    fn match_at_start_not_truncated_left() {
-        let s = render("MATCH right at the start of this paragraph", &["match"], &opts_small());
+        let s = extract(
-        assert!(!s.starts_with("…"), "got {s}");
+            "MATCH right at the start of this paragraph",
+            &["match"],
+            &opts_small(),
+        );
+        assert!(!s.truncated_start);
+        assert_eq!(s.ranges[0].0, 0);
     }
 
     #[test]
-    fn no_match_on_tail_returns_head() {
+    fn no_match_returns_head_without_ranges() {
         let text = "alpha beta gamma delta epsilon zeta eta theta iota kappa";
-        let s = render(text, &["nomatch"], &opts_small());
+        let s = extract(text, &["nomatch"], &opts_small());
-        assert!(!s.contains("<b>"));
+        assert!(s.ranges.is_empty());
-        assert!(s.starts_with("alpha"));
+        assert!(s.window.starts_with("alpha"));
     }
 
     #[test]
-    fn overlapping_terms_do_not_nest_tags() {
+    fn overlapping_terms_coalesce() {
-        // Two terms matching the same span must coalesce.
+        let s = extract("the RUSTC compiler", &["rust", "rustc"], &opts_small());
-        let s = render("the RUSTC compiler", &["rust", "rustc"], &opts_small());
+        assert_eq!(marked(&s), vec!["RUSTC"]);
-        assert!(s.contains("<b>RUSTC</b>"), "got {s}");
+        assert_ranges_valid(&s);
-        // No nested <b> tags.
-        assert!(!s.contains("<b><b>"), "got {s}");
     }
 
     #[test]
-    fn utf8_boundary_safe_truncation() {
+    fn utf8_boundaries_hold_with_multibyte_text() {
-        // Insert multi-byte chars near the window boundary.
         let text = "café café café café café café café café café café";
-        let s = render(text, &["nope"], &opts_small());
+        let s = extract(text, &["café"], &opts_small());
-        // Returned string must be valid UTF-8 (push_str guarantees this only
+        assert!(!s.ranges.is_empty());
-        // if we sliced on char boundaries). Assert by round-trip.
+        assert_ranges_valid(&s);
-        assert_eq!(s.as_str(), &s.clone());
+        for m in marked(&s) {
+            assert_eq!(m, "café");
+        }
     }
 
     #[test]
-    fn match_near_right_edge_is_fully_shown() {
+    fn match_near_right_edge_is_fully_included() {
         let prefix = "x".repeat(30);
         let text = format!("{}{}", prefix, "LONGMATCHTERMTEXT");
-        let s = render(&text, &["LONGMATCHTERMTEXT"], &opts_small());
+        let s = extract(&text, &["LONGMATCHTERMTEXT"], &opts_small());
-        assert!(s.contains("<b>LONGMATCHTERMTEXT</b>"), "got {s}");
+        assert_eq!(marked(&s), vec!["LONGMATCHTERMTEXT"]);
     }
 
     #[test]
     fn empty_query_term_ignored() {
-        let s = render("hello world", &["", "world"], &opts_small());
+        let s = extract("hello world", &["", "world"], &opts_small());
-        assert!(s.contains("<b>world</b>"), "got {s}");
+        assert_eq!(marked(&s), vec!["world"]);
+    }
+
+    #[test]
+    fn window_around_basic() {
+        let text = "prefix ".repeat(20) + "NEEDLE" + &" suffix".repeat(20);
+        let at = text.find("NEEDLE").unwrap();
+        let s = window_around(&text, (at, at + 6), &opts_small());
+        assert_eq!(marked(&s), vec!["NEEDLE"]);
+        assert!(s.truncated_start && s.truncated_end);
+        assert_ranges_valid(&s);
+    }
+
+    #[test]
+    fn window_around_clamps_out_of_bounds() {
+        let s = window_around("tiny", (2, 999), &opts_small());
+        assert_eq!(s.window, "tiny");
+        assert_eq!(s.ranges, vec![(2, 4)]);
+        // Fully out-of-range → no ranges, but never a panic.
+        let s = window_around("tiny", (999, 1000), &opts_small());
+        assert!(s.ranges.is_empty());
+    }
+
+    #[test]
+    fn window_around_aligns_multibyte_boundaries() {
+        let text = "ééééééééé needle ééééééééé";
+        // Deliberately mis-aligned offsets inside multi-byte sequences.
+        let s = window_around(text, (1, 3), &opts_small());
+        assert_ranges_valid(&s);
+    }
+
+    #[test]
+    fn count_occurrences_cases() {
+        assert_eq!(count_occurrences("aaaa", "aaa", true), 1, "non-overlapping");
+        assert_eq!(count_occurrences("abcABC", "abc", true), 1);
+        assert_eq!(count_occurrences("abcABC", "abc", false), 2);
+        assert_eq!(count_occurrences("", "x", true), 0);
+        assert_eq!(count_occurrences("xyz", "", true), 0);
+        assert_eq!(count_occurrences("no hits here", "zzz", false), 0);
+        assert_eq!(count_occurrences("ab ab ab", "ab", true), 3);
     }
 }

crates/quicksearch-core/tests/full_index.rs

@@ -0,0 +1,636 @@
+//! End-to-end phase-1 tests over a real tree and a real database.
+//!
+//! These cover the failure mode that unit tests structurally cannot: a full
+//! run deletes index rows for every path it did not see, so any walk that
+//! quietly reports less than it should destroys data. That damage is
+//! invisible on a first index — `existing_files` is empty, so nothing is
+//! stale — and only appears on the second run.
+
+use std::path::{Path, PathBuf};
+use std::time::{Duration, Instant, SystemTime, UNIX_EPOCH};
+
+use quicksearch_core::config::Config;
+use quicksearch_core::indexing::{IndexingService, IndexingStatus};
+
+fn tmp_dir(tag: &str) -> PathBuf {
+    let mut p = std::env::temp_dir();
+    p.push(format!(
+        "quicksearch-e2e-{}-{}-{}",
+        tag,
+        std::process::id(),
+        SystemTime::now().duration_since(UNIX_EPOCH).unwrap().as_nanos()
+    ));
+    std::fs::create_dir_all(&p).unwrap();
+    p
+}
+
+fn touch(p: &Path, body: &[u8]) {
+    std::fs::create_dir_all(p.parent().unwrap()).unwrap();
+    std::fs::write(p, body).unwrap();
+}
+
+/// Run one full index and wait for it to finish.
+///
+/// Completion is detected via the `last_full_index` marker, which
+/// `run_indexing` writes only on a successful finish. Polling the status
+/// enum instead would race: a small tree finishes between two polls, so
+/// `Idle` is ambiguous between "not started yet" and "already done".
+fn index_once(root: &Path, db: &Path, config: &Config) {
+    if db.exists() {
+        let conn = rusqlite::Connection::open(db).unwrap();
+        conn.execute("DELETE FROM schema_info WHERE key = 'last_full_index'", [])
+            .unwrap();
+    }
+
+    let service = IndexingService::new();
+    service
+        .start_indexing(
+            vec![root.to_string_lossy().into_owned()],
+            db.to_string_lossy().into_owned(),
+            config.clone(),
+        )
+        .unwrap();
+
+    let deadline = Instant::now() + Duration::from_secs(120);
+    let mut done = false;
+    while Instant::now() < deadline {
+        if let IndexingStatus::Error(e) = service.get_status() {
+            panic!("indexing failed: {}", e);
+        }
+        if db.exists() {
+            if let Ok(conn) = rusqlite::Connection::open(db) {
+                if quicksearch_core::db::repo::get_last_full_index(&conn).is_some() {
+                    done = true;
+                    break;
+                }
+            }
+        }
+        std::thread::sleep(Duration::from_millis(10));
+    }
+    assert!(done, "indexing did not finish within the timeout");
+    service.stop_indexing().unwrap();
+}
+
+/// (path, mtime, content_state) for every indexed row, ordered by path.
+fn rows(db: &Path) -> Vec<(String, i64, i64)> {
+    let conn = rusqlite::Connection::open(db).unwrap();
+    let mut stmt = conn
+        .prepare("SELECT path, mtime, content_state FROM files ORDER BY path")
+        .unwrap();
+    let out = stmt
+        .query_map([], |r| Ok((r.get(0)?, r.get(1)?, r.get(2)?)))
+        .unwrap()
+        .map(|r| r.unwrap())
+        .collect();
+    out
+}
+
+fn test_config() -> Config {
+    let config = Config::default();
+    // Keep the run to phase 1 semantics we're asserting on; extraction is
+    // covered elsewhere.
+    config
+}
+
+#[test]
+fn reindexing_an_unchanged_tree_changes_nothing() {
+    let root = tmp_dir("stable");
+    let db_dir = tmp_dir("stable-db");
+    let db = db_dir.join("index.sqlite");
+    let config = test_config();
+
+    touch(&root.join("a.txt"), b"alpha");
+    touch(&root.join("sub/b.txt"), b"bravo");
+    touch(&root.join("sub/deep/c.txt"), b"charlie");
+    touch(&root.join("other/d.md"), b"delta");
+
+    index_once(&root, &db, &config);
+    let first = rows(&db);
+    assert_eq!(first.len(), 4, "all four files indexed");
+
+    index_once(&root, &db, &config);
+    let second = rows(&db);
+
+    // The whole point: a second run over an unchanged tree must not delete
+    // and re-insert anything. A wiped-and-rebuilt row would come back with
+    // content_state reset, throwing away extracted text for no reason.
+    assert_eq!(first, second, "an unchanged tree must re-index to an identical set");
+
+    std::fs::remove_dir_all(&root).ok();
+    std::fs::remove_dir_all(&db_dir).ok();
+}
+
+#[test]
+fn deleted_files_are_removed_and_new_ones_added() {
+    let root = tmp_dir("churn");
+    let db_dir = tmp_dir("churn-db");
+    let db = db_dir.join("index.sqlite");
+    let config = test_config();
+
+    touch(&root.join("keep.txt"), b"keep");
+    touch(&root.join("remove.txt"), b"remove");
+    index_once(&root, &db, &config);
+    assert_eq!(rows(&db).len(), 2);
+
+    std::fs::remove_file(root.join("remove.txt")).unwrap();
+    touch(&root.join("added.txt"), b"added");
+    index_once(&root, &db, &config);
+
+    let names: Vec<String> = rows(&db)
+        .into_iter()
+        .map(|(p, _, _)| Path::new(&p).file_name().unwrap().to_string_lossy().into_owned())
+        .collect();
+    assert_eq!(names, vec!["added.txt", "keep.txt"], "stale cleanup still works");
+
+    std::fs::remove_dir_all(&root).ok();
+    std::fs::remove_dir_all(&db_dir).ok();
+}
+
+#[test]
+fn a_modified_file_is_updated_in_place() {
+    let root = tmp_dir("modify");
+    let db_dir = tmp_dir("modify-db");
+    let db = db_dir.join("index.sqlite");
+    let config = test_config();
+
+    let target = root.join("doc.txt");
+    touch(&target, b"first");
+    index_once(&root, &db, &config);
+    let before = rows(&db);
+    assert_eq!(before.len(), 1);
+
+    // Filesystem mtime has one-second granularity in the stored value, so
+    // move it decisively rather than racing it.
+    touch(&target, b"second body, clearly different");
+    let later = SystemTime::now() + Duration::from_secs(5);
+    filetime_set(&target, later);
+
+    index_once(&root, &db, &config);
+    let after = rows(&db);
+    assert_eq!(after.len(), 1, "still exactly one row");
+    assert_ne!(before[0].1, after[0].1, "mtime was refreshed");
+    assert_eq!(before[0].0, after[0].0, "same path");
+
+    std::fs::remove_dir_all(&root).ok();
+    std::fs::remove_dir_all(&db_dir).ok();
+}
+
+/// Set a file's mtime without pulling in a dependency for it.
+fn filetime_set(path: &Path, when: SystemTime) {
+    let f = std::fs::OpenOptions::new().write(true).open(path).unwrap();
+    f.set_modified(when).unwrap();
+    f.sync_all().unwrap();
+}
+
+#[test]
+#[cfg(unix)]
+fn an_unreadable_directory_does_not_delete_its_rows() {
+    // The scenario this guards: a network share or removable drive that is
+    // briefly unavailable. The walk sees nothing beneath it, which must not
+    // be read as "every file under here was deleted".
+    use std::os::unix::fs::PermissionsExt;
+
+    let root = tmp_dir("blip");
+    let db_dir = tmp_dir("blip-db");
+    let db = db_dir.join("index.sqlite");
+    let config = test_config();
+
+    touch(&root.join("visible.txt"), b"visible");
+    let vault = root.join("vault");
+    touch(&vault.join("secret.txt"), b"secret");
+    touch(&vault.join("nested/deeper.txt"), b"deeper");
+
+    index_once(&root, &db, &config);
+    assert_eq!(rows(&db).len(), 3, "all three indexed while readable");
+
+    std::fs::set_permissions(&vault, std::fs::Permissions::from_mode(0o000)).unwrap();
+    index_once(&root, &db, &config);
+    let during = rows(&db);
+    std::fs::set_permissions(&vault, std::fs::Permissions::from_mode(0o755)).unwrap();
+
+    assert_eq!(
+        during.len(),
+        3,
+        "rows under an unreadable directory must survive, not be deleted"
+    );
+
+    // And once it is readable again, everything still lines up.
+    index_once(&root, &db, &config);
+    assert_eq!(rows(&db).len(), 3);
+
+    std::fs::remove_dir_all(&root).ok();
+    std::fs::remove_dir_all(&db_dir).ok();
+}
+
+#[test]
+fn stopping_mid_run_deletes_nothing() {
+    // Pins the end-to-end property: an interrupted run must never delete the
+    // rows it did not reach.
+    //
+    // Two independent guards currently provide it — `run_indexing` skips
+    // cleanup when the walk did not complete, and `cleanup_stale_index_entries`
+    // re-checks the stop flag before its first delete. This test passes with
+    // either one alone, so it does not prove the former is present; it is here
+    // to catch the day someone removes the last of them.
+    let root = tmp_dir("stop");
+    let db_dir = tmp_dir("stop-db");
+    let db = db_dir.join("index.sqlite");
+    let config = test_config();
+
+    for i in 0..1500 {
+        touch(&root.join(format!("d{}/f{:04}.txt", i % 25, i)), b"body");
+    }
+
+    index_once(&root, &db, &config);
+    let full = rows(&db);
+    assert_eq!(full.len(), 1500);
+
+    // Start again and stop almost immediately, so the walk is cut short.
+    let service = IndexingService::new();
+    service
+        .start_indexing(
+            vec![root.to_string_lossy().into_owned()],
+            db.to_string_lossy().into_owned(),
+            config.clone(),
+        )
+        .unwrap();
+    std::thread::sleep(Duration::from_millis(15));
+    service.stop_indexing().unwrap();
+    drop(service);
+    std::thread::sleep(Duration::from_millis(250));
+
+    let after = rows(&db);
+    assert_eq!(
+        after.len(),
+        1500,
+        "an interrupted run must not delete the rows it never got to"
+    );
+
+    std::fs::remove_dir_all(&root).ok();
+    std::fs::remove_dir_all(&db_dir).ok();
+}
+
+#[test]
+fn a_wide_tree_indexes_every_file_exactly_once() {
+    // Exercises the parallel walk's chunking and termination against a real
+    // database, where a duplicate path would be a UNIQUE violation and a
+    // dropped path would be a missing row.
+    let root = tmp_dir("wide");
+    let db_dir = tmp_dir("wide-db");
+    let db = db_dir.join("index.sqlite");
+    let config = test_config();
+
+    let count = 900;
+    for i in 0..count {
+        touch(&root.join(format!("d{}/f{:04}.txt", i % 13, i)), b"body");
+    }
+
+    index_once(&root, &db, &config);
+    assert_eq!(rows(&db).len(), count, "every file indexed exactly once");
+
+    index_once(&root, &db, &config);
+    assert_eq!(rows(&db).len(), count, "and the second run is stable");
+
+    std::fs::remove_dir_all(&root).ok();
+    std::fs::remove_dir_all(&db_dir).ok();
+}
+
+/// Like `index_once`, but over several roots at once — the per-root
+/// pipeline path.
+fn index_roots_once(roots: &[&Path], db: &Path, config: &Config) {
+    if db.exists() {
+        let conn = rusqlite::Connection::open(db).unwrap();
+        conn.execute("DELETE FROM schema_info WHERE key = 'last_full_index'", [])
+            .unwrap();
+    }
+    let service = IndexingService::new();
+    service
+        .start_indexing(
+            roots.iter().map(|r| r.to_string_lossy().into_owned()).collect(),
+            db.to_string_lossy().into_owned(),
+            config.clone(),
+        )
+        .unwrap();
+    let deadline = Instant::now() + Duration::from_secs(120);
+    let mut done = false;
+    while Instant::now() < deadline {
+        if let IndexingStatus::Error(e) = service.get_status() {
+            panic!("indexing failed: {}", e);
+        }
+        if db.exists() {
+            if let Ok(conn) = rusqlite::Connection::open(db) {
+                if quicksearch_core::db::repo::get_last_full_index(&conn).is_some() {
+                    done = true;
+                    break;
+                }
+            }
+        }
+        std::thread::sleep(Duration::from_millis(10));
+    }
+    assert!(done, "indexing did not finish within the timeout");
+    service.stop_indexing().unwrap();
+}
+
+#[test]
+fn two_roots_walk_extract_and_clean_independently() {
+    let root_a = tmp_dir("multi-a");
+    let root_b = tmp_dir("multi-b");
+    let db_dir = tmp_dir("multi-db");
+    let db = db_dir.join("index.sqlite");
+    let config = test_config();
+
+    // Imbalanced roots so the round-robin writer sees a firehose and a
+    // trickle in the same run.
+    for i in 0..60 {
+        touch(&root_a.join(format!("a{:03}.txt", i)), b"alpha corpus xylophone");
+    }
+    for i in 0..5 {
+        touch(&root_b.join(format!("b{:03}.txt", i)), b"bravo corpus quagmire");
+    }
+
+    index_roots_once(&[&root_a, &root_b], &db, &config);
+
+    let conn = rusqlite::Connection::open(&db).unwrap();
+    let total: i64 = conn
+        .query_row("SELECT COUNT(*) FROM files", [], |r| r.get(0))
+        .unwrap();
+    assert_eq!(total, 65, "both roots fully walked");
+    let pending: i64 = conn
+        .query_row("SELECT COUNT(*) FROM files WHERE content_state = 0", [], |r| r.get(0))
+        .unwrap();
+    assert_eq!(pending, 0, "per-root extraction drained both roots");
+    // Content from EACH root is searchable.
+    for term in ["\"xylophone\"", "\"quagmire\""] {
+        let hits: i64 = conn
+            .query_row(
+                "SELECT COUNT(*) FROM searchabletext WHERE searchabletext MATCH ?1",
+                [term],
+                |r| r.get(0),
+            )
+            .unwrap();
+        assert!(hits > 0, "content from both roots must be indexed ({})", term);
+    }
+    drop(conn);
+
+    // Stale cleanup is global: deleting a file from the trickle root must
+    // remove exactly that row on the next multi-root run.
+    std::fs::remove_file(root_b.join("b000.txt")).unwrap();
+    index_roots_once(&[&root_a, &root_b], &db, &config);
+    let conn = rusqlite::Connection::open(&db).unwrap();
+    let total: i64 = conn
+        .query_row("SELECT COUNT(*) FROM files", [], |r| r.get(0))
+        .unwrap();
+    assert_eq!(total, 64, "stale row swept across roots");
+
+    std::fs::remove_dir_all(&root_a).ok();
+    std::fs::remove_dir_all(&root_b).ok();
+    std::fs::remove_dir_all(&db_dir).ok();
+}
+
+// ---------------------------------------------------------------------------
+// Inline extraction: the walk finishes files whose head is the whole file.
+//
+// `hash_length` is what decides how much of a file the walk reads, so setting
+// it to 0 leaves an empty head, nothing can be extracted inline, and the run
+// degrades to the pure two-pass behaviour. That makes it the control against
+// which the optimised path must produce an identical index.
+// ---------------------------------------------------------------------------
+
+/// Everything about a file's indexed content that a user can observe: its
+/// state, the stored snippet body, and its property rows.
+fn content_rows(db: &Path) -> Vec<(String, i64, Option<String>, Option<i64>, String)> {
+    let conn = rusqlite::Connection::open(db).unwrap();
+    let mut stmt = conn
+        .prepare(
+            "SELECT f.path, f.content_state, f.failure_msg, d.text_len,
+                    COALESCE(GROUP_CONCAT(p.key || '=' || p.value, ','), '')
+               FROM files f
+               LEFT JOIN documents_text d ON d.file_id = f.id
+               LEFT JOIN properties p ON p.file_id = f.id
+              GROUP BY f.id
+              ORDER BY f.path",
+        )
+        .unwrap();
+    let out = stmt
+        .query_map([], |r| {
+            Ok((r.get(0)?, r.get(1)?, r.get(2)?, r.get(3)?, r.get(4)?))
+        })
+        .unwrap()
+        .map(|r| r.unwrap())
+        .collect();
+    out
+}
+
+/// The decompressed body stored for a file, if any.
+fn stored_text(db: &Path, suffix: &str) -> Option<String> {
+    let conn = rusqlite::Connection::open(db).unwrap();
+    let blob: Option<Vec<u8>> = conn
+        .query_row(
+            "SELECT d.text_zstd FROM documents_text d
+               JOIN files f ON f.id = d.file_id
+              WHERE f.path LIKE '%' || ?1",
+            [suffix],
+            |r| r.get(0),
+        )
+        .ok();
+    blob.map(|b| String::from_utf8(zstd::decode_all(&b[..]).unwrap()).unwrap())
+}
+
+/// A tree that exercises every branch of the inline decision at once.
+fn seed_mixed_tree(root: &Path) {
+    let big = "lorem ipsum dolor sit amet ".repeat(600); // ~16 KiB, past any head
+    touch(&root.join("small.txt"), b"a small plaintext body with xylophone in it");
+    touch(&root.join("large.txt"), big.as_bytes());
+    touch(&root.join("empty.txt"), b"");
+    // Invalid UTF-8 with a .txt extension: claimed by the plaintext extractor,
+    // but not decodable, so it must be reported as a failure either way.
+    touch(&root.join("bad.txt"), &[0x68, 0x69, 0xff, 0xfe, 0x00, 0x41]);
+    // No extension `infer` or `mime_guess` recognises: no extractor claims it.
+    touch(&root.join("blob.bin"), &[0x00, 0x01, 0x02, 0xfd, 0xfe, 0xff]);
+    touch(&root.join("nested/deep/note.md"), b"# heading\n\nquagmire body text\n");
+}
+
+#[test]
+fn inline_extraction_produces_an_identical_index_to_the_two_pass_path() {
+    let root = tmp_dir("inline-equiv");
+    let db_dir = tmp_dir("inline-equiv-db");
+    seed_mixed_tree(&root);
+
+    // Control: hash_length 0 => empty head => nothing can be inlined.
+    let mut control = Config::default();
+    control.processing.hash_length = 0;
+    let db_control = db_dir.join("control.sqlite");
+    index_once(&root, &db_control, &control);
+
+    // Optimised: the default head covers every small file in the tree.
+    let optimised = Config::default();
+    let db_opt = db_dir.join("optimised.sqlite");
+    index_once(&root, &db_opt, &optimised);
+
+    assert_eq!(
+        content_rows(&db_control),
+        content_rows(&db_opt),
+        "inlining during the walk must not change a single indexed byte"
+    );
+
+    // And the bodies themselves round-trip identically, not just their lengths.
+    for f in ["small.txt", "large.txt", "note.md"] {
+        assert_eq!(
+            stored_text(&db_control, f),
+            stored_text(&db_opt, f),
+            "stored body differs for {}",
+            f
+        );
+    }
+
+    std::fs::remove_dir_all(&root).ok();
+    std::fs::remove_dir_all(&db_dir).ok();
+}
+
+#[test]
+fn the_head_boundary_decides_inlining_without_changing_the_result() {
+    let root = tmp_dir("inline-boundary");
+    let db_dir = tmp_dir("inline-boundary-db");
+
+    // Exactly at the limit, and one byte past it.
+    let mut config = Config::default();
+    config.processing.hash_length = 64;
+    let at = "x".repeat(64);
+    let past = "y".repeat(65);
+    touch(&root.join("at.txt"), at.as_bytes());
+    touch(&root.join("past.txt"), past.as_bytes());
+
+    let db = db_dir.join("index.sqlite");
+    index_once(&root, &db, &config);
+
+    // Both are fully extracted; the boundary only decides *which pass* did it.
+    let conn = rusqlite::Connection::open(&db).unwrap();
+    let pending: i64 = conn
+        .query_row("SELECT COUNT(*) FROM files WHERE content_state != 1", [], |r| r.get(0))
+        .unwrap();
+    assert_eq!(pending, 0, "both sides of the boundary end up extracted");
+    drop(conn);
+
+    assert_eq!(stored_text(&db, "at.txt").as_deref(), Some(at.as_str()));
+    assert_eq!(stored_text(&db, "past.txt").as_deref(), Some(past.as_str()));
+
+    std::fs::remove_dir_all(&root).ok();
+    std::fs::remove_dir_all(&db_dir).ok();
+}
+
+#[test]
+fn undecodable_small_files_are_reported_as_failures_not_silently_skipped() {
+    let root = tmp_dir("inline-badutf8");
+    let db_dir = tmp_dir("inline-badutf8-db");
+    let db = db_dir.join("index.sqlite");
+
+    touch(&root.join("bad.txt"), &[0x68, 0x69, 0xff, 0xfe]);
+    index_once(&root, &db, &Config::default());
+
+    let conn = rusqlite::Connection::open(&db).unwrap();
+    let (state, msg): (i64, Option<String>) = conn
+        .query_row(
+            "SELECT content_state, failure_msg FROM files WHERE path LIKE '%bad.txt'",
+            [],
+            |r| Ok((r.get(0)?, r.get(1)?)),
+        )
+        .unwrap();
+    // Inlining must not swallow the error: the walk declines to record it, so
+    // the content pass still opens the file and stores a reason.
+    assert_eq!(state, 2, "undecodable content is FAILED, not DONE or NA");
+    assert!(
+        msg.unwrap_or_default().contains("bad.txt"),
+        "the failure names the file"
+    );
+
+    std::fs::remove_dir_all(&root).ok();
+    std::fs::remove_dir_all(&db_dir).ok();
+}
+
+#[test]
+fn an_empty_file_is_done_with_no_snippet_sidecar() {
+    let root = tmp_dir("inline-empty");
+    let db_dir = tmp_dir("inline-empty-db");
+    let db = db_dir.join("index.sqlite");
+
+    touch(&root.join("empty.txt"), b"");
+    index_once(&root, &db, &Config::default());
+
+    let conn = rusqlite::Connection::open(&db).unwrap();
+    let (state, sidecars): (i64, i64) = conn
+        .query_row(
+            "SELECT f.content_state, (SELECT COUNT(*) FROM documents_text d WHERE d.file_id = f.id)
+               FROM files f WHERE f.path LIKE '%empty.txt'",
+            [],
+            |r| Ok((r.get(0)?, r.get(1)?)),
+        )
+        .unwrap();
+    assert_eq!(state, 1, "an empty file is extracted, not failed");
+    assert_eq!(sidecars, 0, "no zstd frame for an empty body");
+
+    std::fs::remove_dir_all(&root).ok();
+    std::fs::remove_dir_all(&db_dir).ok();
+}
+
+#[test]
+fn the_content_extension_filter_still_excludes_small_text_files() {
+    let root = tmp_dir("inline-filter");
+    let db_dir = tmp_dir("inline-filter-db");
+    let db = db_dir.join("index.sqlite");
+
+    let mut config = Config::default();
+    config.indexing.content_extensions = vec!["md".into()];
+    touch(&root.join("kept.md"), b"kept quagmire body");
+    touch(&root.join("skipped.txt"), b"skipped xylophone body");
+    index_once(&root, &db, &config);
+
+    let conn = rusqlite::Connection::open(&db).unwrap();
+    let states: Vec<(String, i64)> = conn
+        .prepare("SELECT path, content_state FROM files ORDER BY path")
+        .unwrap()
+        .query_map([], |r| Ok((r.get(0)?, r.get(1)?)))
+        .unwrap()
+        .map(|r| r.unwrap())
+        .collect();
+    for (path, state) in &states {
+        if path.ends_with("kept.md") {
+            assert_eq!(*state, 1, "an allowed extension is extracted");
+        } else {
+            assert_eq!(*state, 3, "a filtered extension is NA, never inlined");
+        }
+    }
+    drop(conn);
+    assert_eq!(stored_text(&db, "skipped.txt"), None, "no body stored for a filtered file");
+
+    std::fs::remove_dir_all(&root).ok();
+    std::fs::remove_dir_all(&db_dir).ok();
+}
+
+#[test]
+fn contentless_mode_still_indexes_inlined_files_without_storing_bodies() {
+    let root = tmp_dir("inline-contentless");
+    let db_dir = tmp_dir("inline-contentless-db");
+    let db = db_dir.join("index.sqlite");
+
+    let mut config = Config::default();
+    config.processing.store_text_for_snippets = false;
+    touch(&root.join("small.txt"), b"searchable xylophone body");
+    index_once(&root, &db, &config);
+
+    let conn = rusqlite::Connection::open(&db).unwrap();
+    let sidecars: i64 = conn
+        .query_row("SELECT COUNT(*) FROM documents_text", [], |r| r.get(0))
+        .unwrap();
+    assert_eq!(sidecars, 0, "contentless mode stores no bodies");
+    let hits: i64 = conn
+        .query_row(
+            "SELECT COUNT(*) FROM searchabletext WHERE searchabletext MATCH '\"xylophone\"'",
+            [],
+            |r| r.get(0),
+        )
+        .unwrap();
+    assert_eq!(hits, 1, "an inlined file is still searchable in contentless mode");
+
+    std::fs::remove_dir_all(&root).ok();
+    std::fs::remove_dir_all(&db_dir).ok();
+}

crates/quicksearch-core/tests/snippet_perf.rs

@@ -207,12 +207,7 @@ fn snippet_paths_perf_comparison() {
     let b_reps = 10;
     let start_b = Instant::now();
     let mut rows_b_total = 0usize;
-    let opts = snippet::Options {
+    let opts = snippet::Options { approx_chars: 64 };
-        pre: "<b>",
-        post: "</b>",
-        ellipsis: "...",
-        approx_chars: 64,
-    };
     for _ in 0..b_reps {
         for q in QUERIES {
             // Contentless FTS5 returns NULL for stored columns (that's the
@@ -245,7 +240,7 @@ fn snippet_paths_perf_comparison() {
                     }
                     None => String::new(),
                 };
-                let _snip = snippet::render(&text, &[q], &opts);
+                let _snip = snippet::extract(&text, &[q], &opts);
                 rows_b_total += 1;
             }
         }

crates/quicksearch-gui/Cargo.toml

@@ -2,14 +2,52 @@
 name = "quicksearch-gui"
 version.workspace = true
 edition.workspace = true
+license.workspace = true
+authors.workspace = true
+repository.workspace = true
+description = "Fast full-text search across your files: desktop app and terminal search tool."
 
 [[bin]]
 name = "quicksearch"
 path = "src/main.rs"
 
+# Terminal search lives in its own binary because the two want opposite
+# Windows subsystems: a GUI built as a console app flashes a console window on
+# every launch, and a console tool built as a GUI app cannot write to the shell
+# that invoked it (cmd and PowerShell do not even wait for it). Splitting is
+# the only arrangement that is correct in both cases. On Unix the distinction
+# does not exist and `quicksearch` still does both.
+[[bin]]
+name = "quicksearch-cli"
+path = "src/cli_main.rs"
+
 [dependencies]
 quicksearch-core = { path = "../quicksearch-core" }
 
-dioxus = { version = "0.5.1", features = ["desktop"] }
+eframe = { version = "0.32", default-features = false, features = [
-dioxus-desktop = "0.5.1"
+    "glow",
-tokio = { version = "1.0", features = ["time", "signal"] }
+    "persistence",
+] }
+egui = "0.32"
+egui_extras = "0.32"
+rfd = "0.15"
+open = "5"
+chrono = { version = "0.4", default-features = false, features = ["clock"] }
+
+# Display backends, which only exist on Linux/BSD. `default-features = false`
+# has to be repeated: feature resolution unions the two stanzas, so a single
+# permissive one would switch defaults back on for every target.
+[target.'cfg(all(unix, not(target_os = "macos")))'.dependencies]
+eframe = { version = "0.32", default-features = false, features = [
+    "wayland",
+    "x11",
+] }
+
+# Console attachment for the GUI binary (which has no stdio when launched from
+# Explorer) and VT-mode enabling for the CLI binary. 0.59 matches what eframe
+# and rfd already resolve, so no extra crate is compiled.
+[target.'cfg(windows)'.dependencies]
+windows-sys = { version = "0.59", features = [
+    "Win32_Foundation",
+    "Win32_System_Console",
+] }

crates/quicksearch-gui/assets/app.js

@@ -1,232 +0,0 @@
-// QuickSearch Application JavaScript
-
-// Enhanced UI interactions
-document.addEventListener('DOMContentLoaded', function() {
-    console.log('QuickSearch UI loaded');
-    
-    // Add loading states to buttons
-    function addLoadingState(button, originalText) {
-        button.disabled = true;
-        button.innerHTML = '<span class="loading"></span>' + originalText;
-    }
-    
-    function removeLoadingState(button, originalText) {
-        button.disabled = false;
-        button.innerHTML = originalText;
-    }
-    
-    // Enhanced form interactions
-    const forms = document.querySelectorAll('form');
-    forms.forEach(form => {
-        form.addEventListener('submit', function(e) {
-            const submitButton = form.querySelector('button[type="submit"]');
-            if (submitButton) {
-                addLoadingState(submitButton, submitButton.textContent);
-            }
-        });
-    });
-    
-    // Keyboard shortcuts
-    document.addEventListener('keydown', function(e) {
-        // Ctrl+F to focus search
-        if (e.ctrlKey && e.key === 'f') {
-            e.preventDefault();
-            const searchInput = document.querySelector('input[type="text"]');
-            if (searchInput) {
-                searchInput.focus();
-                searchInput.select();
-            }
-        }
-        
-        // Escape to clear search
-        if (e.key === 'Escape') {
-            const searchInput = document.querySelector('input[type="text"]');
-            if (searchInput && searchInput === document.activeElement) {
-                searchInput.value = '';
-                searchInput.blur();
-            }
-        }
-    });
-    
-    // Enhanced table interactions
-    function enhanceTable(table) {
-        // Add click-to-copy functionality for table cells
-        const cells = table.querySelectorAll('td');
-        cells.forEach(cell => {
-            cell.addEventListener('click', function() {
-                const text = cell.textContent.trim();
-                if (text && navigator.clipboard) {
-                    navigator.clipboard.writeText(text).then(() => {
-                        // Visual feedback
-                        cell.style.backgroundColor = '#4CAF50';
-                        cell.style.color = 'white';
-                        setTimeout(() => {
-                            cell.style.backgroundColor = '';
-                            cell.style.color = '';
-                        }, 200);
-                    });
-                }
-            });
-        });
-        
-        // Add sortable columns (basic implementation)
-        const headers = table.querySelectorAll('th');
-        headers.forEach((header, index) => {
-            header.style.cursor = 'pointer';
-            header.addEventListener('click', () => sortTable(table, index));
-        });
-    }
-    
-    // Simple table sorting
-    function sortTable(table, columnIndex) {
-        const tbody = table.querySelector('tbody');
-        const rows = Array.from(tbody.querySelectorAll('tr'));
-        
-        rows.sort((a, b) => {
-            const aVal = a.cells[columnIndex]?.textContent.trim() || '';
-            const bVal = b.cells[columnIndex]?.textContent.trim() || '';
-            
-            // Try numeric sort first
-            const aNum = parseFloat(aVal);
-            const bNum = parseFloat(bVal);
-            
-            if (!isNaN(aNum) && !isNaN(bNum)) {
-                return aNum - bNum;
-            }
-            
-            // Fall back to string sort
-            return aVal.localeCompare(bVal);
-        });
-        
-        // Clear tbody and re-append sorted rows
-        tbody.innerHTML = '';
-        rows.forEach(row => tbody.appendChild(row));
-    }
-    
-    // Auto-enhance any tables that appear
-    const observer = new MutationObserver(function(mutations) {
-        mutations.forEach(function(mutation) {
-            mutation.addedNodes.forEach(function(node) {
-                if (node.nodeType === 1) { // Element node
-                    const tables = node.querySelectorAll ? node.querySelectorAll('table') : [];
-                    tables.forEach(enhanceTable);
-                    
-                    if (node.tagName === 'TABLE') {
-                        enhanceTable(node);
-                    }
-                }
-            });
-        });
-    });
-    
-    observer.observe(document.body, { childList: true, subtree: true });
-    
-    // Enhance existing tables
-    document.querySelectorAll('table').forEach(enhanceTable);
-});
-
-// Utility functions for Rust integration
-window.QuickSearch = {
-    // Function to show toast notifications
-    showToast: function(message, type = 'info') {
-        const toast = document.createElement('div');
-        toast.className = `toast toast-${type}`;
-        toast.textContent = message;
-        toast.style.cssText = `
-            position: fixed;
-            top: 20px;
-            right: 20px;
-            padding: 12px 20px;
-            border-radius: 6px;
-            color: white;
-            font-weight: 600;
-            z-index: 2000;
-            animation: slideIn 0.3s ease;
-        `;
-        
-        // Set background based on type
-        const colors = {
-            info: '#2196F3',
-            success: '#4CAF50',
-            warning: '#FF9800',
-            error: '#f44336'
-        };
-        toast.style.backgroundColor = colors[type] || colors.info;
-        
-        document.body.appendChild(toast);
-        
-        setTimeout(() => {
-            toast.style.animation = 'slideOut 0.3s ease';
-            setTimeout(() => {
-                document.body.removeChild(toast);
-            }, 300);
-        }, 3000);
-    },
-    
-    // Function to update status display
-    updateStatus: function(status) {
-        const statusDisplay = document.querySelector('.status-display');
-        if (statusDisplay) {
-            statusDisplay.textContent = status;
-        }
-    },
-    
-    // Function to highlight search terms in results
-    highlightSearchTerms: function(searchTerm, container) {
-        if (!searchTerm || !container) return;
-        
-        const walker = document.createTreeWalker(
-            container,
-            NodeFilter.SHOW_TEXT,
-            null,
-            false
-        );
-        
-        const textNodes = [];
-        let node;
-        while (node = walker.nextNode()) {
-            textNodes.push(node);
-        }
-        
-        textNodes.forEach(textNode => {
-            const parent = textNode.parentNode;
-            if (parent.tagName === 'B') return; // Skip already highlighted
-            
-            const text = textNode.textContent;
-            const regex = new RegExp(`(${searchTerm})`, 'gi');
-            
-            if (regex.test(text)) {
-                const highlightedHTML = text.replace(regex, '<mark>$1</mark>');
-                const wrapper = document.createElement('span');
-                wrapper.innerHTML = highlightedHTML;
-                parent.replaceChild(wrapper, textNode);
-            }
-        });
-    }
-};
-
-// Add custom CSS for toasts and animations
-const style = document.createElement('style');
-style.textContent = `
-    @keyframes slideIn {
-        from { transform: translateX(100%); opacity: 0; }
-        to { transform: translateX(0); opacity: 1; }
-    }
-    
-    @keyframes slideOut {
-        from { transform: translateX(0); opacity: 1; }
-        to { transform: translateX(100%); opacity: 0; }
-    }
-    
-    mark {
-        background: #ffeb3b;
-        padding: 2px 4px;
-        border-radius: 3px;
-        font-weight: bold;
-    }
-    
-    .toast {
-        box-shadow: 0 4px 12px rgba(0,0,0,0.2);
-    }
-`;
-document.head.appendChild(style);

crates/quicksearch-gui/assets/icons/quicksearch_icon.svg

@@ -0,0 +1,14 @@
+<svg width="240" height="240" viewBox="0 0 240 240" role="img" xmlns="http://www.w3.org/2000/svg"><title style="fill:rgb(0, 0, 0);stroke:none;color:rgb(11, 11, 11);stroke-width:1px;stroke-linecap:butt;stroke-linejoin:miter;opacity:1;font-family:&quot;Anthropic Sans&quot;, -apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, sans-serif;font-size:16px;font-weight:400;text-anchor:start;dominant-baseline:auto">Fast full-text search icon, square crop</title><desc style="fill:rgb(0, 0, 0);stroke:none;color:rgb(11, 11, 11);stroke-width:1px;stroke-linecap:butt;stroke-linejoin:miter;opacity:1;font-family:&quot;Anthropic Sans&quot;, -apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, sans-serif;font-size:16px;font-weight:400;text-anchor:start;dominant-baseline:auto">A square app icon: centered text lines indenting around a magnifying glass with a lightning bolt inside.</desc>
+<rect x="0" y="0" width="240" height="240" rx="52" fill="#1e293b" style="fill:rgb(30, 41, 59);stroke:none;color:rgb(11, 11, 11);stroke-width:1px;stroke-linecap:butt;stroke-linejoin:miter;opacity:1;font-family:&quot;Anthropic Sans&quot;, -apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, sans-serif;font-size:16px;font-weight:400;text-anchor:start;dominant-baseline:auto"/>
+<g stroke="#475569" stroke-width="10" stroke-linecap="round" style="fill:rgb(0, 0, 0);stroke:rgb(71, 85, 105);color:rgb(11, 11, 11);stroke-width:10px;stroke-linecap:round;stroke-linejoin:miter;opacity:1;font-family:&quot;Anthropic Sans&quot;, -apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, sans-serif;font-size:16px;font-weight:400;text-anchor:start;dominant-baseline:auto">
+<line x1="42" y1="58" x2="198" y2="58" style="fill:rgb(0, 0, 0);stroke:rgb(71, 85, 105);color:rgb(11, 11, 11);stroke-width:10px;stroke-linecap:round;stroke-linejoin:miter;opacity:1;font-family:&quot;Anthropic Sans&quot;, -apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, sans-serif;font-size:16px;font-weight:400;text-anchor:start;dominant-baseline:auto"/>
+<line x1="42" y1="92" x2="62" y2="92" style="fill:rgb(0, 0, 0);stroke:rgb(71, 85, 105);color:rgb(11, 11, 11);stroke-width:10px;stroke-linecap:round;stroke-linejoin:miter;opacity:1;font-family:&quot;Anthropic Sans&quot;, -apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, sans-serif;font-size:16px;font-weight:400;text-anchor:start;dominant-baseline:auto"/>
+<line x1="42" y1="126" x2="56" y2="126" style="fill:rgb(0, 0, 0);stroke:rgb(71, 85, 105);color:rgb(11, 11, 11);stroke-width:10px;stroke-linecap:round;stroke-linejoin:miter;opacity:1;font-family:&quot;Anthropic Sans&quot;, -apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, sans-serif;font-size:16px;font-weight:400;text-anchor:start;dominant-baseline:auto"/>
+<line x1="42" y1="160" x2="68" y2="160" style="fill:rgb(0, 0, 0);stroke:rgb(71, 85, 105);color:rgb(11, 11, 11);stroke-width:10px;stroke-linecap:round;stroke-linejoin:miter;opacity:1;font-family:&quot;Anthropic Sans&quot;, -apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, sans-serif;font-size:16px;font-weight:400;text-anchor:start;dominant-baseline:auto"/>
+<line x1="42" y1="194" x2="168" y2="194" style="fill:rgb(0, 0, 0);stroke:rgb(71, 85, 105);color:rgb(11, 11, 11);stroke-width:10px;stroke-linecap:round;stroke-linejoin:miter;opacity:1;font-family:&quot;Anthropic Sans&quot;, -apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, sans-serif;font-size:16px;font-weight:400;text-anchor:start;dominant-baseline:auto"/>
+</g>
+<circle cx="130" cy="118" r="46" fill="#0ea5e9" fill-opacity="0.18" style="fill:rgb(14, 165, 233);stroke:none;color:rgb(11, 11, 11);stroke-width:1px;stroke-linecap:butt;stroke-linejoin:miter;opacity:1;font-family:&quot;Anthropic Sans&quot;, -apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, sans-serif;font-size:16px;font-weight:400;text-anchor:start;dominant-baseline:auto"/>
+<circle cx="130" cy="118" r="46" fill="none" stroke="#38bdf8" stroke-width="14" style="fill:none;stroke:rgb(56, 189, 248);color:rgb(11, 11, 11);stroke-width:14px;stroke-linecap:butt;stroke-linejoin:miter;opacity:1;font-family:&quot;Anthropic Sans&quot;, -apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, sans-serif;font-size:16px;font-weight:400;text-anchor:start;dominant-baseline:auto"/>
+<line x1="163" y1="151" x2="195" y2="183" stroke="#38bdf8" stroke-width="16" stroke-linecap="round" style="fill:rgb(0, 0, 0);stroke:rgb(56, 189, 248);color:rgb(11, 11, 11);stroke-width:16px;stroke-linecap:round;stroke-linejoin:miter;opacity:1;font-family:&quot;Anthropic Sans&quot;, -apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, sans-serif;font-size:16px;font-weight:400;text-anchor:start;dominant-baseline:auto"/>
+<path d="M137 84 L108 125 L128 125 L121 154 L152 111 L132 111 Z" fill="#facc15" style="fill:rgb(250, 204, 21);stroke:none;color:rgb(11, 11, 11);stroke-width:1px;stroke-linecap:butt;stroke-linejoin:miter;opacity:1;font-family:&quot;Anthropic Sans&quot;, -apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, sans-serif;font-size:16px;font-weight:400;text-anchor:start;dominant-baseline:auto"/>
+</svg>

crates/quicksearch-gui/assets/index.html

@@ -1,21 +0,0 @@
-<!DOCTYPE html>
-<html lang="en">
-<head>
-    <meta charset="utf-8">
-    <meta name="viewport" content="width=device-width, initial-scale=1">
-    <title>QuickSearch - File Indexer & Search</title>
-    
-    <!-- External CSS -->
-    <link rel="stylesheet" href="assets://styles.css">
-    
-    <!-- Favicon (optional) -->
-    <link rel="icon" type="image/x-icon" href="data:image/svg+xml,<svg xmlns='http://www.w3.org/2000/svg' viewBox='0 0 24 24'><path fill='%234CAF50' d='M15.5 14h-.79l-.28-.27C15.41 12.59 16 11.11 16 9.5 16 5.91 13.09 3 9.5 3S3 5.91 3 9.5 5.91 16 9.5 16c1.61 0 3.09-.59 4.23-1.57l.27.28v.79l5 4.99L20.49 19l-4.99-5zm-6 0C7.01 14 5 11.99 5 9.5S7.01 5 9.5 5 14 7.01 14 9.5 11.99 14 9.5 14z'/></svg>">
-</head>
-<body>
-    <!-- Main app container - Dioxus will render into this -->
-    <div id="main"></div>
-    
-    <!-- External JavaScript -->
-    <script src="assets://app.js"></script>
-</body>
-</html>

crates/quicksearch-gui/assets/styles.css

@@ -1,299 +0,0 @@
-/* QuickSearch Application Styles */
-body {
-    font-family: 'Segoe UI', Tahoma, Geneva, Verdana, sans-serif;
-    margin: 0;
-    padding: 20px;
-    background: linear-gradient(135deg, #667eea 0%, #764ba2 100%);
-    min-height: 100vh;
-}
-
-.app-container {
-    max-width: 1200px;
-    margin: 0 auto;
-    background: white;
-    border-radius: 12px;
-    box-shadow: 0 8px 32px rgba(0,0,0,0.1);
-    overflow: hidden;
-}
-
-.app-header {
-    background: linear-gradient(90deg, #4CAF50 0%, #45a049 100%);
-    color: white;
-    padding: 20px;
-    text-align: center;
-}
-
-.app-content {
-    padding: 20px;
-}
-
-.section {
-    background: #f8f9fa;
-    border-radius: 8px;
-    padding: 20px;
-    margin-bottom: 20px;
-    border-left: 4px solid #4CAF50;
-}
-
-.form-control {
-    width: 100%;
-    padding: 12px;
-    border: 2px solid #e0e0e0;
-    border-radius: 6px;
-    font-size: 14px;
-    transition: border-color 0.3s ease;
-    box-sizing: border-box;
-}
-
-.form-control:focus {
-    outline: none;
-    border-color: #4CAF50;
-    box-shadow: 0 0 0 3px rgba(76, 175, 80, 0.1);
-}
-
-.btn {
-    padding: 12px 24px;
-    border: none;
-    border-radius: 6px;
-    cursor: pointer;
-    font-size: 14px;
-    font-weight: 600;
-    text-transform: uppercase;
-    letter-spacing: 0.5px;
-    transition: all 0.3s ease;
-    margin-right: 10px;
-    display: inline-block;
-}
-
-.btn-primary {
-    background: linear-gradient(90deg, #4CAF50 0%, #45a049 100%);
-    color: white;
-}
-
-.btn-primary:hover:not(:disabled) {
-    transform: translateY(-2px);
-    box-shadow: 0 4px 12px rgba(76, 175, 80, 0.3);
-}
-
-.btn-danger {
-    background: linear-gradient(90deg, #f44336 0%, #d32f2f 100%);
-    color: white;
-}
-
-.btn-danger:hover:not(:disabled) {
-    transform: translateY(-2px);
-    box-shadow: 0 4px 12px rgba(244, 67, 54, 0.3);
-}
-
-.btn-info {
-    background: linear-gradient(90deg, #2196F3 0%, #1976D2 100%);
-    color: white;
-}
-
-.btn-info:hover:not(:disabled) {
-    transform: translateY(-2px);
-    box-shadow: 0 4px 12px rgba(33, 150, 243, 0.3);
-}
-
-.btn:disabled {
-    opacity: 0.6;
-    cursor: not-allowed;
-    transform: none !important;
-    box-shadow: none !important;
-}
-
-.status-display {
-    background: #1a1a1a;
-    color: #00ff00;
-    padding: 15px;
-    border-radius: 6px;
-    font-family: 'Courier New', monospace;
-    white-space: pre-wrap;
-    font-size: 13px;
-    border: 1px solid #333;
-    overflow-x: auto;
-}
-
-.search-results {
-    background: white;
-    border-radius: 8px;
-    overflow: hidden;
-    box-shadow: 0 2px 8px rgba(0,0,0,0.1);
-    margin-top: 20px;
-}
-
-.results-table {
-    max-height: 400px;
-    overflow: auto;
-    border: 1px solid #ddd;
-    border-radius: 6px;
-}
-
-.results-table table {
-    width: 100%;
-    border-collapse: collapse;
-}
-
-.results-table th {
-    background: #4CAF50;
-    color: white;
-    padding: 12px;
-    text-align: left;
-    font-weight: 600;
-    position: sticky;
-    top: 0;
-    border-bottom: 2px solid #45a049;
-}
-
-.results-table td {
-    padding: 10px 12px;
-    border-bottom: 1px solid #e0e0e0;
-    word-break: break-all;
-}
-
-.results-table tbody tr:hover {
-    background: #f5f5f5;
-}
-
-.results-table tbody tr:nth-child(even) {
-    background: #fafafa;
-}
-
-.results-table tbody tr:nth-child(even):hover {
-    background: #f0f0f0;
-}
-
-/* Clickable path cells */
-.path-cell.clickable {
-    cursor: pointer;
-    color: #1976d2;
-    font-weight: 500;
-    transition: all 0.2s ease;
-    position: relative;
-}
-
-.path-cell.clickable:hover {
-    background: #e3f2fd !important;
-    color: #0d47a1;
-    text-decoration: underline;
-}
-
-.path-cell.clickable:active {
-    background: #bbdefb !important;
-    transform: scale(0.98);
-}
-
-
-.error-message {
-    background: #ffebee;
-    color: #c62828;
-    padding: 12px;
-    border-radius: 6px;
-    border-left: 4px solid #f44336;
-    margin: 10px 0;
-}
-
-.modal-backdrop {
-    position: fixed;
-    top: 0;
-    left: 0;
-    width: 100%;
-    height: 100%;
-    background: rgba(0,0,0,0.5);
-    backdrop-filter: blur(4px);
-    display: flex;
-    align-items: center;
-    justify-content: center;
-    z-index: 1000;
-}
-
-.modal-dialog {
-    background: white;
-    padding: 30px;
-    border-radius: 12px;
-    max-width: 600px;
-    box-shadow: 0 20px 60px rgba(0,0,0,0.3);
-    animation: modalSlideIn 0.3s ease;
-}
-
-@keyframes modalSlideIn {
-    from {
-        opacity: 0;
-        transform: translateY(-20px);
-    }
-    to {
-        opacity: 1;
-        transform: translateY(0);
-    }
-}
-
-h1, h2, h3 {
-    margin-top: 0;
-    color: #333;
-}
-
-label {
-    display: block;
-    margin-bottom: 8px;
-    font-weight: 600;
-    color: #555;
-}
-
-.form-group {
-    margin-bottom: 20px;
-}
-
-/* Highlight matched text in search results */
-.results-table b {
-    background: #ffeb3b;
-    padding: 2px 4px;
-    border-radius: 3px;
-    font-weight: bold;
-    color: #333;
-}
-
-/* Loading spinner */
-.loading {
-    display: inline-block;
-    width: 20px;
-    height: 20px;
-    border: 3px solid #f3f3f3;
-    border-top: 3px solid #4CAF50;
-    border-radius: 50%;
-    animation: spin 1s linear infinite;
-    margin-right: 10px;
-}
-
-@keyframes spin {
-    0% { transform: rotate(0deg); }
-    100% { transform: rotate(360deg); }
-}
-
-/* Responsive design */
-@media (max-width: 768px) {
-    body {
-        padding: 10px;
-    }
-    
-    .app-container {
-        border-radius: 8px;
-    }
-    
-    .app-content {
-        padding: 15px;
-    }
-    
-    .form-control {
-        font-size: 16px; /* Prevents zoom on iOS */
-    }
-    
-    .results-table {
-        font-size: 12px;
-    }
-    
-    .btn {
-        width: 100%;
-        margin-bottom: 10px;
-        margin-right: 0;
-    }
-}

crates/quicksearch-gui/src/app.rs

@@ -0,0 +1,721 @@
+//! Application shell: tab strip, per-frame event drains, debounce,
+//! status bar, and config-change routing.
+
+use std::time::{Duration, Instant};
+
+use quicksearch_core::cli::{index_counts, IndexCounts};
+use quicksearch_core::config::{diff_actions, nested_roots, Config};
+use quicksearch_core::coordinator::{IndexMode, IndexerState, WatcherStatus};
+use quicksearch_core::indexing::{ConfigChange, IndexingStatus, RootPhase};
+use quicksearch_core::search::SearchOptions;
+use quicksearch_core::watcher::WatchError;
+
+use crate::backend::Backend;
+use crate::duplicates_tab::{DupState, DuplicatesTab};
+use crate::format::{fmt_interval, group_thousands};
+use crate::logs_tab::LogsTab;
+use crate::manage_tab::ManageTab;
+use crate::options::OptionsWindow;
+use crate::search_tab::SearchTab;
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+enum Tab {
+    Search,
+    Manage,
+    Duplicates,
+    Logs,
+}
+
+pub struct QuickSearchApp {
+    cfg: Config,
+    backend: Backend,
+    tab: Tab,
+    search: SearchTab,
+    manage: ManageTab,
+    dups: DuplicatesTab,
+    logs: LogsTab,
+    options: OptionsWindow,
+    /// Cached idle counts for the status bar, refreshed at most every 5 s.
+    counts: Option<(Instant, IndexCounts)>,
+    /// Set when applying a config that invalidates the stored index.
+    rebuild_prompt: Option<Vec<ConfigChange>>,
+    /// Set while the "delete the index?" confirmation is open.
+    clear_prompt: bool,
+    /// Nested roots found in the loaded config (startup validation); shown
+    /// as a modal over the Manage tab until dismissed.
+    nested_prompt: Option<Vec<(String, String)>>,
+    /// Set when the watcher gave up on the directory budget and live
+    /// updates are off; see [`QuickSearchApp::check_watch_cap_warning`].
+    watch_cap_prompt: Option<WatchError>,
+    config_error: Option<String>,
+}
+
+impl QuickSearchApp {
+    /// `initial_query` pre-fills the search box and fires a search on the
+    /// first frame. It carries the positional arguments the binary was given,
+    /// which on Windows is the only thing the GUI can do with them — terminal
+    /// output belongs to `quicksearch-cli` there.
+    pub fn new(
+        cc: &eframe::CreationContext<'_>,
+        cfg: Config,
+        config_error: Option<String>,
+        initial_query: Option<String>,
+    ) -> Result<QuickSearchApp, String> {
+        // Compact styling: results density is the whole point.
+        cc.egui_ctx.style_mut(|style| {
+            style.spacing.item_spacing = egui::vec2(6.0, 3.0);
+            style.spacing.button_padding = egui::vec2(6.0, 2.0);
+        });
+        cc.egui_ctx.set_zoom_factor(clamp_scale(cfg.ui.scale));
+
+        let backend = Backend::start(&cfg, cc.egui_ctx.clone())?;
+        let fuzzy = cfg.search.fuzzy_default;
+        // Startup validation: a hand-edited config can nest roots, which
+        // per-root pipelines can't accept. Redirect straight to the folder
+        // list with an explanatory modal; the coordinator refuses runs
+        // until it's fixed.
+        let nested = nested_roots(&cfg.paths.indexing_paths);
+        let (tab, nested_prompt) = if nested.is_empty() {
+            (Tab::Search, None)
+        } else {
+            (Tab::Manage, Some(nested))
+        };
+        let mut search = SearchTab::new(fuzzy);
+        if let Some(query) = initial_query {
+            search.seed(query);
+        }
+        Ok(QuickSearchApp {
+            cfg,
+            backend,
+            tab,
+            search,
+            manage: ManageTab::new(),
+            dups: DuplicatesTab::new(),
+            logs: LogsTab::new(),
+            options: OptionsWindow::new(),
+            counts: None,
+            rebuild_prompt: None,
+            clear_prompt: false,
+            nested_prompt,
+            watch_cap_prompt: None,
+            config_error,
+        })
+    }
+
+    fn search_options(&self) -> SearchOptions {
+        SearchOptions {
+            fuzzy: self.search.fuzzy,
+            fuzzy_max_edits: self.cfg.search.fuzzy_max_edits,
+            limit: self.cfg.search.display_limit,
+            batch: self.cfg.search.results_per_page.max(1),
+            session_ignores: self.search.session_ignores.clone(),
+        }
+    }
+
+    fn start_search(&mut self) {
+        let generation = self
+            .backend
+            .search()
+            .search(&self.search.query, self.search_options());
+        self.search.on_search_started(generation);
+    }
+
+    fn start_duplicates_scan(&mut self, ctx: &egui::Context) {
+        self.dups.state = DupState::Loading;
+        let cfg = self.cfg.clone();
+        self.backend.start_duplicates(&cfg, ctx.clone());
+    }
+
+    /// Save + route an edited config to the running services.
+    fn apply_new_config(&mut self, ctx: &egui::Context, mut new: Config) {
+        if let Some((child, parent)) = nested_roots(&new.paths.indexing_paths).first() {
+            self.config_error = Some(format!(
+                "Not applied: indexed folder {} is nested under {}",
+                child, parent
+            ));
+            return;
+        }
+        // Warned-root memory only means anything for folders still indexed.
+        // Pruning here is what makes removing and re-adding a folder warn
+        // again rather than staying silently suppressed forever.
+        new.ui
+            .watch_cap_warned_roots
+            .retain(|root| new.paths.indexing_paths.contains(root));
+        let actions = diff_actions(&self.cfg, &new);
+        if let Err(e) = new.save() {
+            self.config_error = Some(e);
+        }
+        if (new.ui.scale - self.cfg.ui.scale).abs() > f32::EPSILON {
+            ctx.set_zoom_factor(clamp_scale(new.ui.scale));
+        }
+        if actions.search_db_changed {
+            self.backend
+                .search()
+                .set_db_path(new.resolved_database_path());
+            self.counts = None;
+        }
+        self.backend.coordinator.apply_config(new.clone());
+        if actions.requires_rebuild {
+            if self.backend.coordinator.state().mode == IndexMode::Auto {
+                // Automatic mode is hands-off: reconcile immediately, no
+                // prompt. Root-only changes need just a full run — the
+                // walk indexes new roots and the stale sweep drops removed
+                // ones. Anything else (tokenizer, hashing, filters, hidden
+                // files) invalidates stored data and gets the real wipe.
+                let roots_only = {
+                    let mut probe = new.clone();
+                    probe.paths.indexing_paths = self.cfg.paths.indexing_paths.clone();
+                    !diff_actions(&self.cfg, &probe).requires_rebuild
+                };
+                if roots_only {
+                    self.backend.coordinator.reindex_now();
+                } else {
+                    self.backend.coordinator.rebuild_index();
+                }
+            } else {
+                let changes = self
+                    .backend
+                    .coordinator
+                    .check_config_validation(&new)
+                    .ok()
+                    .flatten()
+                    .unwrap_or_default();
+                self.rebuild_prompt = Some(changes);
+            }
+        }
+        self.cfg = new;
+        self.manage.invalidate_editors();
+    }
+
+    fn drain_events(&mut self) {
+        // Streamed search results.
+        loop {
+            match self.backend.search_rx.try_recv() {
+                Ok(update) => self
+                    .search
+                    .apply_update(update, self.cfg.search.display_limit),
+                Err(_) => break,
+            }
+        }
+        // Duplicates worker.
+        if let Some(rx) = &self.backend.dup_job {
+            match rx.try_recv() {
+                Ok(Ok(groups)) => {
+                    self.dups.state = DupState::Loaded(groups);
+                    self.backend.dup_job = None;
+                }
+                Ok(Err(e)) => {
+                    self.dups.state = DupState::Error(e);
+                    self.backend.dup_job = None;
+                }
+                Err(std::sync::mpsc::TryRecvError::Empty) => {}
+                Err(std::sync::mpsc::TryRecvError::Disconnected) => {
+                    self.dups.state = DupState::Error("duplicates scan aborted".into());
+                    self.backend.dup_job = None;
+                }
+            }
+        }
+    }
+
+    fn tick_debounce(&mut self, ctx: &egui::Context) {
+        let Some(edited_at) = self.search.pending_edit else {
+            return;
+        };
+        let debounce = Duration::from_millis(self.cfg.search.debounce_ms);
+        let elapsed = edited_at.elapsed();
+        if elapsed >= debounce {
+            self.search.pending_edit = None;
+            self.start_search();
+        } else {
+            ctx.request_repaint_after(debounce - elapsed);
+        }
+    }
+
+    /// Raise the "live updates are disabled" modal when the watcher has
+    /// given up on the directory budget and at least one indexed folder has
+    /// not been warned about yet.
+    ///
+    /// Keyed on roots rather than a single dismissed flag: a restart should
+    /// stay quiet, but adding a folder changes the trade-off and deserves
+    /// the warning again.
+    fn check_watch_cap_warning(&mut self, state: &IndexerState) {
+        let WatcherStatus::Disabled { reason } = &state.watcher else {
+            // Recovered (e.g. the user trimmed the folder list) — retract a
+            // modal that is no longer true.
+            self.watch_cap_prompt = None;
+            return;
+        };
+        // Only the budget limits warrant a modal. Other failures are
+        // transient and not the user's to act on; they are named in the
+        // status line's tooltip and logged to the Logs tab.
+        if !matches!(
+            reason,
+            WatchError::TooManyDirectories { .. } | WatchError::KernelLimit { .. }
+        ) {
+            return;
+        }
+        if self.watch_cap_prompt.is_some() {
+            return;
+        }
+        let unwarned = self
+            .cfg
+            .paths
+            .indexing_paths
+            .iter()
+            .any(|root| !self.cfg.ui.watch_cap_warned_roots.contains(root));
+        if unwarned {
+            self.watch_cap_prompt = Some(reason.clone());
+        }
+    }
+
+    fn status_bar(&mut self, ctx: &egui::Context) {
+        let state = self.backend.coordinator.state();
+        self.manage.observe(&state.activity);
+        self.check_watch_cap_warning(&state);
+
+        egui::TopBottomPanel::bottom("status-bar").show(ctx, |ui| {
+            ui.horizontal(|ui| {
+                match &state.activity {
+                    IndexingStatus::Idle => {
+                        let mode = match state.mode {
+                            IndexMode::Auto => "Auto",
+                            IndexMode::ManualStopped => "Manual",
+                            IndexMode::ManualRunning => "Manual",
+                        };
+                        let stale = self
+                            .counts
+                            .map(|(at, _)| at.elapsed() > Duration::from_secs(5))
+                            .unwrap_or(true);
+                        if stale {
+                            let db = self.cfg.resolved_database_path();
+                            let counts = index_counts(&db.to_string_lossy()).unwrap_or(
+                                IndexCounts {
+                                    files: 0,
+                                    content_done: 0,
+                                    content_pending: 0,
+                                },
+                            );
+                            self.counts = Some((Instant::now(), counts));
+                        }
+                        let files = self.counts.map(|(_, c)| c.files).unwrap_or(0);
+                        ui.label(
+                            egui::RichText::new(format!(
+                                "Idle · {} · {} files indexed",
+                                mode,
+                                group_thousands(files.max(0) as u64)
+                            ))
+                            .small(),
+                        );
+                    }
+                    IndexingStatus::Error(e) => {
+                        ui.colored_label(
+                            ui.visuals().error_fg_color,
+                            egui::RichText::new(format!("Indexing error: {}", e)).small(),
+                        );
+                    }
+                    IndexingStatus::Stopping => {
+                        ui.label(egui::RichText::new("Stopping indexing…").small());
+                    }
+                    IndexingStatus::Running { roots, .. } => {
+                        let done = roots
+                            .iter()
+                            .filter(|r| r.phase == RootPhase::Done)
+                            .count();
+                        let processed: usize =
+                            roots.iter().map(|r| r.walked + r.extracted).sum();
+                        let totals_known = roots.iter().all(|r| r.walk_total.is_some());
+                        let denominator: usize = roots
+                            .iter()
+                            .map(|r| r.walk_total.unwrap_or(0) + r.extract_total)
+                            .sum();
+
+                        let mut text = if totals_known && denominator > 0 {
+                            let frac =
+                                (processed as f64 / denominator as f64).min(1.0);
+                            format!(
+                                "Indexing {} / {} ({:.0}%)",
+                                group_thousands(processed as u64),
+                                group_thousands(denominator as u64),
+                                frac * 100.0
+                            )
+                        } else {
+                            format!(
+                                "Indexing · {} files",
+                                group_thousands(processed as u64)
+                            )
+                        };
+                        if roots.len() > 1 {
+                            text.push_str(&format!(" · {}/{} roots done", done, roots.len()));
+                        }
+                        if let Some(rate) = self.manage.speed.files_per_sec() {
+                            text.push_str(&format!(" · {}", crate::format::fmt_rate(rate)));
+                        }
+                        let active: usize = roots.iter().map(|r| r.active_workers).sum();
+                        let total_workers: usize =
+                            roots.iter().map(|r| r.total_workers).sum();
+                        if total_workers > 0 {
+                            text.push_str(&format!(" · {}/{} workers", active, total_workers));
+                        }
+                        ui.label(egui::RichText::new(text).small());
+                        if totals_known && denominator > 0 {
+                            let frac =
+                                (processed as f32 / denominator as f32).clamp(0.0, 1.0);
+                            ui.add(egui::ProgressBar::new(frac).desired_width(120.0));
+                        } else {
+                            ui.add(egui::Spinner::new().size(12.0));
+                        }
+                    }
+                }
+
+                // Right corner: search result count.
+                ui.with_layout(egui::Layout::right_to_left(egui::Align::Center), |ui| {
+                    if self.tab == Tab::Search {
+                        if let Some(label) = self.search.result_count_label() {
+                            ui.label(egui::RichText::new(label).small().weak());
+                        }
+                    }
+                });
+            });
+        });
+
+        // Keep painting while anything is moving.
+        if !matches!(state.activity, IndexingStatus::Idle | IndexingStatus::Error(_)) {
+            ctx.request_repaint_after(Duration::from_millis(250));
+        }
+        // Watcher registration walks every root, so its verdict can land
+        // minutes after startup. Without this the warning would wait for
+        // the user to happen to move the mouse.
+        if matches!(state.watcher, WatcherStatus::Starting) {
+            ctx.request_repaint_after(Duration::from_millis(500));
+        }
+    }
+
+    fn rebuild_prompt_ui(&mut self, ctx: &egui::Context) {
+        let Some(changes) = &self.rebuild_prompt else {
+            return;
+        };
+        let changes = changes.clone();
+        let mut close = false;
+        egui::Window::new("Rebuild index?")
+            .collapsible(false)
+            .resizable(false)
+            .default_width(560.0)
+            .anchor(egui::Align2::CENTER_CENTER, [0.0, 0.0])
+            .show(ctx, |ui| {
+                ui.label("These settings differ from what the index was built with:");
+                ui.add_space(4.0);
+                if changes.is_empty() {
+                    ui.monospace("indexing settings changed");
+                }
+                for change in &changes {
+                    ui.strong(format!("{}:", change.key));
+                    // Multi-line values (roots, patterns, extensions) are
+                    // newline-joined — side-by-side columns keep before and
+                    // after readable instead of one run-on arrow line.
+                    ui.columns(2, |cols| {
+                        cols[0].label(
+                            egui::RichText::new("index was built with").small().weak(),
+                        );
+                        cols[0].monospace(display_value(&change.stored));
+                        cols[1].label(egui::RichText::new("config now says").small().weak());
+                        cols[1].monospace(display_value(&change.current));
+                    });
+                    ui.add_space(6.0);
+                }
+                ui.label(
+                    egui::RichText::new(
+                        "A full rebuild applies them everywhere. Until then, existing \
+                         entries keep the old settings.",
+                    )
+                    .small()
+                    .weak(),
+                );
+                ui.horizontal(|ui| {
+                    if ui.button("Rebuild now").clicked() {
+                        self.backend.coordinator.rebuild_index();
+                        close = true;
+                    }
+                    if ui.button("Later").clicked() {
+                        close = true;
+                    }
+                });
+            });
+        if close {
+            self.rebuild_prompt = None;
+        }
+    }
+}
+
+impl QuickSearchApp {
+    fn nested_prompt_ui(&mut self, ctx: &egui::Context) {
+        let Some(pairs) = &self.nested_prompt else {
+            return;
+        };
+        let pairs = pairs.clone();
+        let mut close = false;
+        egui::Window::new("Indexed folders may not be nested")
+            .collapsible(false)
+            .resizable(false)
+            .anchor(egui::Align2::CENTER_CENTER, [0.0, 0.0])
+            .show(ctx, |ui| {
+                ui.label(
+                    "Each root is indexed by its own worker pool, so one root \
+                     may not contain another. Fix the folder list below:",
+                );
+                for (child, parent) in &pairs {
+                    ui.monospace(format!("{}  ⊂  {}", child, parent));
+                }
+                ui.label(
+                    egui::RichText::new(
+                        "Indexing stays paused until the overlap is removed and \
+                         the list is applied.",
+                    )
+                    .small()
+                    .weak(),
+                );
+                if ui.button("Fix folders").clicked() {
+                    close = true;
+                }
+            });
+        if close {
+            self.nested_prompt = None;
+            self.tab = Tab::Manage;
+        }
+    }
+
+    fn watch_cap_prompt_ui(&mut self, ctx: &egui::Context) {
+        let Some(reason) = &self.watch_cap_prompt else {
+            return;
+        };
+        let reason = reason.clone();
+        let mut close = false;
+        egui::Window::new("Live index updating is disabled")
+            .collapsible(false)
+            .resizable(false)
+            .anchor(egui::Align2::CENTER_CENTER, [0.0, 0.0])
+            .show(ctx, |ui| {
+                ui.set_max_width(420.0);
+                match &reason {
+                    WatchError::TooManyDirectories { cap, .. } => {
+                        ui.label(format!(
+                            "Your indexed folders contain more than {} directories. The \
+                             system limits how many folders can be watched for changes at \
+                             once, so QuickSearch cannot update the index as files change.",
+                            group_thousands(*cap as u64),
+                        ));
+                    }
+                    WatchError::KernelLimit { registered } => {
+                        ui.label(format!(
+                            "The system ran out of folder watches after {} directories, so \
+                             QuickSearch cannot update the index as files change.",
+                            group_thousands(*registered as u64),
+                        ));
+                    }
+                    WatchError::Other(msg) => {
+                        ui.label(format!("Live updates are unavailable: {}", msg));
+                    }
+                }
+                ui.add_space(4.0);
+                ui.label(format!(
+                    "The index is rebuilt every {} instead. Searches keep working; \
+                     recent changes may take that long to appear.",
+                    fmt_interval(self.cfg.indexing.reindex_interval_minutes),
+                ));
+                ui.label(
+                    egui::RichText::new(
+                        "To restore live updates, index fewer folders or exclude large \
+                         subfolders under Filters on the Manage Index tab.",
+                    )
+                    .small()
+                    .weak(),
+                );
+                if ui.button("OK").clicked() {
+                    close = true;
+                }
+            });
+        if close {
+            self.watch_cap_prompt = None;
+            for root in &self.cfg.paths.indexing_paths {
+                if !self.cfg.ui.watch_cap_warned_roots.contains(root) {
+                    self.cfg.ui.watch_cap_warned_roots.push(root.clone());
+                }
+            }
+            if let Err(e) = self.cfg.save() {
+                self.config_error = Some(e);
+            }
+        }
+    }
+
+    fn clear_prompt_ui(&mut self, ctx: &egui::Context) {
+        if !self.clear_prompt {
+            return;
+        }
+        let mut close = false;
+        egui::Window::new("Clear index?")
+            .collapsible(false)
+            .resizable(false)
+            .anchor(egui::Align2::CENTER_CENTER, [0.0, 0.0])
+            .show(ctx, |ui| {
+                ui.label("This deletes the search index database. Your files are not touched.");
+                ui.label(
+                    egui::RichText::new(
+                        "Indexing switches to manual until you start it again or return                          to automatic mode.",
+                    )
+                    .small()
+                    .weak(),
+                );
+                ui.horizontal(|ui| {
+                    if ui
+                        .button(
+                            egui::RichText::new("Delete index")
+                                .color(ui.visuals().error_fg_color),
+                        )
+                        .clicked()
+                    {
+                        self.backend.coordinator.clear_index();
+                        self.counts = None;
+                        self.dups.state = DupState::NotLoaded;
+                        close = true;
+                    }
+                    if ui.button("Cancel").clicked() {
+                        close = true;
+                    }
+                });
+            });
+        if close {
+            self.clear_prompt = false;
+        }
+    }
+}
+
+
+/// A stored/current config value for the rebuild prompt; list values are
+/// already newline-joined and render as-is, empty means unset.
+fn display_value(value: &str) -> String {
+    if value.trim().is_empty() {
+        "(none)".to_string()
+    } else {
+        value.to_string()
+    }
+}
+
+/// Keep the configured UI scale within sane, recoverable bounds.
+fn clamp_scale(scale: f32) -> f32 {
+    if scale.is_finite() {
+        scale.clamp(0.5, 2.5)
+    } else {
+        1.1
+    }
+}
+
+impl eframe::App for QuickSearchApp {
+    fn update(&mut self, ctx: &egui::Context, _frame: &mut eframe::Frame) {
+        self.drain_events();
+        self.tick_debounce(ctx);
+        self.status_bar(ctx);
+
+        let previous_tab = self.tab;
+        egui::TopBottomPanel::top("tab-strip").show(ctx, |ui| {
+            ui.horizontal(|ui| {
+                ui.selectable_value(&mut self.tab, Tab::Search, "Search");
+                ui.selectable_value(&mut self.tab, Tab::Manage, "Manage Index");
+                ui.selectable_value(&mut self.tab, Tab::Duplicates, "Duplicates");
+                ui.selectable_value(&mut self.tab, Tab::Logs, "Logs");
+                ui.with_layout(egui::Layout::right_to_left(egui::Align::Center), |ui| {
+                    if ui.button("⚙").on_hover_text("Options").clicked() {
+                        if self.options.open {
+                            self.options.open = false;
+                        } else {
+                            self.options.open_with(&self.cfg);
+                        }
+                    }
+                });
+            });
+        });
+        // Entering the Duplicates tab kicks off a fresh scan.
+        if self.tab == Tab::Duplicates && previous_tab != Tab::Duplicates {
+            self.start_duplicates_scan(ctx);
+        }
+
+        if let Some(err) = &self.config_error {
+            let err = err.clone();
+            egui::TopBottomPanel::top("config-error").show(ctx, |ui| {
+                ui.horizontal(|ui| {
+                    ui.colored_label(
+                        ui.visuals().error_fg_color,
+                        format!("Config problem: {} (using defaults)", err),
+                    );
+                    if ui.small_button("Dismiss").clicked() {
+                        self.config_error = None;
+                    }
+                });
+            });
+        }
+
+        egui::CentralPanel::default().show(ctx, |ui| match self.tab {
+            Tab::Search => {
+                let actions = self.search.ui(ui);
+                if let Some(fuzzy) = actions.save_fuzzy_default {
+                    self.cfg.search.fuzzy_default = fuzzy;
+                    if let Err(e) = self.cfg.save() {
+                        self.config_error = Some(e);
+                    }
+                }
+                if let Some(pattern) = actions.persist_ignore {
+                    let mut new_cfg = self.cfg.clone();
+                    if !new_cfg.indexing.ignore_patterns.contains(&pattern) {
+                        new_cfg.indexing.ignore_patterns.push(pattern);
+                        self.apply_new_config(ctx, new_cfg);
+                    }
+                }
+                if actions.rerun {
+                    self.start_search();
+                }
+            }
+            Tab::Manage => {
+                let state = self.backend.coordinator.state();
+                let actions = self.manage.ui(ui, &state, &self.cfg);
+                if actions.start_now {
+                    self.backend.coordinator.reindex_now();
+                }
+                if actions.stop {
+                    self.backend.coordinator.set_mode(IndexMode::ManualStopped);
+                }
+                if actions.auto {
+                    self.backend.coordinator.set_mode(IndexMode::Auto);
+                }
+                if actions.clear_index {
+                    self.clear_prompt = true;
+                }
+                if actions.start_now || actions.stop || actions.auto {
+                    // Keep repainting while the command lands so the state
+                    // change is visible without wiggling the mouse — fast
+                    // runs otherwise flash by between frames.
+                    ui.ctx().request_repaint_after(Duration::from_millis(100));
+                }
+                if let Some(new_cfg) = actions.apply_config {
+                    self.apply_new_config(ctx, new_cfg);
+                }
+            }
+            Tab::Duplicates => {
+                let actions = self.dups.ui(ui);
+                if actions.refresh {
+                    self.start_duplicates_scan(ctx);
+                }
+            }
+            Tab::Logs => self.logs.ui(ui),
+        });
+
+        if let Some(new_cfg) = self.options.ui(ctx, &self.cfg) {
+            self.apply_new_config(ctx, new_cfg);
+        }
+        self.rebuild_prompt_ui(ctx);
+        self.clear_prompt_ui(ctx);
+        self.nested_prompt_ui(ctx);
+        self.watch_cap_prompt_ui(ctx);
+    }
+
+    fn on_exit(&mut self, _gl: Option<&eframe::glow::Context>) {
+        self.backend.shutdown();
+    }
+}

crates/quicksearch-gui/src/backend.rs

@@ -0,0 +1,71 @@
+//! Wiring between the egui thread and the core services.
+//!
+//! All communication is non-blocking from the UI's point of view:
+//! searches stream over an mpsc receiver drained each frame, indexing
+//! state is polled, and the duplicates query runs on a throwaway worker
+//! thread. Core threads wake the UI through `ctx.request_repaint()`.
+
+use std::sync::{mpsc, Arc};
+
+use quicksearch_core::config::Config;
+use quicksearch_core::coordinator::IndexCoordinator;
+use quicksearch_core::search::{DuplicateGroup, SearchService, SearchUpdate};
+use quicksearch_core::shutdown;
+
+pub struct Backend {
+    pub coordinator: Arc<IndexCoordinator>,
+    pub search: Option<SearchService>,
+    pub search_rx: mpsc::Receiver<SearchUpdate>,
+    pub dup_job: Option<mpsc::Receiver<Result<Vec<DuplicateGroup>, String>>>,
+}
+
+impl Backend {
+    pub fn start(config: &Config, ctx: egui::Context) -> Result<Backend, String> {
+        let coordinator = Arc::new(IndexCoordinator::start(config.clone())?);
+        if let Err(e) = shutdown::install_signal_handler(coordinator.clone()) {
+            quicksearch_core::log_warn!("failed to install signal handler: {}", e);
+        }
+
+        let repaint_ctx = ctx.clone();
+        let (search, search_rx) = SearchService::new(
+            config.resolved_database_path(),
+            Arc::new(move || repaint_ctx.request_repaint()),
+        );
+
+        Ok(Backend {
+            coordinator,
+            search: Some(search),
+            search_rx,
+            dup_job: None,
+        })
+    }
+
+    pub fn search(&self) -> &SearchService {
+        self.search.as_ref().expect("search service alive")
+    }
+
+    /// Kick off (or restart) the duplicates listing on a worker thread.
+    pub fn start_duplicates(&mut self, config: &Config, ctx: egui::Context) {
+        let (tx, rx) = mpsc::channel();
+        let db = config.resolved_database_path();
+        std::thread::spawn(move || {
+            let result = quicksearch_core::search::find_duplicate_groups(
+                &db.to_string_lossy(),
+                500,
+                0,
+            );
+            let _ = tx.send(result);
+            ctx.request_repaint();
+        });
+        self.dup_job = Some(rx);
+    }
+
+    /// Join the search worker and stop the coordinator. Called once from
+    /// `on_exit`.
+    pub fn shutdown(&mut self) {
+        if let Some(search) = self.search.take() {
+            search.shutdown();
+        }
+        self.coordinator.shutdown();
+    }
+}

crates/quicksearch-gui/src/cli.rs

@@ -0,0 +1,230 @@
+//! Terminal query mode: `quicksearch [FLAGS] <query terms...>` runs the
+//! same ranked cascade the GUI uses and prints results to stdout. With no
+//! positional arguments the binary opens the GUI instead.
+
+use std::io::IsTerminal;
+use std::sync::atomic::AtomicU64;
+
+use quicksearch_core::config::Config;
+use quicksearch_core::db;
+use quicksearch_core::query::split::split_for_cascade;
+use quicksearch_core::search::{cascade, SearchHit, SearchOptions};
+
+use crate::format::{fmt_mtime, human_size};
+
+pub(crate) const USAGE: &str = "\
+QuickSearch: indexed file search
+
+USAGE:
+    quicksearch                          open the GUI
+    quicksearch [FLAGS] <query terms>    search from the terminal
+                                         (Windows: quicksearch-cli)
+
+FLAGS:
+    --fuzzy         also run the fuzzy filename/full-text passes
+    --limit <N>     maximum results (default: [search].display_limit)
+    --long          rank, size, mtime, and snippets instead of bare paths
+    -h, --help      this help
+
+Query syntax matches the GUI: plain words form one phrase; filters like
+type:Document, modified:>=2024-01-01, path:/dir, mime:application/pdf,
+name:frag combine with it.";
+
+/// Parse argv; `Some(exit_code)` when the invocation was CLI-mode (query
+/// or --help), `None` to open the GUI.
+///
+/// Invariant: terminal mode never builds an [`IndexCoordinator`], so it
+/// starts no filesystem watcher, no background threads, and consumes no
+/// inotify watches — a one-shot query must not leave anything running or
+/// compete for the per-user watch budget with a running GUI. It opens the
+/// database, queries, prints, and exits. Keep it that way: the coordinator
+/// belongs to the GUI path in `backend.rs` alone.
+///
+/// [`IndexCoordinator`]: quicksearch_core::coordinator::IndexCoordinator
+pub fn maybe_run_cli() -> Option<i32> {
+    let args: Vec<String> = std::env::args().skip(1).collect();
+
+    let mut fuzzy = false;
+    let mut long = false;
+    let mut limit: Option<usize> = None;
+    let mut terms: Vec<String> = Vec::new();
+
+    let mut it = args.into_iter();
+    while let Some(arg) = it.next() {
+        match arg.as_str() {
+            "-h" | "--help" => {
+                println!("{}", USAGE);
+                return Some(0);
+            }
+            "--fuzzy" => fuzzy = true,
+            "--long" => long = true,
+            "--limit" => match it.next().and_then(|v| v.parse().ok()) {
+                Some(n) => limit = Some(n),
+                None => {
+                    eprintln!("--limit requires a number\n\n{}", USAGE);
+                    return Some(2);
+                }
+            },
+            other if other.starts_with("--limit=") => {
+                match other["--limit=".len()..].parse() {
+                    Ok(n) => limit = Some(n),
+                    Err(_) => {
+                        eprintln!("--limit requires a number\n\n{}", USAGE);
+                        return Some(2);
+                    }
+                }
+            }
+            other if other.starts_with('-') && terms.is_empty() => {
+                // Unknown flags without a query fall through to the GUI
+                // (they may be eframe/winit flags).
+                return None;
+            }
+            other => terms.push(other.to_string()),
+        }
+    }
+
+    if terms.is_empty() {
+        return None;
+    }
+    Some(run_query(&terms.join(" "), fuzzy, limit, long))
+}
+
+fn run_query(query: &str, fuzzy: bool, limit: Option<usize>, long: bool) -> i32 {
+    let config = match Config::load() {
+        Ok(c) => c,
+        Err(e) => {
+            eprintln!("config: {}", e);
+            return 2;
+        }
+    };
+    let db_path = config.resolved_database_path();
+    // Read-write purely so SQLite may create the WAL shared-memory file
+    // when no other process has the index open; nothing is written.
+    let conn = match db::open_existing(&db_path.to_string_lossy(), true) {
+        Ok(c) => c,
+        Err(e) => {
+            eprintln!(
+                "No usable index at {}; run the GUI once to build it.\n({})",
+                db_path.display(),
+                e
+            );
+            return 2;
+        }
+    };
+
+    let split = match split_for_cascade(query) {
+        Ok(s) => s,
+        Err(e) => {
+            eprintln!("query: {}", e);
+            return 2;
+        }
+    };
+
+    if fuzzy {
+        if let Some(warning) = config.search.fuzzy_edits_warning() {
+            eprintln!("warning: {}", warning);
+        }
+    }
+    let options = SearchOptions {
+        fuzzy,
+        fuzzy_max_edits: config.search.fuzzy_max_edits,
+        limit: limit.unwrap_or(config.search.display_limit),
+        batch: config.search.results_per_page.max(1),
+        session_ignores: Vec::new(),
+    };
+    let latest = AtomicU64::new(1);
+    let mut hits: Vec<SearchHit> = Vec::new();
+    let outcome = cascade::run(&conn, &split, &options, 1, &latest, &mut |batch| {
+        hits.extend(batch)
+    });
+
+    match outcome {
+        Ok(Some(outcome)) => {
+            let color = long && std::io::stdout().is_terminal() && enable_vt();
+            for hit in &hits {
+                if long {
+                    println!(
+                        "{:6.3}  {:>9}  {}  {}",
+                        hit.rank,
+                        human_size(hit.size),
+                        fmt_mtime(hit.mtime),
+                        hit.path
+                    );
+                    if let Some(snip) = &hit.snippet {
+                        println!("        {}", render_snippet(snip, color));
+                    }
+                } else {
+                    println!("{}", hit.path);
+                }
+            }
+            if outcome.limited {
+                eprintln!("(truncated at {} results; raise with --limit)", hits.len());
+            }
+            0
+        }
+        Ok(None) => 0, // unreachable: nothing cancels a CLI search
+        Err(e) => {
+            eprintln!("search: {}", e);
+            2
+        }
+    }
+}
+
+/// Whether ANSI escapes will actually render.
+///
+/// Always true where the terminal is ANSI by nature. On Windows the console
+/// only interprets escapes once `ENABLE_VIRTUAL_TERMINAL_PROCESSING` is set:
+/// Windows Terminal and Windows 11 have it already, older conhost needs it
+/// turned on, and anything that refuses gets plain text rather than a screen
+/// full of `\x1b[1m`.
+#[cfg(not(windows))]
+fn enable_vt() -> bool {
+    true
+}
+
+#[cfg(windows)]
+fn enable_vt() -> bool {
+    use windows_sys::Win32::Foundation::INVALID_HANDLE_VALUE;
+    use windows_sys::Win32::System::Console::{
+        GetConsoleMode, GetStdHandle, SetConsoleMode, ENABLE_VIRTUAL_TERMINAL_PROCESSING,
+        STD_OUTPUT_HANDLE,
+    };
+
+    unsafe {
+        let handle = GetStdHandle(STD_OUTPUT_HANDLE);
+        if handle.is_null() || handle == INVALID_HANDLE_VALUE {
+            return false;
+        }
+        let mut mode = 0;
+        if GetConsoleMode(handle, &mut mode) == 0 {
+            return false;
+        }
+        mode & ENABLE_VIRTUAL_TERMINAL_PROCESSING != 0
+            || SetConsoleMode(handle, mode | ENABLE_VIRTUAL_TERMINAL_PROCESSING) != 0
+    }
+}
+
+/// One-line snippet with matches emphasized (ANSI bold on TTYs).
+fn render_snippet(snip: &quicksearch_core::snippet::Snippet, color: bool) -> String {
+    let mut out = String::new();
+    if snip.truncated_start {
+        out.push('…');
+    }
+    let mut cursor = 0;
+    for &(start, end) in &snip.ranges {
+        out.push_str(&snip.window[cursor..start]);
+        if color {
+            out.push_str("\x1b[1m");
+            out.push_str(&snip.window[start..end]);
+            out.push_str("\x1b[0m");
+        } else {
+            out.push_str(&snip.window[start..end]);
+        }
+        cursor = end;
+    }
+    out.push_str(&snip.window[cursor..]);
+    if snip.truncated_end {
+        out.push('…');
+    }
+    out.replace(['\n', '\r'], " ")
+}

crates/quicksearch-gui/src/cli_main.rs

@@ -0,0 +1,21 @@
+//! `quicksearch-cli <query>` — terminal search, and nothing else.
+//!
+//! A console-subsystem binary, so redirection, pipes, exit codes, and the
+//! shell waiting for the process all behave normally. `src/cli.rs` and
+//! `src/format.rs` are shared with the GUI binary by compiling them into both;
+//! neither touches egui, so there is nothing to split out into a library.
+
+mod cli;
+// The GUI uses more of this module than the CLI does.
+#[allow(dead_code)]
+mod format;
+
+fn main() {
+    // `maybe_run_cli` returns `None` for "no query given", which the combined
+    // binary treats as "open the GUI". This one has no GUI to fall back to.
+    let code = cli::maybe_run_cli().unwrap_or_else(|| {
+        eprintln!("{}", cli::USAGE);
+        2
+    });
+    std::process::exit(code);
+}

crates/quicksearch-gui/src/duplicates_tab.rs

@@ -0,0 +1,111 @@
+//! The Duplicates tab: groups of files sharing a content hash.
+
+use quicksearch_core::search::DuplicateGroup;
+
+use crate::format::{group_thousands, human_size};
+use crate::platform;
+
+pub enum DupState {
+    NotLoaded,
+    Loading,
+    Loaded(Vec<DuplicateGroup>),
+    Error(String),
+}
+
+pub struct DuplicatesTab {
+    pub state: DupState,
+}
+
+/// What the tab asks the app to do after this frame.
+#[derive(Default)]
+pub struct DuplicatesActions {
+    pub refresh: bool,
+}
+
+impl DuplicatesTab {
+    pub fn new() -> DuplicatesTab {
+        DuplicatesTab {
+            state: DupState::NotLoaded,
+        }
+    }
+
+    pub fn ui(&mut self, ui: &mut egui::Ui) -> DuplicatesActions {
+        let mut actions = DuplicatesActions::default();
+
+        ui.horizontal(|ui| {
+            let loading = matches!(self.state, DupState::Loading);
+            if ui.add_enabled(!loading, egui::Button::new("Refresh")).clicked() {
+                actions.refresh = true;
+            }
+            if loading {
+                ui.add(egui::Spinner::new().size(16.0));
+                ui.label("Scanning for duplicates…");
+            }
+        });
+        ui.separator();
+
+        match &self.state {
+            DupState::NotLoaded => {
+                ui.label(
+                    egui::RichText::new("Press Refresh to scan the index for duplicate files.")
+                        .weak(),
+                );
+            }
+            DupState::Loading => {}
+            DupState::Error(e) => {
+                ui.colored_label(ui.visuals().error_fg_color, e);
+            }
+            DupState::Loaded(groups) => {
+                if groups.is_empty() {
+                    ui.label("No duplicate files found.");
+                    return actions;
+                }
+                if groups.len() == 500 {
+                    ui.label(
+                        egui::RichText::new("Showing the 500 largest groups.").small().weak(),
+                    );
+                }
+                egui::ScrollArea::vertical().auto_shrink([false; 2]).show(ui, |ui| {
+                    for (i, group) in groups.iter().enumerate() {
+                        let name = group
+                            .members
+                            .first()
+                            .map(|m| m.1.as_str())
+                            .unwrap_or("(unknown)");
+                        let title = format!(
+                            "{} × {}: {} reclaimable ({} total)",
+                            group_thousands(group.count as u64),
+                            name,
+                            human_size(group.redundant_size.max(0) as u64),
+                            human_size(group.total_size.max(0) as u64),
+                        );
+                        egui::CollapsingHeader::new(title).id_salt(i).show(ui, |ui| {
+                            for (_, _, path, size, _) in &group.members {
+                                ui.horizontal(|ui| {
+                                    ui.label(human_size(*size));
+                                    let response = ui
+                                        .add(egui::Label::new(egui::RichText::new(path).monospace())
+                                            .sense(egui::Sense::click()));
+                                    if response.double_clicked() {
+                                        platform::open_file(path);
+                                    }
+                                    response.context_menu(|ui| {
+                                        if ui.button("Open").clicked() {
+                                            platform::open_file(path);
+                                            ui.close();
+                                        }
+                                        if ui.button("Open containing folder").clicked() {
+                                            platform::reveal_in_folder(path);
+                                            ui.close();
+                                        }
+                                    });
+                                });
+                            }
+                        });
+                    }
+                });
+            }
+        }
+        actions
+    }
+}

crates/quicksearch-gui/src/format.rs

@@ -0,0 +1,202 @@
+//! Small display formatters shared across tabs.
+
+/// Human-readable byte size: `999 B`, `1.2 KB`, `4.7 MB`, `1.3 GB`.
+pub fn human_size(bytes: u64) -> String {
+    const UNITS: [&str; 5] = ["B", "KB", "MB", "GB", "TB"];
+    let mut value = bytes as f64;
+    let mut unit = 0;
+    while value >= 1000.0 && unit < UNITS.len() - 1 {
+        value /= 1000.0;
+        unit += 1;
+    }
+    if unit == 0 {
+        format!("{} B", bytes)
+    } else {
+        format!("{:.1} {}", value, UNITS[unit])
+    }
+}
+
+/// `YYYY-MM-DD HH:MM` in local time; raw seconds if out of range.
+pub fn fmt_mtime(unix_secs: i64) -> String {
+    use chrono::TimeZone;
+    match chrono::Local.timestamp_opt(unix_secs, 0) {
+        chrono::LocalResult::Single(dt) => dt.format("%Y-%m-%d %H:%M").to_string(),
+        _ => unix_secs.to_string(),
+    }
+}
+
+/// Relative time for recent events, absolute for old ones: "just now",
+/// "5 min ago", "3 h ago", else `YYYY-MM-DD HH:MM`. Gives instant
+/// feedback that an action (like a fast index run) actually happened.
+pub fn fmt_ago(unix_secs: u64) -> String {
+    let now = std::time::SystemTime::now()
+        .duration_since(std::time::UNIX_EPOCH)
+        .map(|d| d.as_secs())
+        .unwrap_or(0);
+    let age = now.saturating_sub(unix_secs);
+    if age < 60 {
+        "just now".to_string()
+    } else if age < 3600 {
+        format!("{} min ago", age / 60)
+    } else if age < 86_400 {
+        format!("{} h ago", age / 3600)
+    } else {
+        fmt_mtime(unix_secs as i64)
+    }
+}
+
+/// A configured interval as a phrase to drop after "every": `90 min`,
+/// `24 h`, `3 days`. Used where the periodic reindex is the only thing
+/// refreshing the index, so the user can judge how stale it may get.
+pub fn fmt_interval(minutes: u64) -> String {
+    if minutes == 0 {
+        // The scheduler treats 0 as always-due.
+        return "run".to_string();
+    }
+    if minutes < 60 {
+        return format!("{} min", minutes);
+    }
+    if minutes.is_multiple_of(1440) {
+        let days = minutes / 1440;
+        return if days == 1 {
+            // "24 h" reads better than "1 day" for the shipped default.
+            "24 h".to_string()
+        } else {
+            format!("{} days", days)
+        };
+    }
+    if minutes.is_multiple_of(60) {
+        return format!("{} h", minutes / 60);
+    }
+    format!("{} h {} min", minutes / 60, minutes % 60)
+}
+
+/// Group thousands for counts: `1,234,567`.
+pub fn group_thousands(n: u64) -> String {
+    let digits = n.to_string();
+    let mut out = String::with_capacity(digits.len() + digits.len() / 3);
+    for (i, c) in digits.chars().enumerate() {
+        if i > 0 && (digits.len() - i) % 3 == 0 {
+            out.push(',');
+        }
+        out.push(c);
+    }
+    out
+}
+
+/// Files/sec display. Never renders a nonzero rate as "0.0": slow rates
+/// switch to a per-minute figure.
+pub fn fmt_rate(files_per_sec: f64) -> String {
+    if files_per_sec <= 0.0 {
+        "0 files/s".to_string()
+    } else if files_per_sec >= 10.0 {
+        format!("{:.0} files/s", files_per_sec)
+    } else if files_per_sec >= 1.0 {
+        format!("{:.1} files/s", files_per_sec)
+    } else {
+        format!("{:.0} files/min", (files_per_sec * 60.0).max(1.0))
+    }
+}
+
+/// Search duration: milliseconds under a second, seconds above.
+pub fn fmt_elapsed(d: std::time::Duration) -> String {
+    let ms = d.as_millis();
+    if ms >= 1000 {
+        format!("{:.1} s", d.as_secs_f64())
+    } else {
+        format!("{} ms", ms)
+    }
+}
+
+/// Middle-truncate a path to at most `max_chars` characters.
+pub fn middle_truncate(s: &str, max_chars: usize) -> String {
+    let chars: Vec<char> = s.chars().collect();
+    if chars.len() <= max_chars || max_chars < 5 {
+        return s.to_string();
+    }
+    let keep = max_chars - 1;
+    let head = keep / 2;
+    let tail = keep - head;
+    let mut out: String = chars[..head].iter().collect();
+    out.push('…');
+    out.extend(&chars[chars.len() - tail..]);
+    out
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn sizes() {
+        assert_eq!(human_size(0), "0 B");
+        assert_eq!(human_size(999), "999 B");
+        assert_eq!(human_size(1200), "1.2 KB");
+        assert_eq!(human_size(4_700_000), "4.7 MB");
+        assert_eq!(human_size(1_300_000_000), "1.3 GB");
+    }
+
+    #[test]
+    fn intervals() {
+        assert_eq!(fmt_interval(0), "run");
+        assert_eq!(fmt_interval(1), "1 min");
+        assert_eq!(fmt_interval(59), "59 min");
+        assert_eq!(fmt_interval(60), "1 h");
+        assert_eq!(fmt_interval(90), "1 h 30 min");
+        assert_eq!(fmt_interval(120), "2 h");
+        assert_eq!(fmt_interval(1440), "24 h", "the shipped default");
+        assert_eq!(fmt_interval(2880), "2 days");
+        assert_eq!(fmt_interval(10_080), "7 days");
+    }
+
+    #[test]
+    fn thousands() {
+        assert_eq!(group_thousands(0), "0");
+        assert_eq!(group_thousands(999), "999");
+        assert_eq!(group_thousands(1000), "1,000");
+        assert_eq!(group_thousands(1_234_567), "1,234,567");
+    }
+
+    #[test]
+    fn rates_never_show_zero_for_nonzero() {
+        assert_eq!(fmt_rate(0.0), "0 files/s");
+        assert_eq!(fmt_rate(2543.0), "2543 files/s");
+        // Not 3.14: clippy reads that as a botched `PI` and denies it.
+        assert_eq!(fmt_rate(3.12), "3.1 files/s");
+        assert_eq!(fmt_rate(0.4), "24 files/min");
+        assert_eq!(fmt_rate(0.001), "1 files/min", "floor at 1/min, never 0.0");
+    }
+
+    #[test]
+    fn elapsed_units() {
+        use std::time::Duration;
+        assert_eq!(fmt_elapsed(Duration::from_millis(0)), "0 ms");
+        assert_eq!(fmt_elapsed(Duration::from_millis(7)), "7 ms");
+        assert_eq!(fmt_elapsed(Duration::from_millis(999)), "999 ms");
+        assert_eq!(fmt_elapsed(Duration::from_millis(1000)), "1.0 s");
+        assert_eq!(fmt_elapsed(Duration::from_millis(2340)), "2.3 s");
+    }
+
+    #[test]
+    fn ago_buckets() {
+        let now = std::time::SystemTime::now()
+            .duration_since(std::time::UNIX_EPOCH)
+            .unwrap()
+            .as_secs();
+        assert_eq!(fmt_ago(now), "just now");
+        assert_eq!(fmt_ago(now - 59), "just now");
+        assert_eq!(fmt_ago(now - 120), "2 min ago");
+        assert_eq!(fmt_ago(now - 7200), "2 h ago");
+        assert!(fmt_ago(now - 200_000).contains('-'), "old = absolute date");
+    }
+
+    #[test]
+    fn truncation() {
+        assert_eq!(middle_truncate("short", 20), "short");
+        let t = middle_truncate("/very/long/path/to/some/file.txt", 15);
+        assert!(t.chars().count() <= 15);
+        assert!(t.contains('…'));
+        assert!(t.starts_with("/very"));
+        assert!(t.ends_with("e.txt"));
+    }
+}

crates/quicksearch-gui/src/frontend.rs

@@ -1,385 +0,0 @@
-#![allow(non_snake_case)]
-
-use std::sync::Arc;
-use std::collections::VecDeque;
-use std::time::Instant;
-use dioxus::prelude::*;
-use quicksearch_core::indexing::{IndexingService, IndexingStatus};
-use quicksearch_core::config::Config;
-
-#[derive(Debug, Clone)]
-struct SpeedDataPoint {
-    timestamp: Instant,
-    files_processed: usize,
-}
-
-struct SpeedTracker {
-    data_points: VecDeque<SpeedDataPoint>,
-}
-
-impl SpeedTracker {
-    fn new() -> Self {
-        Self {
-            data_points: VecDeque::new(),
-        }
-    }
-
-    fn add_data_point(&mut self, files_processed: usize) {
-        let now = Instant::now();
-        self.data_points.push_back(SpeedDataPoint {
-            timestamp: now,
-            files_processed,
-        });
-        
-        // Prune data points older than 1 second
-        while let Some(front) = self.data_points.front() {
-            if now.duration_since(front.timestamp).as_secs_f64() > 1.0 {
-                self.data_points.pop_front();
-            } else {
-                break;
-            }
-        }
-    }
-
-    fn calculate_files_per_second(&self) -> Option<f64> {
-        if self.data_points.len() < 2 {
-            return None;
-        }
-
-        let newest = self.data_points.back()?;
-        let oldest = self.data_points.front()?;
-        
-        let time_span = newest.timestamp.duration_since(oldest.timestamp).as_secs_f64();
-        if time_span < 0.1 { // Avoid division by very small numbers
-            return None;
-        }
-        
-        let files_diff = newest.files_processed.saturating_sub(oldest.files_processed);
-        Some(files_diff as f64 / time_span)
-    }
-}
-
-#[derive(Props, Clone)]
-pub struct AppProps {
-    pub indexing_service: Arc<IndexingService>,
-    pub config: Config,
-}
-
-impl PartialEq for AppProps {
-    fn eq(&self, other: &Self) -> bool {
-        Arc::ptr_eq(&self.indexing_service, &other.indexing_service)
-            && self.config.paths.indexing_paths == other.config.paths.indexing_paths
-            && self.config.paths.database_path == other.config.paths.database_path
-    }
-}
-
-pub fn App(props: AppProps) -> Element {
-    // Multi-root support: the GUI's single line edit holds one path per line.
-    // Empty lines are ignored. Core still de-duplicates and handles nested roots.
-    let mut indexing_path = use_signal(|| props.config.paths.indexing_paths.join("\n"));
-    let mut db_path = use_signal(|| props.config.paths.database_path.clone());
-    let mut status_text = use_signal(|| "Idle".to_string());
-    let mut show_config_dialog = use_signal(|| false);
-    let mut config_changes = use_signal(|| Vec::<String>::new());
-    let speed_tracker = use_signal(|| SpeedTracker::new());
-    
-
-    let indexing_service_for_start = props.indexing_service.clone();
-    let indexing_service_for_start_dialog = props.indexing_service.clone();
-    let indexing_service_for_stop = props.indexing_service.clone();
-    let indexing_service_for_timer = props.indexing_service.clone();
-    let config_for_start = props.config.clone();
-    let config_for_dialog = props.config.clone();
-
-
-    // Automatic status updates every second
-    {
-        let mut status_text_clone = status_text.clone();
-        let mut speed_tracker_clone = speed_tracker.clone();
-        let service_clone = indexing_service_for_timer.clone();
-        use_future(move || {
-            let service = service_clone.clone();
-            async move {
-                loop {
-                    tokio::time::sleep(tokio::time::Duration::from_millis(50)).await;
-                    
-                    let status = service.get_status();
-                    let status_str = match status {
-                        IndexingStatus::Idle => {
-                            // Reset speed tracker when idle
-                            speed_tracker_clone.set(SpeedTracker::new());
-                            "Idle".to_string()
-                        },
-                        IndexingStatus::CountingFiles {
-                            current_file,
-                            start_time,
-                            ..
-                        } => {
-                            let elapsed = start_time.elapsed();
-                            let current_file_display = current_file
-                                .as_ref()
-                                .map(|f| format!("{}", f))
-                                .unwrap_or_else(|| "...".to_string());
-                            format!(
-                                "Phase 0 - Counting paths (shell) - {:.1}s elapsed\n{}",
-                                elapsed.as_secs_f64(),
-                                current_file_display
-                            )
-                        }
-                        IndexingStatus::RunningFileIndex { files_processed, total_files, current_file, start_time } => {
-                            // Add data point to speed tracker
-                            speed_tracker_clone.with_mut(|tracker| {
-                                tracker.add_data_point(files_processed);
-                            });
-                            
-                            let elapsed = start_time.elapsed();
-                            let current_file_display = current_file
-                                .as_ref()
-                                .map(|f| format!("Current: {}", f))
-                                .unwrap_or_default();
-                            
-                            // Calculate speed
-                            let speed_display = speed_tracker_clone.with(|tracker| {
-                                tracker.calculate_files_per_second()
-                                    .map(|fps| format!(" - {:.1} files/sec", fps))
-                                    .unwrap_or_default()
-                            });
-                            
-                            if let Some(total) = total_files {
-                                let percentage = if total > 0 { 
-                                    (files_processed as f64 / total as f64 * 100.0) as u32 
-                                } else { 0 };
-                                format!(
-                                    "Phase 1 - File Index: {}/{} files ({}%) - {:.1}s elapsed{}\n{}",
-                                    files_processed,
-                                    total,
-                                    percentage,
-                                    elapsed.as_secs_f64(),
-                                    speed_display,
-                                    current_file_display
-                                )
-                            } else {
-                                format!(
-                                    "Phase 1 - File Index: {} files processed - {:.1}s elapsed{}\n{}",
-                                    files_processed,
-                                    elapsed.as_secs_f64(),
-                                    speed_display,
-                                    current_file_display
-                                )
-                            }
-                        }
-                        IndexingStatus::RunningTextIndex { files_processed, current_file, start_time } => {
-                            // Add data point to speed tracker
-                            speed_tracker_clone.with_mut(|tracker| {
-                                tracker.add_data_point(files_processed);
-                            });
-                            
-                            let elapsed = start_time.elapsed();
-                            let current_file_display = current_file
-                                .as_ref()
-                                .map(|f| format!("Current: {}", f))
-                                .unwrap_or_default();
-                            
-                            // Calculate speed
-                            let speed_display = speed_tracker_clone.with(|tracker| {
-                                tracker.calculate_files_per_second()
-                                    .map(|fps| format!(" - {:.1} files/sec", fps))
-                                    .unwrap_or_default()
-                            });
-                            format!(
-                                "Phase 2 - Text Index: {} files processed - {:.1}s elapsed{}\n{}",
-                                files_processed,
-                                elapsed.as_secs_f64(),
-                                speed_display,
-                                current_file_display
-                            )
-                        }
-                        IndexingStatus::Stopping => "Indexing Stopped".to_string(),
-                        IndexingStatus::Error(ref e) => format!("Error: {}", e),
-                    };
-                    status_text_clone.set(status_str);
-                }
-            }
-        });
-    }
-
-    rsx! {
-        div { 
-            class: "app-container",
-            
-            div {
-                class: "app-header",
-                h1 { "QuickSearch File Indexer" }
-            }
-            
-            div {
-                class: "app-content",
-            
-            div { 
-                class: "section",
-                h2 { "Indexing Controls" }
-                
-                div {
-                    class: "form-group",
-                    label { "Paths to index (one per line):" }
-                    textarea {
-                        class: "form-control",
-                        rows: "3",
-                        value: "{indexing_path}",
-                        oninput: move |evt| indexing_path.set(evt.value())
-                    }
-                }
-                
-                div { 
-                    class: "form-group",
-                    label { "Database path:" }
-                    input { 
-                        class: "form-control",
-                        r#type: "text",
-                        value: "{db_path}",
-                        oninput: move |evt| db_path.set(evt.value())
-                    }
-                }
-                
-                div {
-                    class: "form-group",
-                    button { 
-                        class: "btn btn-primary",
-                        onclick: move |_| {
-                            let service = indexing_service_for_start.clone();
-                            let config = config_for_start.clone();
-                            let path_text = indexing_path().clone();
-                            let paths: Vec<String> = path_text
-                                .lines()
-                                .map(|l| l.trim().to_string())
-                                .filter(|l| !l.is_empty())
-                                .collect();
-                            let db = db_path().clone();
-                            
-                            if paths.is_empty() {
-                                status_text.set("Enter at least one path to index (one per line).".to_string());
-                            } else {
-                                // The config_validation key still stores a single joined string.
-                                let joined = paths.join("\n");
-                                match service.check_config_validation(&db, &config, &joined) {
-                                    Ok(Some(changes)) => {
-                                        config_changes.set(changes);
-                                        show_config_dialog.set(true);
-                                    }
-                                    Ok(None) => {
-                                        let _ = service.start_indexing(paths, db, config);
-                                    }
-                                    Err(e) => {
-                                        status_text.set(format!("Configuration validation error: {}", e));
-                                    }
-                                }
-                            }
-                        },
-                        "Start Indexing"
-                    }
-                    button { 
-                        class: "btn btn-danger",
-                        onclick: move |_| {
-                            let _ = indexing_service_for_stop.stop_indexing();
-                        },
-                        "Stop Indexing"
-                    }
-                }
-            }
-            
-            div {
-                class: "section",
-                h2 { "Status" }
-                pre { 
-                    class: "status-display",
-                    "{status_text}"
-                }
-            }
-            
-            crate::search::Search {
-                indexing_service: props.indexing_service.clone(),
-                db_path: db_path().clone()
-            }
-            
-            } // Close app-content
-        }
-        
-        // Configuration validation dialog
-        if show_config_dialog() {
-            div {
-                class: "modal-backdrop",
-                div {
-                    class: "modal-dialog",
-                    h3 { 
-                        style: "margin-top: 0; color: #d32f2f;",
-                        "⚠️ Configuration Changes Detected"
-                    }
-                    p { 
-                        style: "margin: 15px 0;",
-                        "The following configuration changes require deleting and rebuilding the search index:"
-                    }
-                    ul {
-                        style: "margin: 15px 0; padding-left: 20px;",
-                        for change in config_changes().iter() {
-                            li { 
-                                style: "margin: 5px 0; font-family: monospace; background-color: #f5f5f5; padding: 5px; border-radius: 3px;",
-                                "{change}"
-                            }
-                        }
-                    }
-                    p {
-                        style: "margin: 15px 0; font-weight: bold;",
-                        "This will delete the existing index and rebuild it from scratch."
-                    }
-                    div {
-                        style: "display: flex; gap: 10px; margin-top: 20px;",
-                        button {
-                            style: "padding: 10px 20px; background-color: #d32f2f; color: white; border: none; border-radius: 5px; cursor: pointer;",
-                            onclick: move |_| {
-                                let service = indexing_service_for_start_dialog.clone();
-                                let config = config_for_dialog.clone();
-                                let path_text = indexing_path().clone();
-                            let paths: Vec<String> = path_text
-                                .lines()
-                                .map(|l| l.trim().to_string())
-                                .filter(|l| !l.is_empty())
-                                .collect();
-                                let db = db_path().clone();
-                                
-                                show_config_dialog.set(false);
-                                status_text.set("Stopping indexing and deleting database...".to_string());
-                                
-                                // Delete database file and restart indexing
-                                let service_clone = service.clone();
-                                let paths_clone = paths.clone();
-                                let db_clone = db.clone();
-                                let config_clone = config.clone();
-                                let mut status_clone = status_text.clone();
-
-                                spawn(async move {
-                                    match service_clone.delete_index_for_rebuild(&db_clone) {
-                                        Ok(()) => {
-                                            status_clone.set("Database deleted. Starting fresh indexing...".to_string());
-                                            let _ = service_clone.start_indexing(paths_clone, db_clone, config_clone);
-                                        }
-                                        Err(e) => {
-                                            status_clone.set(format!("Error deleting database: {}", e));
-                                        }
-                                    }
-                                });
-                            },
-                            "Yes, Rebuild Index"
-                        }
-                        button {
-                            style: "padding: 10px 20px; background-color: #666; color: white; border: none; border-radius: 5px; cursor: pointer;",
-                            onclick: move |_| {
-                                show_config_dialog.set(false);
-                            },
-                            "Cancel"
-                        }
-                    }
-                }
-            }
-        }
-    }
-}

crates/quicksearch-gui/src/logs_tab.rs

@@ -0,0 +1,255 @@
+//! The Logs tab: what the terminal would have shown.
+//!
+//! Everything here comes from [`quicksearch_core::log`], which background
+//! threads write through instead of printing. Launched from a desktop
+//! launcher there is no terminal to read, and this is the only place a
+//! "cannot read that folder" warning is visible.
+
+use quicksearch_core::log::{self, Level, LogLine};
+
+use crate::format::group_thousands;
+
+/// How often to repaint while the tab is open. Log lines arrive on indexer
+/// and watcher threads, which have no reason to wake the UI, so an idle
+/// window would otherwise sit on a stale list until the mouse moved.
+const REFRESH_MS: u64 = 500;
+
+pub struct LogsTab {
+    /// Copy of the ring, refreshed only when the recorded count moves —
+    /// cloning a few thousand lines every frame would be silly.
+    lines: Vec<LogLine>,
+    /// [`log::recorded`] as of the last refresh.
+    seen: u64,
+    dropped: u64,
+    filter: String,
+    warnings_only: bool,
+    /// Keep the newest line in view. Scrolling up releases the view anyway
+    /// (egui unsticks a scroll area the user moves, and re-sticks it when
+    /// they return to the bottom); unticking this stops it following at all.
+    follow: bool,
+}
+
+impl LogsTab {
+    pub fn new() -> LogsTab {
+        LogsTab {
+            lines: Vec::new(),
+            seen: 0,
+            dropped: 0,
+            filter: String::new(),
+            warnings_only: false,
+            follow: true,
+        }
+    }
+
+    fn refresh(&mut self) {
+        self.lines = log::snapshot();
+        self.seen = log::recorded();
+        self.dropped = log::dropped();
+    }
+
+    pub fn ui(&mut self, ui: &mut egui::Ui) {
+        if log::recorded() != self.seen {
+            self.refresh();
+        }
+        ui.ctx()
+            .request_repaint_after(std::time::Duration::from_millis(REFRESH_MS));
+
+        // Indices rather than references: the control row below takes the
+        // filter and follow flags mutably, and a borrow of `self.lines`
+        // held across it would conflict. One frame of lag after a
+        // keystroke, which repaints immediately anyway.
+        let needle = self.filter.to_lowercase();
+        let shown: Vec<usize> = self
+            .lines
+            .iter()
+            .enumerate()
+            .filter(|(_, l)| keep(l, &needle, self.warnings_only))
+            .map(|(i, _)| i)
+            .collect();
+
+        let mut cleared = false;
+        ui.horizontal(|ui| {
+            ui.checkbox(&mut self.follow, "Follow")
+                .on_hover_text("Scroll to the newest line as it arrives");
+            ui.checkbox(&mut self.warnings_only, "Warnings only");
+            ui.add(
+                egui::TextEdit::singleline(&mut self.filter)
+                    .desired_width(200.0)
+                    .hint_text("Filter"),
+            );
+            if ui
+                .add_enabled(!shown.is_empty(), egui::Button::new("Copy"))
+                .on_hover_text("Copy the lines shown below to the clipboard")
+                .clicked()
+            {
+                let joined = shown
+                    .iter()
+                    .map(|&i| self.lines[i].text.as_str())
+                    .collect::<Vec<_>>()
+                    .join("\n");
+                ui.ctx().copy_text(joined);
+            }
+            if ui
+                .add_enabled(!self.lines.is_empty(), egui::Button::new("Clear"))
+                .clicked()
+            {
+                log::clear();
+                cleared = true;
+            }
+
+            ui.with_layout(egui::Layout::right_to_left(egui::Align::Center), |ui| {
+                let count = if shown.len() == self.lines.len() {
+                    format!("{} lines", group_thousands(self.lines.len() as u64))
+                } else {
+                    format!(
+                        "{} of {} lines",
+                        group_thousands(shown.len() as u64),
+                        group_thousands(self.lines.len() as u64)
+                    )
+                };
+                ui.label(egui::RichText::new(count).small().weak());
+            });
+        });
+        if cleared {
+            // `shown` indexes lines that no longer exist.
+            self.refresh();
+            return;
+        }
+        if self.dropped > 0 {
+            ui.label(
+                egui::RichText::new(format!(
+                    "{} earlier lines were dropped; the newest {} are kept.",
+                    group_thousands(self.dropped),
+                    group_thousands(log::CAPACITY as u64),
+                ))
+                .small()
+                .weak(),
+            );
+        }
+        ui.separator();
+
+        if self.lines.is_empty() {
+            ui.label(
+                egui::RichText::new(
+                    "Nothing logged yet. Warnings from indexing, watching folders and \
+                     opening files appear here — the same lines the terminal would show.",
+                )
+                .weak(),
+            );
+            return;
+        }
+        if shown.is_empty() {
+            ui.label(egui::RichText::new("No lines match the filter.").weak());
+            return;
+        }
+
+        // Long paths extend into a horizontal scroll rather than wrapping:
+        // `show_rows` only draws the visible slice, and that costs nothing
+        // only while every row is exactly one line tall.
+        ui.style_mut().wrap_mode = Some(egui::TextWrapMode::Extend);
+        let row_height = ui.text_style_height(&egui::TextStyle::Monospace);
+        egui::ScrollArea::both()
+            .auto_shrink([false; 2])
+            .stick_to_bottom(self.follow)
+            .show_rows(ui, row_height, shown.len(), |ui, range| {
+                for &i in &shown[range] {
+                    let line = &self.lines[i];
+                    ui.horizontal(|ui| {
+                        ui.label(
+                            egui::RichText::new(fmt_clock(line.at))
+                                .monospace()
+                                .weak(),
+                        );
+                        let text = egui::RichText::new(&line.text).monospace();
+                        match line.level {
+                            Level::Warn => {
+                                ui.colored_label(ui.visuals().warn_fg_color, text);
+                            }
+                            Level::Info => {
+                                ui.label(text);
+                            }
+                        }
+                    });
+                }
+            });
+    }
+}
+
+/// Whether a line survives the tab's two filters. `needle` is expected
+/// already lowercased — it is the same for every line, so folding it once
+/// per frame beats folding it per line.
+fn keep(line: &LogLine, needle: &str, warnings_only: bool) -> bool {
+    if warnings_only && line.level != Level::Warn {
+        return false;
+    }
+    needle.is_empty() || line.text.to_lowercase().contains(needle)
+}
+
+/// `HH:MM:SS` local time. The date is deliberately absent: these lines are
+/// read while something is going wrong now, and a full stamp on every row
+/// would crowd out the message.
+fn fmt_clock(unix_secs: u64) -> String {
+    use chrono::TimeZone;
+    // Saturating rather than `as`: that cast wraps a huge value into a
+    // negative one, which is a perfectly valid 1969 timestamp and would
+    // render as a plausible time instead of falling back.
+    let secs = i64::try_from(unix_secs).unwrap_or(i64::MAX);
+    match chrono::Local.timestamp_opt(secs, 0) {
+        chrono::LocalResult::Single(dt) => dt.format("%H:%M:%S").to_string(),
+        _ => "--:--:--".to_string(),
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::{fmt_clock, keep};
+    use quicksearch_core::log::{Level, LogLine};
+
+    fn line(level: Level, text: &str) -> LogLine {
+        LogLine {
+            at: 1_700_000_000,
+            level,
+            text: text.to_string(),
+        }
+    }
+
+    #[test]
+    fn an_empty_filter_keeps_everything() {
+        assert!(keep(&line(Level::Info, "anything"), "", false));
+        assert!(keep(&line(Level::Warn, "anything"), "", false));
+    }
+
+    #[test]
+    fn the_filter_ignores_case_on_both_sides() {
+        let l = line(Level::Warn, "Warning: cannot read /Home/Photos");
+        assert!(keep(&l, "photos", false), "needle case must not matter");
+        assert!(keep(&l, "cannot read", false), "nor the line's");
+        assert!(!keep(&l, "videos", false));
+    }
+
+    #[test]
+    fn warnings_only_hides_informational_lines() {
+        assert!(!keep(&line(Level::Info, "shutting down"), "", true));
+        assert!(keep(&line(Level::Warn, "cannot read"), "", true));
+    }
+
+    /// Both filters apply, not either.
+    #[test]
+    fn the_two_filters_compose() {
+        assert!(!keep(&line(Level::Info, "cannot read"), "cannot", true));
+        assert!(!keep(&line(Level::Warn, "cannot read"), "missing", true));
+        assert!(keep(&line(Level::Warn, "cannot read"), "cannot", true));
+    }
+
+    #[test]
+    fn a_clock_stamp_is_fixed_width() {
+        assert_eq!(fmt_clock(0).len(), 8, "epoch renders as a time, not a date");
+        assert_eq!(fmt_clock(1_700_000_000).len(), 8);
+    }
+
+    /// Beyond what a local calendar can represent, the row still lines up.
+    #[test]
+    fn an_out_of_range_stamp_falls_back() {
+        assert_eq!(fmt_clock(u64::MAX), "--:--:--");
+    }
+}

crates/quicksearch-gui/src/main.rs

@@ -1,54 +1,95 @@
-use std::sync::{Arc, OnceLock};
+//! QuickSearch binary: `quicksearch <query>` searches from the terminal;
-use dioxus::prelude::*;
+//! without a query it opens the egui desktop app.
-use quicksearch_core::{config, indexing, shutdown};
+//!
-mod frontend;
+//! On Windows this is the GUI only, built as a window-subsystem app so no
-mod search;
+//! console flashes behind it. Terminal search there is `quicksearch-cli`,
+//! which is a console app and so keeps working pipes, exit codes, and a shell
+//! that waits for it. A query passed here still does something useful: it
+//! seeds the search box.
+#![cfg_attr(windows, windows_subsystem = "windows")]
 
-static INDEXING_SERVICE: OnceLock<Arc<indexing::IndexingService>> = OnceLock::new();
+mod app;
+mod backend;
+#[cfg(not(windows))]
+mod cli;
+mod duplicates_tab;
+mod format;
+mod logs_tab;
+mod manage_tab;
+mod options;
+mod platform;
+mod query_highlight;
+mod search_tab;
+mod tracker;
+
+use quicksearch_core::config::Config;
+
+/// The window icon, shown in the titlebar, taskbar and alt-tab switcher.
+///
+/// X11 takes these pixels directly via `_NET_WM_ICON`. Wayland ignores them and
+/// instead looks up the app id in `/usr/share/applications/`, so the id below has
+/// to match the installed `quicksearch.desktop` for the icon to appear there.
+fn app_icon() -> egui::IconData {
+    eframe::icon_data::from_png_bytes(include_bytes!("../assets/icons/quicksearch-256.png"))
+        .expect("bundled icon is a valid PNG")
+}
+
+/// Leftover positional arguments, joined — used to seed the search box.
+///
+/// Flags are dropped rather than parsed: eframe and winit take some of their
+/// own, and a stray `--foo` should not end up in the query.
+fn seed_query() -> Option<String> {
+    let terms: Vec<String> = std::env::args()
+        .skip(1)
+        .filter(|a| !a.starts_with('-'))
+        .collect();
+    if terms.is_empty() {
+        None
+    } else {
+        Some(terms.join(" "))
+    }
+}
 
 fn main() {
-    let indexing_service = Arc::new(indexing::IndexingService::new());
+    // Must come first: anything below may print, and printing without a
-    INDEXING_SERVICE
+    // stdio handle panics rather than failing quietly.
-        .set(indexing_service.clone())
+    #[cfg(windows)]
-        .expect("Failed to set global indexing service");
+    platform::redirect_null_stdio();
 
-    if let Err(e) = shutdown::install_signal_handler(indexing_service.clone()) {
+    #[cfg(not(windows))]
-        eprintln!("Warning: failed to install signal handler: {}", e);
+    if let Some(code) = cli::maybe_run_cli() {
+        std::process::exit(code);
     }
 
-    LaunchBuilder::desktop()
+    // A broken config file should never keep the window from opening —
-        .with_cfg(
+    // surface the error in-app and run on defaults.
-            dioxus_desktop::Config::new()
+    let (config, config_error) = match Config::load() {
-                .with_custom_head(format!("<style>{}</style>", include_str!("../assets/styles.css")))
+        Ok(c) => (c, None),
-                .with_window(dioxus_desktop::WindowBuilder::new()
+        Err(e) => (Config::default(), Some(e)),
-                    .with_title("QuickSearch - File Indexer & Search")
+    };
-                    .with_resizable(true)
+    let initial_query = seed_query();
-                    .with_inner_size(dioxus_desktop::LogicalSize::new(1000.0, 700.0))
-                )
-        )
-        .launch(app);
-}
 
-
+    let native_options = eframe::NativeOptions {
-fn app() -> Element {
+        viewport: egui::ViewportBuilder::default()
-    let cfg = match config::Config::load() {
+            .with_title("QuickSearch")
-        Ok(c) => c,
+            .with_app_id("quicksearch")
-        Err(e) => {
+            .with_icon(app_icon())
-            eprintln!("Failed to load config: {}", e);
+            .with_inner_size([1000.0, 700.0])
-            return rsx! { div { "Failed to load configuration" } };
+            .with_min_inner_size([640.0, 400.0]),
-        }
+        ..Default::default()
     };
 
-    let indexing_service = INDEXING_SERVICE
+    let result = eframe::run_native(
-        .get()
+        "QuickSearch",
-        .expect("Indexing service not initialized")
+        native_options,
-        .clone();
+        Box::new(move |cc| {
-
+            app::QuickSearchApp::new(cc, config, config_error, initial_query)
-    rsx! {
+                .map(|app| Box::new(app) as Box<dyn eframe::App>)
-        frontend::App {
+                .map_err(|e| e.into())
-            indexing_service: indexing_service,
+        }),
-            config: cfg
+    );
-        }
+    if let Err(e) = result {
+        eprintln!("failed to start GUI: {}", e);
+        std::process::exit(1);
     }
 }
-

crates/quicksearch-gui/src/manage_tab.rs

@@ -0,0 +1,447 @@
+//! The Manage Index tab: detailed status, mode controls, indexed roots,
+//! and the content/ignore filter editors.
+
+use quicksearch_core::config::Config;
+use quicksearch_core::coordinator::{IndexMode, IndexerState, WatcherStatus};
+use quicksearch_core::indexing::{IndexingStatus, RootPhase, RootProgress};
+
+use crate::format::{fmt_interval, fmt_rate, group_thousands, middle_truncate};
+use crate::options::{config_editor_ui, Section};
+use crate::tracker::SpeedTracker;
+
+/// What the tab asks the app to do after this frame.
+#[derive(Default)]
+pub struct ManageActions {
+    pub start_now: bool,
+    pub stop: bool,
+    pub auto: bool,
+    /// Ask the app to confirm and delete the index.
+    pub clear_index: bool,
+    /// A full edited config to apply (roots / filters / indexing knobs).
+    pub apply_config: Option<Config>,
+}
+
+pub struct ManageTab {
+    pub speed: SpeedTracker,
+    /// Multiline editors, one entry per line; synced from config on tab
+    /// entry and parsed back on Apply.
+    ext_filter_text: String,
+    ignore_filter_text: String,
+    new_root: String,
+    /// Inline error from a rejected root add (nested/duplicate).
+    root_error: Option<String>,
+    editors_synced: bool,
+    /// Draft of the indexing/processing knobs edited in-place.
+    draft: Option<Config>,
+}
+
+impl ManageTab {
+    pub fn new() -> ManageTab {
+        ManageTab {
+            speed: SpeedTracker::new(),
+            ext_filter_text: String::new(),
+            ignore_filter_text: String::new(),
+            new_root: String::new(),
+            root_error: None,
+            editors_synced: false,
+            draft: None,
+        }
+    }
+
+    /// Feed the tracker from the polled status (called every frame, on
+    /// every tab, so the status bar rate stays live).
+    pub fn observe(&mut self, status: &IndexingStatus) {
+        match status {
+            IndexingStatus::Running { roots, .. } => {
+                // Monotonic within a run: walks and extractions only grow.
+                let total: usize = roots.iter().map(|r| r.walked + r.extracted).sum();
+                self.speed.record(total);
+            }
+            IndexingStatus::Idle | IndexingStatus::Error(_) => self.speed.reset(),
+            _ => {}
+        }
+    }
+
+    fn sync_editors(&mut self, config: &Config) {
+        if !self.editors_synced {
+            self.ext_filter_text = config.indexing.content_extensions.join("\n");
+            self.ignore_filter_text = config.indexing.ignore_patterns.join("\n");
+            self.draft = Some(config.clone());
+            self.editors_synced = true;
+        }
+    }
+
+    /// Force a re-sync next frame (config changed elsewhere).
+    pub fn invalidate_editors(&mut self) {
+        self.editors_synced = false;
+    }
+
+    pub fn ui(
+        &mut self,
+        ui: &mut egui::Ui,
+        state: &IndexerState,
+        config: &Config,
+    ) -> ManageActions {
+        let mut actions = ManageActions::default();
+        self.sync_editors(config);
+
+        egui::ScrollArea::vertical().auto_shrink([false; 2]).show(ui, |ui| {
+            // --- Status ---------------------------------------------------
+            ui.heading("Status");
+            status_panel(ui, state, &self.speed);
+            watch_panel(ui, state, config);
+            ui.add_space(8.0);
+
+            // --- Controls -------------------------------------------------
+            ui.horizontal(|ui| {
+                let running = !matches!(
+                    state.activity,
+                    IndexingStatus::Idle | IndexingStatus::Error(_)
+                );
+                if ui.add_enabled(!running, egui::Button::new("Start indexing now")).clicked() {
+                    actions.start_now = true;
+                }
+                if ui.add_enabled(running || state.mode == IndexMode::Auto, egui::Button::new("Stop")).clicked() {
+                    actions.stop = true;
+                }
+                if ui
+                    .add_enabled(state.mode != IndexMode::Auto, egui::Button::new("Return to Automatic"))
+                    .clicked()
+                {
+                    actions.auto = true;
+                }
+                let mode = match state.mode {
+                    IndexMode::Auto => "Automatic",
+                    IndexMode::ManualStopped => "Manual (stopped)",
+                    IndexMode::ManualRunning => "Manual (running)",
+                };
+                ui.label(egui::RichText::new(format!("Mode: {}", mode)).weak());
+                ui.separator();
+                if ui
+                    .button(egui::RichText::new("Clear index…").color(ui.visuals().error_fg_color))
+                    .on_hover_text("Delete the index database (asks for confirmation)")
+                    .clicked()
+                {
+                    actions.clear_index = true;
+                }
+                if state.queued_events > 0 {
+                    ui.label(
+                        egui::RichText::new(format!("{} changes queued", state.queued_events))
+                            .small()
+                            .weak(),
+                    );
+                }
+            });
+            ui.separator();
+
+            // --- Indexed roots ---------------------------------------------
+            ui.heading("Indexed folders");
+            let draft = self.draft.as_mut().expect("synced");
+            let mut remove: Option<usize> = None;
+            for (i, root) in draft.paths.indexing_paths.clone().iter().enumerate() {
+                ui.horizontal(|ui| {
+                    // Controls claim the right edge first so a long path can
+                    // never push them out of view; the path truncates into
+                    // whatever width remains (full path on hover).
+                    ui.with_layout(egui::Layout::right_to_left(egui::Align::Center), |ui| {
+                        if ui.small_button("Remove").clicked() {
+                            remove = Some(i);
+                        }
+                        // Per-root walker override; 0 = auto (4 local / 16
+                        // network, detected per root). Applies on the next run.
+                        let mut workers =
+                            draft.indexing.root_workers.get(root).copied().unwrap_or(0);
+                        let response = ui
+                            .add(
+                                egui::DragValue::new(&mut workers)
+                                    .range(0..=64)
+                                    .custom_formatter(|n, _| {
+                                        if n == 0.0 {
+                                            "auto".to_string()
+                                        } else {
+                                            format!("{:.0}", n)
+                                        }
+                                    })
+                                    .custom_parser(|s| {
+                                        let s = s.trim();
+                                        if s.is_empty() || s.eq_ignore_ascii_case("auto") {
+                                            Some(0.0)
+                                        } else {
+                                            s.parse().ok()
+                                        }
+                                    }),
+                            )
+                            .on_hover_text(
+                                "Walker threads for this folder. auto = 4 on local \
+                                 storage, 16 on network mounts. Takes effect on \
+                                 the next indexing run.",
+                            );
+                        if response.changed() {
+                            if workers == 0 {
+                                draft.indexing.root_workers.remove(root);
+                            } else {
+                                draft.indexing.root_workers.insert(root.clone(), workers);
+                            }
+                        }
+                        ui.label(egui::RichText::new("workers:").small().weak());
+
+                        // Path label takes the leftover width, middle-truncated.
+                        ui.with_layout(
+                            egui::Layout::left_to_right(egui::Align::Center),
+                            |ui| {
+                                let font_id = egui::TextStyle::Monospace.resolve(ui.style());
+                                let char_width =
+                                    ui.fonts(|f| f.glyph_width(&font_id, '0')).max(1.0);
+                                let budget =
+                                    ((ui.available_width() / char_width) as usize).max(16);
+                                ui.monospace(middle_truncate(root, budget))
+                                    .on_hover_text(root);
+                            },
+                        );
+                    });
+                });
+            }
+            if let Some(i) = remove {
+                let removed = draft.paths.indexing_paths.remove(i);
+                draft.indexing.root_workers.remove(&removed);
+            }
+            ui.horizontal(|ui| {
+                if ui.button("Add folder…").clicked() {
+                    if let Some(dir) = rfd::FileDialog::new().pick_folder() {
+                        let path = dir.to_string_lossy().into_owned();
+                        try_add_root(draft, path, &mut self.root_error);
+                    }
+                }
+                ui.add(
+                    egui::TextEdit::singleline(&mut self.new_root)
+                        .desired_width(240.0)
+                        .hint_text("or type a path"),
+                );
+                if ui.button("Add").clicked() && !self.new_root.trim().is_empty() {
+                    let path = self.new_root.trim().to_string();
+                    if try_add_root(draft, path, &mut self.root_error) {
+                        self.new_root.clear();
+                    }
+                }
+            });
+            if let Some(err) = &self.root_error {
+                ui.colored_label(ui.visuals().error_fg_color, err);
+            }
+            ui.separator();
+
+            // --- Filters ---------------------------------------------------
+            ui.heading("Content filters");
+            ui.columns(2, |cols| {
+                cols[0].label("Full-text extensions (empty = all supported):");
+                cols[0].add(
+                    egui::TextEdit::multiline(&mut self.ext_filter_text)
+                        .desired_rows(4)
+                        .desired_width(f32::INFINITY)
+                        .hint_text("txt\nmd\npdf"),
+                );
+                cols[1].label("Ignore patterns (excluded entirely):");
+                cols[1].add(
+                    egui::TextEdit::multiline(&mut self.ignore_filter_text)
+                        .desired_rows(4)
+                        .desired_width(f32::INFINITY)
+                        .hint_text(".git\nnode_modules\n*.tmp"),
+                );
+            });
+            ui.separator();
+
+            // --- Indexing options -------------------------------------------
+            ui.heading("Indexing options");
+            config_editor_ui(ui, draft, Section::Indexing);
+            ui.add_space(4.0);
+            config_editor_ui(ui, draft, Section::Processing);
+            ui.add_space(8.0);
+
+            if ui.button("Apply & Save").clicked() {
+                let mut new_config = draft.clone();
+                new_config.indexing.content_extensions = parse_lines(&self.ext_filter_text);
+                new_config.indexing.ignore_patterns = parse_lines(&self.ignore_filter_text);
+                let roots = new_config.paths.indexing_paths.clone();
+                new_config
+                    .indexing
+                    .root_workers
+                    .retain(|root, _| roots.contains(root));
+                actions.apply_config = Some(new_config);
+                self.editors_synced = false;
+            }
+        });
+
+        actions
+    }
+}
+
+/// Append a root to the draft unless it would duplicate or nest with an
+/// existing one; the rejection reason lands in `error`.
+fn try_add_root(draft: &mut Config, candidate: String, error: &mut Option<String>) -> bool {
+    if draft.paths.indexing_paths.contains(&candidate) {
+        *error = Some(format!("{} is already in the list", candidate));
+        return false;
+    }
+    let mut probe = draft.paths.indexing_paths.clone();
+    probe.push(candidate.clone());
+    if let Some((child, parent)) = quicksearch_core::config::nested_roots(&probe).first() {
+        *error = Some(format!(
+            "Not added: {} is nested under {}; indexed folders may not overlap",
+            child, parent
+        ));
+        return false;
+    }
+    draft.paths.indexing_paths.push(candidate);
+    *error = None;
+    true
+}
+
+fn parse_lines(text: &str) -> Vec<String> {
+    text.lines()
+        .map(str::trim)
+        .filter(|l| !l.is_empty())
+        .map(str::to_string)
+        .collect()
+}
+
+/// Live-update health. Permanent counterpart to the one-time modal: the
+/// modal is dismissed and remembered per root, but "live updates are off"
+/// stays true and must remain discoverable.
+fn watch_panel(ui: &mut egui::Ui, state: &IndexerState, config: &Config) {
+    match &state.watcher {
+        // Manual mode already says "stopped" in the controls row; repeating
+        // it here would be noise.
+        WatcherStatus::Off => {}
+        WatcherStatus::Starting => {
+            ui.label(
+                egui::RichText::new("Setting up live updates…")
+                    .small()
+                    .weak(),
+            );
+        }
+        WatcherStatus::Active { dirs } => {
+            ui.label(
+                egui::RichText::new(format!(
+                    "Live updates on, watching {} folders",
+                    group_thousands(*dirs as u64)
+                ))
+                .small()
+                .weak(),
+            );
+        }
+        WatcherStatus::Disabled { reason } => {
+            ui.colored_label(
+                ui.visuals().warn_fg_color,
+                format!(
+                    "⚠ Live updates off; reindexing every {}",
+                    fmt_interval(config.indexing.reindex_interval_minutes)
+                ),
+            )
+            .on_hover_text(reason.to_string());
+        }
+    }
+}
+
+fn status_panel(ui: &mut egui::Ui, state: &IndexerState, speed: &SpeedTracker) {
+    match &state.activity {
+        IndexingStatus::Idle => {
+            // Relative wording makes even a milliseconds-fast run visibly
+            // register ("just now") instead of looking like a dead button.
+            let last = state
+                .last_full_index
+                .map(crate::format::fmt_ago)
+                .unwrap_or_else(|| "never".to_string());
+            ui.label(format!("Idle; last full index: {}", last));
+        }
+        IndexingStatus::Error(e) => {
+            ui.colored_label(ui.visuals().error_fg_color, format!("Error: {}", e));
+        }
+        IndexingStatus::Stopping => {
+            ui.label("Stopping…");
+        }
+        IndexingStatus::Running { roots, .. } => {
+            for root in roots {
+                root_row(ui, root);
+            }
+            if let Some(rate) = speed.files_per_sec() {
+                ui.label(
+                    egui::RichText::new(format!("overall: {}", fmt_rate(rate)))
+                        .small()
+                        .weak(),
+                );
+            }
+        }
+    }
+}
+
+/// One root's progress: path, phase, bar, counters, current file.
+fn root_row(ui: &mut egui::Ui, r: &RootProgress) {
+    // Weak "|" separators split the row into folder | status | numbers.
+    let divider = |ui: &mut egui::Ui| {
+        ui.label(egui::RichText::new("|").weak());
+    };
+    ui.horizontal(|ui| {
+        ui.monospace(middle_truncate(&r.root, 48));
+        divider(ui);
+        match r.phase {
+            RootPhase::Walking => {
+                ui.label("indexing");
+                divider(ui);
+                let workers = format!("{}/{} workers", r.active_workers, r.total_workers);
+                match r.walk_total {
+                    Some(total) if total > 0 => {
+                        let frac = (r.walked as f32 / total as f32).clamp(0.0, 1.0);
+                        ui.label(format!(
+                            "{} / {} ({:.0}%) · {}",
+                            group_thousands(r.walked as u64),
+                            group_thousands(total as u64),
+                            frac * 100.0,
+                            workers
+                        ));
+                        ui.add(egui::ProgressBar::new(frac).desired_width(160.0));
+                    }
+                    _ => {
+                        ui.label(format!(
+                            "{} files · {}",
+                            group_thousands(r.walked as u64),
+                            workers
+                        ));
+                        ui.add(egui::ProgressBar::new(0.0).animate(true).desired_width(160.0));
+                    }
+                }
+            }
+            RootPhase::Extracting => {
+                ui.label("extracting text for search");
+                divider(ui);
+                let frac = if r.extract_total > 0 {
+                    (r.extracted as f32 / r.extract_total as f32).clamp(0.0, 1.0)
+                } else {
+                    1.0
+                };
+                ui.label(format!(
+                    "{} / {} ({:.0}%)",
+                    group_thousands(r.extracted as u64),
+                    group_thousands(r.extract_total as u64),
+                    frac * 100.0
+                ));
+                ui.add(egui::ProgressBar::new(frac).desired_width(160.0));
+            }
+            RootPhase::Done => {
+                // Whole-root totals: `walked` covers every file the walk
+                // saw (including unchanged, skipped ones) and `extracted`
+                // covers all rows with searchable text, not just this
+                // run's new work.
+                ui.label("done");
+                divider(ui);
+                ui.label(format!(
+                    "indexed {}, extracted {}",
+                    group_thousands(r.walked as u64),
+                    group_thousands(r.extracted as u64)
+                ));
+                ui.add(egui::ProgressBar::new(1.0).desired_width(160.0));
+            }
+        }
+    });
+    if let Some(f) = &r.current_file {
+        ui.label(egui::RichText::new(middle_truncate(f, 90)).small().weak());
+    }
+}

crates/quicksearch-gui/src/options.rs

@@ -0,0 +1,251 @@
+//! The Options window and the shared config editor used by both the
+//! window and the Manage Index tab. Edits happen on a draft; Apply
+//! validates, saves, and hands the new config to the app.
+
+use quicksearch_core::config::Config;
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum Section {
+    Indexing,
+    Processing,
+    Search,
+}
+
+pub struct OptionsWindow {
+    pub open: bool,
+    draft: Option<Config>,
+}
+
+impl OptionsWindow {
+    pub fn new() -> OptionsWindow {
+        OptionsWindow {
+            open: false,
+            draft: None,
+        }
+    }
+
+    pub fn open_with(&mut self, current: &Config) {
+        self.open = true;
+        self.draft = Some(current.clone());
+    }
+
+    /// Render; returns a new config when the user applied changes.
+    pub fn ui(&mut self, ctx: &egui::Context, current: &Config) -> Option<Config> {
+        if !self.open {
+            self.draft = None;
+            return None;
+        }
+        if self.draft.is_none() {
+            self.draft = Some(current.clone());
+        }
+        let mut applied = None;
+        let mut open = self.open;
+        let draft = self.draft.as_mut().unwrap();
+
+        egui::Window::new("Options")
+            .open(&mut open)
+            .resizable(false)
+            .default_width(420.0)
+            .show(ctx, |ui| {
+                egui::ScrollArea::vertical().max_height(480.0).show(ui, |ui| {
+                    ui.heading("Paths");
+                    egui::Grid::new("opt-paths").num_columns(2).show(ui, |ui| {
+                        ui.label("Database file");
+                        ui.add(
+                            egui::TextEdit::singleline(&mut draft.paths.database_path)
+                                .desired_width(260.0),
+                        );
+                        ui.end_row();
+                    });
+                    ui.label(
+                        egui::RichText::new(
+                            "Indexed folders are managed on the Manage Index tab.",
+                        )
+                        .small()
+                        .weak(),
+                    );
+                    ui.separator();
+
+                    ui.heading("Indexing");
+                    config_editor_ui(ui, draft, Section::Indexing);
+                    ui.separator();
+
+                    ui.heading("Processing");
+                    config_editor_ui(ui, draft, Section::Processing);
+                    ui.separator();
+
+                    ui.heading("Search");
+                    config_editor_ui(ui, draft, Section::Search);
+                    ui.separator();
+
+                    ui.heading("Interface");
+                    egui::Grid::new("opt-ui").num_columns(2).show(ui, |ui| {
+                        ui.label("UI scale");
+                        ui.add(
+                            egui::Slider::new(&mut draft.ui.scale, 0.5..=2.5)
+                                .step_by(0.05)
+                                .fixed_decimals(2),
+                        )
+                        .on_hover_text(
+                            "Zooms the whole interface: fonts, spacing, and \
+                             widgets. Ctrl +/- and Ctrl 0 adjust it temporarily \
+                             at runtime.",
+                        );
+                        ui.end_row();
+                    });
+                });
+
+                ui.separator();
+                ui.horizontal(|ui| {
+                    if ui.button("Apply & Save").clicked() {
+                        applied = Some(draft.clone());
+                    }
+                    ui.label(
+                        egui::RichText::new(
+                            "Changes to tokenizer, filters, hidden files, or hashing \
+                             prompt an index rebuild.",
+                        )
+                        .small()
+                        .weak(),
+                    );
+                });
+            });
+
+        self.open = open;
+        if !self.open {
+            self.draft = None;
+        }
+        applied
+    }
+}
+
+/// One implementation of the per-section config controls, shared by the
+/// Options window and the Manage tab.
+pub fn config_editor_ui(ui: &mut egui::Ui, config: &mut Config, section: Section) {
+    match section {
+        Section::Indexing => {
+            egui::Grid::new("cfg-indexing").num_columns(2).show(ui, |ui| {
+                ui.label("Automatic indexing");
+                ui.checkbox(&mut config.indexing.auto_index, "watchers + periodic reindex");
+                ui.end_row();
+
+                ui.label("Full reindex every");
+                ui.horizontal(|ui| {
+                    ui.add(
+                        egui::DragValue::new(&mut config.indexing.reindex_interval_minutes)
+                            .range(5..=60 * 24 * 30),
+                    );
+                    ui.label("minutes");
+                });
+                ui.end_row();
+
+                ui.label("Follow symlinks");
+                ui.checkbox(&mut config.indexing.follow_symlinks, "");
+                ui.end_row();
+
+                ui.label("Include hidden files");
+                ui.checkbox(&mut config.indexing.include_hidden, "");
+                ui.end_row();
+            });
+        }
+        Section::Processing => {
+            egui::Grid::new("cfg-processing").num_columns(2).show(ui, |ui| {
+                ui.label("Tokenizer");
+                egui::ComboBox::from_id_salt("cfg-tokenize")
+                    .selected_text(&config.processing.tokenize)
+                    .show_ui(ui, |ui| {
+                        for opt in ["trigram", "unicode61", "porter"] {
+                            ui.selectable_value(
+                                &mut config.processing.tokenize,
+                                opt.to_string(),
+                                opt,
+                            );
+                        }
+                    });
+                ui.end_row();
+
+                ui.label("Hash sample size (bytes)");
+                ui.add(egui::DragValue::new(&mut config.processing.hash_length).range(512..=1_048_576));
+                ui.end_row();
+
+                ui.label("Max stored text (bytes)");
+                ui.add(
+                    egui::DragValue::new(&mut config.processing.maximum_text_size)
+                        .range(1024..=16_777_216),
+                );
+                ui.end_row();
+
+                ui.label("Max text file size (bytes)");
+                ui.add(
+                    egui::DragValue::new(&mut config.processing.maximum_text_file_size)
+                        .range(1024..=1_073_741_824),
+                );
+                ui.end_row();
+
+                ui.label("Batch size");
+                ui.add(egui::DragValue::new(&mut config.processing.batch_size).range(10..=100_000));
+                ui.end_row();
+
+                ui.label("Store text for snippets");
+                ui.checkbox(&mut config.processing.store_text_for_snippets, "")
+                    .on_hover_text(
+                        "Off: smaller index, but no previews, occurrence ranking, \
+                         case verification, or fuzzy full-text search",
+                    );
+                ui.end_row();
+            });
+        }
+        Section::Search => {
+            egui::Grid::new("cfg-search").num_columns(2).show(ui, |ui| {
+                ui.label("Fuzzy stages on by default");
+                ui.checkbox(&mut config.search.fuzzy_default, "");
+                ui.end_row();
+
+                ui.label("Fuzzy edit distance");
+                ui.vertical(|ui| {
+                    ui.add(egui::DragValue::new(&mut config.search.fuzzy_max_edits).range(0..=8))
+                        .on_hover_text(
+                            "Ceiling on the typo budget. The allowance grows with the \
+                             search term, one edit per three characters, up to this \
+                             value, so 2 means \"1 edit for short terms, 2 for longer \
+                             ones\". 0 turns the fuzzy stages off.",
+                        );
+                    if let Some(warning) = config.search.fuzzy_edits_warning() {
+                        ui.label(
+                            egui::RichText::new(warning)
+                                .small()
+                                .color(egui::Color32::from_rgb(220, 150, 40)),
+                        );
+                    }
+                });
+                ui.end_row();
+
+                ui.label("Display limit");
+                ui.add(egui::DragValue::new(&mut config.search.display_limit).range(50..=100_000));
+                ui.end_row();
+
+                ui.label("Stream batch size");
+                ui.add(
+                    egui::DragValue::new(&mut config.search.results_per_page).range(10..=10_000),
+                );
+                ui.end_row();
+
+                ui.label("Debounce (ms)");
+                ui.add(egui::DragValue::new(&mut config.search.debounce_ms).range(0..=2000));
+                ui.end_row();
+
+                ui.label("Fuzzy max edits");
+                ui.add(egui::DragValue::new(&mut config.search.fuzzy_max_edits).range(0..=8))
+                    .on_hover_text(
+                        "Ceiling on fuzzy edit distance (the budget grows one \
+                         edit per three characters of the term). 0 disables \
+                         the fuzzy passes.",
+                    );
+                ui.end_row();
+            });
+            if let Some(warning) = config.search.fuzzy_edits_warning() {
+                ui.colored_label(ui.visuals().warn_fg_color, warning);
+            }
+        }
+    }
+}

crates/quicksearch-gui/src/platform.rs

@@ -0,0 +1,134 @@
+//! Opening files, revealing them in the system file manager, and the one bit
+//! of process setup that has to happen before anything prints.
+
+use std::process::Command;
+
+/// Give the process somewhere to write when it has no stdio.
+///
+/// A window-subsystem binary launched from Explorer has NULL standard handles,
+/// and `println!`/`eprintln!` *panic* when the write fails rather than
+/// dropping the output. Pointing the handles at `NUL` makes those writes
+/// succeed and go nowhere.
+///
+/// Background reporting no longer depends on this — it goes through
+/// [`quicksearch_core::log`], which ignores a failed stderr write and keeps
+/// the line for the Logs tab — but the remaining direct prints (a startup
+/// failure, a panic message) still reach a handle that accepts them.
+///
+/// Handles inherited from a real console are left alone, so running the binary
+/// from a shell still prints normally.
+#[cfg(windows)]
+pub fn redirect_null_stdio() {
+    use std::os::windows::io::IntoRawHandle;
+    use windows_sys::Win32::Foundation::INVALID_HANDLE_VALUE;
+    use windows_sys::Win32::System::Console::{
+        GetStdHandle, SetStdHandle, STD_ERROR_HANDLE, STD_OUTPUT_HANDLE,
+    };
+
+    for id in [STD_OUTPUT_HANDLE, STD_ERROR_HANDLE] {
+        let existing = unsafe { GetStdHandle(id) };
+        if !existing.is_null() && existing != INVALID_HANDLE_VALUE {
+            continue;
+        }
+        if let Ok(file) = std::fs::OpenOptions::new().write(true).open("NUL") {
+            // Deliberately leaked: the handle has to outlive every later
+            // write, which means the whole process.
+            unsafe { SetStdHandle(id, file.into_raw_handle() as _) };
+        }
+    }
+}
+
+/// Open a file with the desktop's default application, detached.
+pub fn open_file(path: &str) {
+    if let Err(e) = open::that_detached(path) {
+        quicksearch_core::log_warn!("open {}: {}", path, e);
+    }
+}
+
+/// Reveal a file in the system file manager with the file selected.
+///
+/// Linux: `org.freedesktop.FileManager1.ShowItems` over the session bus
+/// (supported by every mainstream file manager) via `dbus-send` — no
+/// D-Bus library dependency for one call. Falls back to opening the
+/// parent directory. Windows/macOS use their native select verbs.
+pub fn reveal_in_folder(path: &str) {
+    #[cfg(target_os = "windows")]
+    {
+        use std::os::windows::process::CommandExt;
+        /// Keep a console window from flashing behind the spawn.
+        const CREATE_NO_WINDOW: u32 = 0x0800_0000;
+
+        // explorer.exe parses its own command line rather than using the
+        // standard argv splitting, and wants `/select,` glued to the path as a
+        // single token with quotes around the path only. Passed as two
+        // arguments it ignores the selection and just opens the folder, and
+        // std's quoting would wrap the whole token. `raw_arg` is the only way
+        // to say exactly this.
+        //
+        // Forward slashes are valid everywhere else on Windows but not here,
+        // so normalize first. The exit code is not worth checking: explorer
+        // returns 1 even on success.
+        let native = path.replace('/', "\\");
+        let _ = Command::new("explorer.exe")
+            .raw_arg(format!("/select,\"{}\"", native))
+            .creation_flags(CREATE_NO_WINDOW)
+            .spawn();
+        return;
+    }
+    #[cfg(target_os = "macos")]
+    {
+        let _ = Command::new("open").arg("-R").arg(path).spawn();
+        return;
+    }
+    #[cfg(all(unix, not(target_os = "macos")))]
+    {
+        use std::path::Path;
+        let uri = format!("file://{}", uri_escape_path(path));
+        let ok = Command::new("dbus-send")
+            .args([
+                "--session",
+                "--print-reply",
+                "--dest=org.freedesktop.FileManager1",
+                "/org/freedesktop/FileManager1",
+                "org.freedesktop.FileManager1.ShowItems",
+                &format!("array:string:{}", uri),
+                "string:",
+            ])
+            .output()
+            .map(|out| out.status.success())
+            .unwrap_or(false);
+        if !ok {
+            let parent = Path::new(path).parent().unwrap_or(Path::new("/"));
+            let _ = Command::new("xdg-open").arg(parent).spawn();
+        }
+    }
+}
+
+/// Percent-encode a filesystem path for a file:// URI, keeping `/`.
+#[cfg(all(unix, not(target_os = "macos")))]
+fn uri_escape_path(path: &str) -> String {
+    let mut out = String::with_capacity(path.len());
+    for byte in path.bytes() {
+        match byte {
+            b'A'..=b'Z' | b'a'..=b'z' | b'0'..=b'9' | b'/' | b'-' | b'_' | b'.' | b'~' => {
+                out.push(byte as char)
+            }
+            _ => out.push_str(&format!("%{:02X}", byte)),
+        }
+    }
+    out
+}
+
+#[cfg(test)]
+mod tests {
+    #[cfg(all(unix, not(target_os = "macos")))]
+    #[test]
+    fn uri_escaping() {
+        use super::uri_escape_path;
+        assert_eq!(uri_escape_path("/plain/path.txt"), "/plain/path.txt");
+        assert_eq!(
+            uri_escape_path("/with space/ünïcode&.txt"),
+            "/with%20space/%C3%BCn%C3%AFcode%26.txt"
+        );
+    }
+}

crates/quicksearch-gui/src/query_highlight.rs

@@ -0,0 +1,739 @@
+//! Syntax highlighting for the search box.
+//!
+//! [`classify`] is a pure token walk over [`tokenize_spanned`] output that
+//! mirrors `split_for_cascade` branch for branch — it must never claim
+//! something is a filter (or a wildcard) that the engine treats as plain
+//! text. The egui layer at the bottom turns its segments into a `Galley`
+//! for `TextEdit::layouter`.
+//!
+//! Color scheme: recognized keywords red, their arguments blue, syntax
+//! characters (operators, quotes, live wildcards) green, invalid arguments
+//! in the error color, everything else plain. A complete recognized filter
+//! additionally gets a tinted background chip.
+
+use std::ops::Range;
+use std::sync::Arc;
+
+use egui::text::{LayoutJob, TextFormat};
+use egui::{Color32, Galley, Stroke};
+use quicksearch_core::query::ast::Op;
+use quicksearch_core::query::lexer::{tokenize_spanned, Token};
+use quicksearch_core::query::pattern::RegexQuery;
+use quicksearch_core::query::translator::{build_filter, is_filter_key};
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum Class {
+    Plain,
+    /// The key word of a recognized filter (`type`, `name`, `regex`, …).
+    Keyword,
+    /// Syntax characters doing work: filter operators (`:`, `:>=`, …),
+    /// quote delimiters, and `*` where it is a live wildcard.
+    Operator,
+    /// The value of a recognized filter.
+    Argument,
+    /// The value of a recognized filter that the engine would reject
+    /// (unknown type name, bad date, invalid regex).
+    InvalidArg,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct Seg {
+    pub range: Range<usize>,
+    pub class: Class,
+    /// Part of a complete recognized filter — drawn on the chip tint.
+    pub chip: bool,
+}
+
+/// Classify `text` into contiguous segments tiling `0..text.len()`.
+pub fn classify(text: &str) -> Vec<Seg> {
+    let (tokens, err) = tokenize_spanned(text);
+    let mut em = Emitter {
+        text,
+        cursor: 0,
+        segs: Vec::new(),
+    };
+
+    let tok = |i: usize| tokens.get(i).map(|(t, _)| t);
+    let span = |i: usize| tokens[i].1.clone();
+
+    let mut i = 0usize;
+    // The engine allows one `regex:` per query; later ones are errors.
+    let mut regex_seen = false;
+
+    while i < tokens.len() {
+        match &tokens[i].0 {
+            Token::Word(word) => {
+                if let Some(Token::Op(op1)) = tok(i + 1) {
+                    // Candidate filter: Word(key) Op [Op] (Word|Quoted),
+                    // exactly as split_for_cascade sees it.
+                    let (op, op_end_idx, value_idx) = match tok(i + 2) {
+                        Some(Token::Op(op2)) => (*op2, i + 2, i + 3),
+                        _ => (*op1, i + 1, i + 2),
+                    };
+                    let value = match tok(value_idx) {
+                        Some(Token::Word(v)) | Some(Token::Quoted(v)) => Some(v.clone()),
+                        _ => None,
+                    };
+                    let is_regex = word.eq_ignore_ascii_case("regex");
+                    if let Some(value) = value {
+                        let value_is_word = matches!(tok(value_idx), Some(Token::Word(_)));
+                        if is_regex || is_filter_key(word) {
+                            let valid = if is_regex {
+                                let first = !regex_seen;
+                                regex_seen = true;
+                                first
+                                    && op == Op::Contains
+                                    && RegexQuery::new(&value).is_ok()
+                            } else {
+                                build_filter(word, op, &value, value_is_word).is_ok()
+                            };
+                            em.emit(span(i), Class::Keyword, true, false);
+                            for op_idx in (i + 1)..=op_end_idx {
+                                em.emit(span(op_idx), Class::Operator, true, true);
+                            }
+                            let vspan = span(value_idx);
+                            if !valid {
+                                // One uniform error run reads better than
+                                // error-with-green-sprinkles.
+                                em.emit(vspan, Class::InvalidArg, true, true);
+                            } else if !value_is_word {
+                                em.emit_quoted(vspan, Class::Argument, true);
+                            } else if glob_value_key(word) {
+                                em.emit_word(vspan, Class::Argument, true, true);
+                            } else {
+                                // Stars in other filter values are literal
+                                // characters — no wildcard color.
+                                em.emit(vspan, Class::Argument, true, true);
+                            }
+                            i = value_idx + 1;
+                            continue;
+                        }
+                        // Unrecognized key: the engine reassembles the whole
+                        // chain verbatim (stars stay literal), so everything
+                        // renders plain — that absence of color is how the
+                        // user learns `foo:` is not a filter.
+                        em.emit(span(i), Class::Plain, false, false);
+                        for op_idx in (i + 1)..=op_end_idx {
+                            em.emit(span(op_idx), Class::Plain, false, false);
+                        }
+                        em.emit_glued_value(span(value_idx), tok(value_idx));
+                        i = value_idx + 1;
+                        while let Some(Token::Op(_)) = tok(i) {
+                            em.emit(span(i), Class::Plain, false, false);
+                            i += 1;
+                            if let Some(Token::Word(_)) | Some(Token::Quoted(_)) = tok(i) {
+                                em.emit_glued_value(span(i), tok(i));
+                                i += 1;
+                            }
+                        }
+                        continue;
+                    }
+                    // Key + op with no value yet (mid-typing `type:`).
+                    // Recognized keys color optimistically — instant
+                    // feedback that the key landed — but earn no chip
+                    // until the filter is complete.
+                    let known = is_regex || is_filter_key(word);
+                    let (key_class, op_class) = if known {
+                        (Class::Keyword, Class::Operator)
+                    } else {
+                        (Class::Plain, Class::Plain)
+                    };
+                    em.emit(span(i), key_class, false, false);
+                    for op_idx in (i + 1)..=op_end_idx {
+                        em.emit(span(op_idx), op_class, false, false);
+                    }
+                    i = op_end_idx + 1;
+                    continue;
+                }
+                // A plain word: unquoted stars are live wildcards.
+                em.emit_word(span(i), Class::Plain, false, false);
+            }
+            Token::Quoted(_) => em.emit_quoted(span(i), Class::Plain, false),
+            // Demoted to plain text by the live search path — coloring
+            // them as operators would lie.
+            Token::And | Token::Or | Token::LParen | Token::RParen | Token::Op(_) => {
+                em.emit(span(i), Class::Plain, false, false);
+            }
+        }
+        i += 1;
+    }
+
+    // Trailing lex error: an unterminated quote is a quote-in-progress,
+    // not a mistake — green delimiter, plain tail.
+    if let Some(err) = err {
+        if err.offset < text.len() && text.as_bytes()[err.offset] == b'"' {
+            em.emit(err.offset..err.offset + 1, Class::Operator, false, false);
+        }
+    }
+    em.finish(text.len())
+}
+
+/// Keys whose word-form values interpret `*` as a wildcard (or, for
+/// `regex`, as live pattern syntax).
+fn glob_value_key(key: &str) -> bool {
+    matches!(
+        key.to_ascii_lowercase().as_str(),
+        "name" | "filename" | "regex"
+    )
+}
+
+struct Emitter<'a> {
+    text: &'a str,
+    cursor: usize,
+    segs: Vec<Seg>,
+}
+
+impl Emitter<'_> {
+    /// Fill the gap (whitespace the lexer skipped) up to `pos`.
+    fn gap_to(&mut self, pos: usize, chip: bool) {
+        if pos > self.cursor {
+            self.segs.push(Seg {
+                range: self.cursor..pos,
+                class: Class::Plain,
+                chip,
+            });
+            self.cursor = pos;
+        }
+    }
+
+    /// Emit one span. `gap_chip` tints the whitespace before it — true for
+    /// the interior of a filter (`type : Audio` chips as one run).
+    fn emit(&mut self, range: Range<usize>, class: Class, chip: bool, gap_chip: bool) {
+        self.gap_to(range.start, gap_chip);
+        if range.end > range.start {
+            self.segs.push(Seg {
+                range: range.clone(),
+                class,
+                chip,
+            });
+            self.cursor = range.end;
+        }
+    }
+
+    /// Emit a word span with each `*` as a green wildcard and the pieces
+    /// between in `base`.
+    fn emit_word(&mut self, range: Range<usize>, base: Class, chip: bool, gap_chip: bool) {
+        self.gap_to(range.start, gap_chip);
+        let bytes = self.text.as_bytes();
+        let mut piece_start = range.start;
+        for pos in range.clone() {
+            if bytes[pos] == b'*' {
+                if pos > piece_start {
+                    self.segs.push(Seg {
+                        range: piece_start..pos,
+                        class: base,
+                        chip,
+                    });
+                }
+                self.segs.push(Seg {
+                    range: pos..pos + 1,
+                    class: Class::Operator,
+                    chip,
+                });
+                piece_start = pos + 1;
+            }
+        }
+        if range.end > piece_start {
+            self.segs.push(Seg {
+                range: piece_start..range.end,
+                class: base,
+                chip,
+            });
+        }
+        self.cursor = self.cursor.max(range.end);
+    }
+
+    /// Emit a quoted span (delimiters included): quotes green, content in
+    /// `content`. Inner `""` escapes are just content bytes — no offset
+    /// math needed.
+    fn emit_quoted(&mut self, range: Range<usize>, content: Class, chip: bool) {
+        self.gap_to(range.start, chip);
+        self.segs.push(Seg {
+            range: range.start..range.start + 1,
+            class: Class::Operator,
+            chip,
+        });
+        if range.end - range.start > 2 {
+            self.segs.push(Seg {
+                range: range.start + 1..range.end - 1,
+                class: content,
+                chip,
+            });
+        }
+        if range.end - range.start >= 2 {
+            self.segs.push(Seg {
+                range: range.end - 1..range.end,
+                class: Class::Operator,
+                chip,
+            });
+        }
+        self.cursor = self.cursor.max(range.end);
+    }
+
+    /// A value inside unrecognized-key glue: plain, except quote
+    /// delimiters, which still did real tokenizing work.
+    fn emit_glued_value(&mut self, range: Range<usize>, token: Option<&Token>) {
+        match token {
+            Some(Token::Quoted(_)) => self.emit_quoted(range, Class::Plain, false),
+            _ => self.emit(range, Class::Plain, false, false),
+        }
+    }
+
+    fn finish(mut self, len: usize) -> Vec<Seg> {
+        self.gap_to(len, false);
+        self.segs
+    }
+}
+
+// ---------------------------------------------------------------------------
+// egui layer
+// ---------------------------------------------------------------------------
+
+struct QueryPalette {
+    keyword: Color32,
+    argument: Color32,
+    operator: Color32,
+}
+
+/// GitHub Primer syntax colors — readable on egui's near-black and white
+/// text-field backgrounds. Same convention as `rank_tier_color`.
+fn query_palette(dark_mode: bool) -> QueryPalette {
+    if dark_mode {
+        QueryPalette {
+            keyword: Color32::from_rgb(255, 123, 114),
+            argument: Color32::from_rgb(121, 192, 255),
+            operator: Color32::from_rgb(126, 231, 135),
+        }
+    } else {
+        QueryPalette {
+            keyword: Color32::from_rgb(207, 34, 46),
+            argument: Color32::from_rgb(5, 80, 174),
+            operator: Color32::from_rgb(26, 127, 55),
+        }
+    }
+}
+
+struct QueryFormats {
+    plain: TextFormat,
+    keyword: TextFormat,
+    operator: TextFormat,
+    argument: TextFormat,
+    invalid: TextFormat,
+    chip_bg: Color32,
+}
+
+fn query_formats(ui: &egui::Ui) -> QueryFormats {
+    let font_id = egui::TextStyle::Body.resolve(ui.style());
+    let palette = query_palette(ui.visuals().dark_mode);
+    let base = |color: Color32| TextFormat {
+        font_id: font_id.clone(),
+        color,
+        ..Default::default()
+    };
+    let error = ui.visuals().error_fg_color;
+    QueryFormats {
+        plain: base(ui.visuals().text_color()),
+        keyword: base(palette.keyword),
+        operator: base(palette.operator),
+        argument: base(palette.argument),
+        // The keyword red and the error red are near neighbors in dark
+        // mode; the underline disambiguates at a glance.
+        invalid: TextFormat {
+            underline: Stroke::new(1.0, error),
+            ..base(error)
+        },
+        // Slightly weaker than the snippet highlight's 0.4 so the colored
+        // text on top stays crisp.
+        chip_bg: ui.visuals().selection.bg_fill.gamma_multiply(0.35),
+    }
+}
+
+impl QueryFormats {
+    fn format_for(&self, seg: &Seg) -> TextFormat {
+        let mut fmt = match seg.class {
+            Class::Plain => self.plain.clone(),
+            Class::Keyword => self.keyword.clone(),
+            Class::Operator => self.operator.clone(),
+            Class::Argument => self.argument.clone(),
+            Class::InvalidArg => self.invalid.clone(),
+        };
+        if seg.chip {
+            fmt.background = self.chip_bg;
+        }
+        fmt
+    }
+}
+
+/// Classification cache: tokenizing is cheap but validating a `regex:`
+/// argument compiles the regex, and the layouter runs every frame — so
+/// segments are recomputed only when the text changes. The `LayoutJob` is
+/// rebuilt each frame (colors follow the live theme) and epaint's own
+/// galley cache dedupes the actual layout work by job hash.
+#[derive(Default)]
+pub struct HighlightCache {
+    text: String,
+    segs: Vec<Seg>,
+}
+
+pub fn galley(ui: &egui::Ui, cache: &mut HighlightCache, text: &str) -> Arc<Galley> {
+    if cache.text != text {
+        cache.text = text.to_owned();
+        cache.segs = classify(text);
+    }
+    let fmts = query_formats(ui);
+    let mut job = LayoutJob::default();
+    for seg in &cache.segs {
+        job.append(&text[seg.range.clone()], 0.0, fmts.format_for(seg));
+    }
+    ui.fonts(|f| f.layout_job(job))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// Readable projection: (slice, class, chip) per segment.
+    fn segs(text: &str) -> Vec<(String, Class, bool)> {
+        assert_tiles(text);
+        classify(text)
+            .into_iter()
+            .map(|s| (text[s.range.clone()].to_string(), s.class, s.chip))
+            .collect()
+    }
+
+    /// Segments must tile 0..len exactly: contiguous, ascending, complete.
+    fn assert_tiles(text: &str) {
+        let segs = classify(text);
+        let mut cursor = 0usize;
+        for s in &segs {
+            assert_eq!(s.range.start, cursor, "gap or overlap in {:?}: {:?}", text, segs);
+            assert!(s.range.end > s.range.start, "empty seg in {:?}", text);
+            cursor = s.range.end;
+        }
+        assert_eq!(cursor, text.len(), "segments must cover {:?}", text);
+    }
+
+    use Class::*;
+
+    fn owned(v: Vec<(&str, Class, bool)>) -> Vec<(String, Class, bool)> {
+        v.into_iter().map(|(s, c, b)| (s.to_string(), c, b)).collect()
+    }
+
+    #[test]
+    fn empty_input_yields_no_segments() {
+        assert!(classify("").is_empty());
+    }
+
+    #[test]
+    fn plain_words_stay_plain() {
+        assert_eq!(
+            segs("budget report"),
+            owned(vec![
+                ("budget", Plain, false),
+                (" ", Plain, false),
+                ("report", Plain, false),
+            ])
+        );
+    }
+
+    #[test]
+    fn every_recognized_filter_chips() {
+        for input in [
+            "type:Audio",
+            "modified:>=2024-01-01",
+            "mtime:<2023-12-01",
+            "path:/home/me",
+            "folder:/x",
+            "includefolder:/x",
+            "name:report",
+            "filename:report",
+            "mime:application/pdf",
+            "regex:foo",
+        ] {
+            let all = segs(input);
+            assert!(
+                all.iter().all(|(_, _, chip)| *chip),
+                "{:?}: whole filter must chip: {:?}",
+                input,
+                all
+            );
+            assert_eq!(all[0].1, Keyword, "{:?}", input);
+            assert_eq!(all[1].1, Operator, "{:?}", input);
+            assert!(
+                all[2..].iter().all(|(_, c, _)| *c == Argument || *c == Operator),
+                "{:?}: {:?}",
+                input,
+                all
+            );
+        }
+    }
+
+    #[test]
+    fn keys_are_case_insensitive() {
+        assert_eq!(segs("TYPE:Audio")[0], ("TYPE".to_string(), Keyword, true));
+        assert_eq!(
+            segs("Modified:>=2024-01-01")[0],
+            ("Modified".to_string(), Keyword, true)
+        );
+    }
+
+    #[test]
+    fn colon_comparator_runs_are_one_green_stretch() {
+        assert_eq!(
+            segs("modified:>=2024-01-01"),
+            owned(vec![
+                ("modified", Keyword, true),
+                (":", Operator, true),
+                (">=", Operator, true),
+                ("2024-01-01", Argument, true),
+            ])
+        );
+    }
+
+    #[test]
+    fn unrecognized_keys_stay_plain() {
+        for input in ["foo:bar", "12:30", "foo:bar:baz"] {
+            assert!(
+                segs(input).iter().all(|(_, c, chip)| *c == Plain && !chip),
+                "{:?}: {:?}",
+                input,
+                segs(input)
+            );
+        }
+        // Stars in glue are literal to the engine — no green.
+        assert!(segs("foo:ba*r").iter().all(|(_, c, _)| *c == Plain));
+    }
+
+    #[test]
+    fn drive_letters_do_not_split() {
+        assert_eq!(
+            segs(r"path:C:\Users\me"),
+            owned(vec![
+                ("path", Keyword, true),
+                (":", Operator, true),
+                (r"C:\Users\me", Argument, true),
+            ])
+        );
+        assert_eq!(
+            segs(r"C:\data"),
+            owned(vec![(r"C:\data", Plain, false)])
+        );
+    }
+
+    #[test]
+    fn quoted_phrases_get_green_delimiters() {
+        assert_eq!(
+            segs("\"exact phrase\""),
+            owned(vec![
+                ("\"", Operator, false),
+                ("exact phrase", Plain, false),
+                ("\"", Operator, false),
+            ])
+        );
+        // Inner "" escapes are content bytes.
+        assert_eq!(
+            segs("\"a\"\"b\""),
+            owned(vec![
+                ("\"", Operator, false),
+                ("a\"\"b", Plain, false),
+                ("\"", Operator, false),
+            ])
+        );
+        // Quoted stars are literal — content stays plain.
+        assert!(segs("\"a*b\"").iter().all(|(s, c, _)| s == "\""
+            || *c == Plain));
+    }
+
+    #[test]
+    fn quoted_filter_values_are_blue_with_green_quotes() {
+        assert_eq!(
+            segs("path:\"/home/me/My Docs\""),
+            owned(vec![
+                ("path", Keyword, true),
+                (":", Operator, true),
+                ("\"", Operator, true),
+                ("/home/me/My Docs", Argument, true),
+                ("\"", Operator, true),
+            ])
+        );
+        // Empty quoted value: two delimiters, no content seg, no panic.
+        assert_eq!(
+            segs("path:\"\""),
+            owned(vec![
+                ("path", Keyword, true),
+                (":", Operator, true),
+                ("\"", Operator, true),
+                ("\"", Operator, true),
+            ])
+        );
+    }
+
+    #[test]
+    fn unterminated_quote_is_a_quote_in_progress() {
+        assert_eq!(
+            segs("\"unclosed phrase"),
+            owned(vec![
+                ("\"", Operator, false),
+                ("unclosed phrase", Plain, false),
+            ])
+        );
+        // Filters before the open quote keep their colors.
+        let all = segs("type:Audio \"x");
+        assert_eq!(all[0], ("type".to_string(), Keyword, true));
+        assert_eq!(all[4], ("\"".to_string(), Operator, false));
+        assert_eq!(all[5], ("x".to_string(), Plain, false));
+    }
+
+    #[test]
+    fn trailing_bare_keys_color_optimistically_without_chip() {
+        assert_eq!(
+            segs("type:"),
+            owned(vec![("type", Keyword, false), (":", Operator, false)])
+        );
+        assert_eq!(
+            segs("modified:>="),
+            owned(vec![
+                ("modified", Keyword, false),
+                (":", Operator, false),
+                (">=", Operator, false),
+            ])
+        );
+        assert_eq!(
+            segs("foo:"),
+            owned(vec![("foo", Plain, false), (":", Plain, false)])
+        );
+    }
+
+    #[test]
+    fn stars_in_words_and_name_values_go_green() {
+        assert_eq!(
+            segs("rep*ort"),
+            owned(vec![
+                ("rep", Plain, false),
+                ("*", Operator, false),
+                ("ort", Plain, false),
+            ])
+        );
+        assert_eq!(
+            segs("name:re*.txt"),
+            owned(vec![
+                ("name", Keyword, true),
+                (":", Operator, true),
+                ("re", Argument, true),
+                ("*", Operator, true),
+                (".txt", Argument, true),
+            ])
+        );
+        // Edge and doubled stars keep tiling intact.
+        assert_tiles("*foo");
+        assert_tiles("foo*");
+        assert_tiles("**");
+        assert_tiles("*");
+        // In non-glob filter values the star is a literal character.
+        assert_eq!(
+            segs("path:/da*ta")[2],
+            ("/da*ta".to_string(), Argument, true)
+        );
+    }
+
+    #[test]
+    fn invalid_arguments_go_error_uniformly() {
+        // (`regex:(` is not here: `(` lexes as a paren, so that input is an
+        // *incomplete* filter — bare-key optimism applies, not an error.)
+        for input in ["type:NotAThing", "modified:>=tomorrow", "regex:[", "type:Doc*"] {
+            let all = segs(input);
+            assert_eq!(all[0].1, Keyword, "{:?}", input);
+            let last = all.last().unwrap();
+            assert_eq!(last.1, InvalidArg, "{:?}: {:?}", input, all);
+            assert!(last.2, "invalid values keep the chip: {:?}", input);
+        }
+        // name:= is an unsupported op → its value is invalid too.
+        let all = segs("name=x");
+        assert_eq!(all.last().unwrap().1, InvalidArg);
+    }
+
+    #[test]
+    fn valid_regex_argument_is_blue_with_green_stars() {
+        assert_eq!(
+            segs("regex:foo.*bar"),
+            owned(vec![
+                ("regex", Keyword, true),
+                (":", Operator, true),
+                ("foo.", Argument, true),
+                ("*", Operator, true),
+                ("bar", Argument, true),
+            ])
+        );
+    }
+
+    #[test]
+    fn a_second_regex_filter_is_invalid() {
+        let all = segs("regex:foo regex:bar");
+        assert_eq!(all[2], ("foo".to_string(), Argument, true));
+        assert_eq!(all.last().unwrap(), &("bar".to_string(), InvalidArg, true));
+    }
+
+    #[test]
+    fn multi_filter_queries_chip_separately() {
+        let all = segs("type:Document budget modified:>=2024-01-01");
+        // The word and the whitespace around it stay un-chipped.
+        assert_eq!(
+            all.iter()
+                .filter(|(_, _, chip)| !chip)
+                .map(|(s, _, _)| s.as_str())
+                .collect::<Vec<_>>(),
+            vec![" ", "budget", " "]
+        );
+    }
+
+    #[test]
+    fn spaced_filters_chip_their_interior_gaps() {
+        // `type : Audio` is still a filter to the lexer/splitter.
+        assert_eq!(
+            segs("type : Audio"),
+            owned(vec![
+                ("type", Keyword, true),
+                (" ", Plain, true),
+                (":", Operator, true),
+                (" ", Plain, true),
+                ("Audio", Argument, true),
+            ])
+        );
+    }
+
+    #[test]
+    fn demoted_operators_stay_plain() {
+        assert!(
+            segs("(alpha AND beta) OR gamma")
+                .iter()
+                .all(|(_, c, chip)| *c == Plain && !chip)
+        );
+        // Dangling comparators are literal text.
+        assert!(segs("a > b").iter().all(|(_, c, _)| *c == Plain));
+        // Leading operator, nothing else.
+        assert!(segs(">foo").iter().all(|(_, c, _)| *c == Plain));
+    }
+
+    #[test]
+    fn adjacency_between_filter_and_quote() {
+        // `Audio"q"`: the word ends at the quote; the filter is complete
+        // and the quoted phrase stands alone.
+        let all = segs("type:Audio\"q\"");
+        assert_eq!(all[2], ("Audio".to_string(), Argument, true));
+        assert_eq!(all[3], ("\"".to_string(), Operator, false));
+    }
+
+    #[test]
+    fn unicode_offsets_hold_up() {
+        assert_tiles("\"José\" type:Audio naïve*file");
+        let all = segs("naïve*café");
+        assert_eq!(
+            all,
+            owned(vec![
+                ("naïve", Plain, false),
+                ("*", Operator, false),
+                ("café", Plain, false),
+            ])
+        );
+    }
+}

crates/quicksearch-gui/src/search.rs

@@ -1,527 +0,0 @@
-#![allow(non_snake_case)]
-
-use std::sync::Arc;
-use std::time::Instant;
-use dioxus::prelude::*;
-use quicksearch_core::indexing::{IndexingService, SearchResult};
-use quicksearch_core::search_sql::{build_count, build_select, SearchArgs};
-
-/// One page of results. Tuned to keep DOM size bounded — rendering ten
-/// thousand `<tr>` nodes wedges WebKit for tens of seconds.
-const PAGE_SIZE: u32 = 50;
-
-#[derive(Props, Clone)]
-pub struct SearchProps {
-    pub indexing_service: Arc<IndexingService>,
-    pub db_path: String,
-}
-
-impl PartialEq for SearchProps {
-    fn eq(&self, other: &Self) -> bool {
-        Arc::ptr_eq(&self.indexing_service, &other.indexing_service) && self.db_path == other.db_path
-    }
-}
-
-pub fn Search(props: SearchProps) -> Element {
-    let mut search_type = use_signal(|| "fulltext".to_string());
-    let mut search_term = use_signal(|| String::new());
-    let mut fulltext_exact = use_signal(|| false);
-    let mut fulltext_case_sensitive = use_signal(|| false);
-    let mut search_results = use_signal(|| Vec::<SearchResult>::new());
-    let mut search_error = use_signal(|| None::<String>);
-    let mut is_searching = use_signal(|| false);
-    let mut last_search_time = use_signal(|| None::<f64>);
-    let mut show_corruption_dialog = use_signal(|| false);
-    let mut current_page = use_signal(|| 1u32);
-    let mut total_count = use_signal(|| None::<u64>);
-    let mut last_args = use_signal(|| None::<SearchArgs>);
-    let mut goto_input = use_signal(|| String::new());
-
-    let service = props.indexing_service.clone();
-    let db_path = props.db_path.clone();
-
-    // Spawn a search task. `refresh_count` is true for fresh searches and
-    // false for in-place page navigation (the cached total still applies).
-    let run_query = {
-        let service = service.clone();
-        let db_path = db_path.clone();
-        move |args: SearchArgs, page: u32, refresh_count: bool| {
-            let service = service.clone();
-            let db_path = db_path.clone();
-            spawn(async move {
-                is_searching.set(true);
-                search_error.set(None);
-                last_search_time.set(None);
-                let start = Instant::now();
-
-                let count_sql = if refresh_count {
-                    match build_count(&args) {
-                        Ok(s) => Some(s),
-                        Err(e) => {
-                            search_error.set(Some(e));
-                            is_searching.set(false);
-                            return;
-                        }
-                    }
-                } else {
-                    None
-                };
-
-                let offset = page.saturating_sub(1).saturating_mul(PAGE_SIZE);
-                // Validate early for the filename/duplicates branch so we
-                // surface parse errors before dispatching the blocking task.
-                let precomputed_select_sql = if args.search_type == "fulltext" {
-                    None
-                } else {
-                    match build_select(&args, PAGE_SIZE, offset) {
-                        Ok(s) => Some(s),
-                        Err(e) => {
-                            search_error.set(Some(e));
-                            is_searching.set(false);
-                            return;
-                        }
-                    }
-                };
-
-                // Fulltext takes the snippet-aware path (decompresses
-                // documents_text and highlights in Rust); filename +
-                // duplicates go through the plain SQL executor.
-                let svc1 = service.clone();
-                let db1 = db_path.clone();
-                let args_for_select = args.clone();
-                let select_handle = tokio::task::spawn_blocking(move || {
-                    if args_for_select.search_type == "fulltext" {
-                        svc1.execute_fulltext_search(&db1, &args_for_select, PAGE_SIZE, offset)
-                    } else {
-                        let sql = precomputed_select_sql
-                            .expect("non-fulltext select SQL was prebuilt above");
-                        svc1.execute_search(&db1, &sql)
-                    }
-                });
-
-                let count_handle = count_sql.map(|sql| {
-                    let svc2 = service.clone();
-                    let db2 = db_path.clone();
-                    tokio::task::spawn_blocking(move || svc2.execute_search(&db2, &sql))
-                });
-
-                let select_run = select_handle.await;
-                let count_run = match count_handle {
-                    Some(h) => Some(h.await),
-                    None => None,
-                };
-
-                let elapsed = start.elapsed().as_secs_f64();
-
-                if let Some(c) = count_run {
-                    match c {
-                        Ok(Ok(rs)) => {
-                            let n = rs
-                                .first()
-                                .and_then(|r| r.rows.first())
-                                .and_then(|r| r.values.first())
-                                .and_then(|s| s.parse::<u64>().ok())
-                                .unwrap_or(0);
-                            total_count.set(Some(n));
-                        }
-                        Ok(Err(e)) => {
-                            handle_query_error(
-                                e,
-                                elapsed,
-                                search_error,
-                                show_corruption_dialog,
-                                last_search_time,
-                                is_searching,
-                            );
-                            return;
-                        }
-                        Err(e) => {
-                            search_error.set(Some(format!("Task execution error: {}", e)));
-                            last_search_time.set(Some(elapsed));
-                            is_searching.set(false);
-                            return;
-                        }
-                    }
-                }
-
-                match select_run {
-                    Ok(Ok(rs)) => {
-                        search_results.set(rs);
-                        current_page.set(page);
-                        last_args.set(Some(args));
-                        last_search_time.set(Some(elapsed));
-                    }
-                    Ok(Err(e)) => {
-                        handle_query_error(
-                            e,
-                            elapsed,
-                            search_error,
-                            show_corruption_dialog,
-                            last_search_time,
-                            is_searching,
-                        );
-                        return;
-                    }
-                    Err(e) => {
-                        search_error.set(Some(format!("Task execution error: {}", e)));
-                        last_search_time.set(Some(elapsed));
-                    }
-                }
-                is_searching.set(false);
-            });
-        }
-    };
-
-    let do_fresh_search = {
-        let run_query = run_query.clone();
-        move || {
-            let args = SearchArgs {
-                search_type: search_type(),
-                term: search_term(),
-                fulltext_exact: fulltext_exact(),
-                fulltext_case_sensitive: fulltext_case_sensitive(),
-            };
-            run_query(args, 1, true);
-        }
-    };
-
-    let do_goto_page = {
-        let run_query = run_query.clone();
-        move |target: u32| {
-            if let Some(args) = last_args() {
-                let total = total_count()
-                    .map(|n| pages_for(n))
-                    .unwrap_or(1)
-                    .max(1);
-                let clamped = target.clamp(1, total);
-                if clamped != current_page() {
-                    run_query(args, clamped, false);
-                }
-            }
-        }
-    };
-
-    let total_pages = total_count().map(pages_for).unwrap_or(0);
-
-    rsx! {
-        div {
-            class: "section",
-            h2 { "Search Database" }
-
-            div {
-                class: "form-group",
-                label { "Search Type: " }
-                select {
-                    class: "form-control",
-                    value: "{search_type}",
-                    onchange: move |evt| search_type.set(evt.value()),
-                    option { value: "fulltext", "Full Text Search" }
-                    option { value: "filename", "Filename Search" }
-                    option { value: "duplicates", "Find Duplicate Files" }
-                }
-            }
-
-            if search_type() == "fulltext" {
-                div {
-                    class: "form-group",
-                    style: "display: flex; flex-direction: column; gap: 6px;",
-                    span { style: "font-weight: 600;", "Full text options" }
-                    label {
-                        style: "display: flex; align-items: center; gap: 8px; cursor: pointer;",
-                        input {
-                            r#type: "checkbox",
-                            checked: fulltext_exact(),
-                            onchange: move |evt| fulltext_exact.set(evt.checked()),
-                        }
-                        "Exact phrase match"
-                    }
-                    label {
-                        style: "display: flex; align-items: center; gap: 8px; cursor: pointer;",
-                        input {
-                            r#type: "checkbox",
-                            checked: fulltext_case_sensitive(),
-                            onchange: move |evt| fulltext_case_sensitive.set(evt.checked()),
-                        }
-                        "Case-sensitive match"
-                    }
-                }
-            }
-
-            if search_type() != "duplicates" {
-                div {
-                    class: "form-group",
-                    label { "Search Term: " }
-                    input {
-                        class: "form-control",
-                        r#type: "text",
-                        value: "{search_term}",
-                        oninput: move |evt| search_term.set(evt.value()),
-                        onkeydown: {
-                            let do_fresh_search = do_fresh_search.clone();
-                            move |evt| {
-                                if evt.code() == dioxus::events::Code::Enter {
-                                    do_fresh_search();
-                                }
-                            }
-                        }
-                    }
-                }
-            }
-
-            div {
-                style: "display: flex; align-items: center; gap: 10px;",
-                button {
-                    class: "btn btn-info",
-                    disabled: is_searching(),
-                    onclick: {
-                        let do_fresh_search = do_fresh_search.clone();
-                        move |_| { do_fresh_search(); }
-                    },
-                    "Search"
-                }
-
-                if is_searching() {
-                    div {
-                        class: "loading",
-                        title: "Searching..."
-                    }
-                } else if let Some(elapsed) = last_search_time() {
-                    span {
-                        style: "color: #666; font-size: 0.9em;",
-                        "Search completed in {elapsed:.3}s"
-                    }
-                }
-            }
-
-            if let Some(error) = search_error() {
-                div {
-                    class: "error-message",
-                    "Error: {error}"
-                }
-            }
-
-            // Results panel: pagination header + bounded table. Only renders
-            // when at least one search has completed (last_args is Some).
-            if last_args().is_some() {
-                div {
-                    class: "search-results",
-                    {
-                        let total_str = match total_count() {
-                            Some(n) => format!("{}", n),
-                            None => "?".to_string(),
-                        };
-                        let page_first = (current_page().saturating_sub(1) as u64) * PAGE_SIZE as u64 + 1;
-                        let page_last_calc = page_first + search_results().first().map(|r| r.rows.len() as u64).unwrap_or(0).saturating_sub(1);
-                        let header = if total_count() == Some(0) {
-                            "No results.".to_string()
-                        } else {
-                            format!(
-                                "Showing {}-{} of {} (page {} of {})",
-                                page_first,
-                                page_last_calc,
-                                total_str,
-                                current_page(),
-                                total_pages
-                            )
-                        };
-                        rsx! { h3 { "{header}" } }
-                    }
-
-                    // Pagination controls. Hidden if there's only one page.
-                    if total_pages > 1 {
-                        div {
-                            style: "display: flex; align-items: center; gap: 8px; margin: 8px 0;",
-                            button {
-                                class: "btn",
-                                disabled: is_searching() || current_page() <= 1,
-                                onclick: {
-                                    let do_goto_page = do_goto_page.clone();
-                                    move |_| do_goto_page(1)
-                                },
-                                "« First"
-                            }
-                            button {
-                                class: "btn",
-                                disabled: is_searching() || current_page() <= 1,
-                                onclick: {
-                                    let do_goto_page = do_goto_page.clone();
-                                    move |_| do_goto_page(current_page().saturating_sub(1))
-                                },
-                                "‹ Prev"
-                            }
-                            button {
-                                class: "btn",
-                                disabled: is_searching() || current_page() >= total_pages,
-                                onclick: {
-                                    let do_goto_page = do_goto_page.clone();
-                                    move |_| do_goto_page(current_page().saturating_add(1))
-                                },
-                                "Next ›"
-                            }
-                            button {
-                                class: "btn",
-                                disabled: is_searching() || current_page() >= total_pages,
-                                onclick: {
-                                    let do_goto_page = do_goto_page.clone();
-                                    move |_| do_goto_page(total_pages)
-                                },
-                                "Last »"
-                            }
-                            span { "Go to:" }
-                            input {
-                                r#type: "number",
-                                style: "width: 70px;",
-                                value: "{goto_input}",
-                                oninput: move |evt| goto_input.set(evt.value()),
-                                onkeydown: {
-                                    let do_goto_page = do_goto_page.clone();
-                                    move |evt| {
-                                        if evt.code() == dioxus::events::Code::Enter {
-                                            if let Ok(p) = goto_input().trim().parse::<u32>() {
-                                                do_goto_page(p);
-                                                goto_input.set(String::new());
-                                            }
-                                        }
-                                    }
-                                }
-                            }
-                        }
-                    }
-
-                    if !search_results().is_empty() && !search_results()[0].rows.is_empty() {
-                        div {
-                            class: "results-table",
-                            table {
-                                thead {
-                                    tr {
-                                        for column in search_results()[0].columns.iter() {
-                                            th { "{column}" }
-                                        }
-                                    }
-                                }
-                                tbody {
-                                    for row in search_results()[0].rows.iter() {
-                                        tr {
-                                            for (col_index, value) in row.values.iter().enumerate() {
-                                                if search_results()[0].columns.get(col_index).map(|s| s.as_str()) == Some("path") {
-                                                    {
-                                                        let value_owned = value.clone();
-                                                        let service_owned = props.indexing_service.clone();
-                                                        rsx! {
-                                                            td {
-                                                                class: "path-cell clickable",
-                                                                onclick: move |_| {
-                                                                    let path = value_owned.clone();
-                                                                    let service_clone = service_owned.clone();
-                                                                    spawn(async move {
-                                                                        if let Err(e) = service_clone.open_file_explorer(&path) {
-                                                                            eprintln!("Failed to open file explorer: {}", e);
-                                                                        }
-                                                                    });
-                                                                },
-                                                                title: "Click to open in file explorer",
-                                                                dangerous_inner_html: "{value}"
-                                                            }
-                                                        }
-                                                    }
-                                                } else {
-                                                    td {
-                                                        dangerous_inner_html: "{value}"
-                                                    }
-                                                }
-                                            }
-                                        }
-                                    }
-                                }
-                            }
-                        }
-                    }
-                }
-            }
-
-            if show_corruption_dialog() {
-                div {
-                    class: "modal-backdrop",
-                    div {
-                        class: "modal-dialog",
-                        h3 {
-                            style: "margin-top: 0; color: #d32f2f;",
-                            "⚠️ Database Corruption Detected"
-                        }
-                        p {
-                            style: "margin: 15px 0;",
-                            "The database appears to be corrupted or malformed. This can happen due to unexpected shutdowns or disk issues."
-                        }
-                        p {
-                            style: "margin: 15px 0; font-weight: bold;",
-                            "Would you like to delete the corrupted database and create a new one? This will require re-indexing your files."
-                        }
-                        div {
-                            style: "display: flex; gap: 10px; margin-top: 20px;",
-                            button {
-                                style: "padding: 10px 20px; background-color: #d32f2f; color: white; border: none; border-radius: 5px; cursor: pointer;",
-                                onclick: move |_| {
-                                    let service = props.indexing_service.clone();
-                                    let db = props.db_path.clone();
-
-                                    show_corruption_dialog.set(false);
-                                    search_error.set(Some("Deleting corrupted database...".to_string()));
-
-                                    spawn(async move {
-                                        match service.delete_index_for_rebuild(&db) {
-                                            Ok(()) => {
-                                                search_error.set(Some("Database deleted. You can now start indexing again.".to_string()));
-                                            }
-                                            Err(e) => {
-                                                search_error.set(Some(format!("Error deleting database: {}", e)));
-                                            }
-                                        }
-                                    });
-                                },
-                                "Yes, Delete & Rebuild"
-                            }
-                            button {
-                                style: "padding: 10px 20px; background-color: #666; color: white; border: none; border-radius: 5px; cursor: pointer;",
-                                onclick: move |_| {
-                                    show_corruption_dialog.set(false);
-                                },
-                                "Cancel"
-                            }
-                        }
-                    }
-                }
-            }
-        }
-    }
-}
-
-/// Number of pages required to fit `total` rows at [`PAGE_SIZE`] per page.
-/// Returns 0 for 0 rows so callers can branch on "no results yet".
-fn pages_for(total: u64) -> u32 {
-    if total == 0 {
-        0
-    } else {
-        ((total - 1) / PAGE_SIZE as u64) as u32 + 1
-    }
-}
-
-/// Centralized error-router for the two query branches that can fail
-/// identically (count and select). Sets the error/timing/spinner signals
-/// and pops the corruption dialog when warranted.
-fn handle_query_error(
-    e: String,
-    elapsed: f64,
-    mut search_error: Signal<Option<String>>,
-    mut show_corruption_dialog: Signal<bool>,
-    mut last_search_time: Signal<Option<f64>>,
-    mut is_searching: Signal<bool>,
-) {
-    if e.starts_with("DATABASE_CORRUPTED:") {
-        search_error.set(Some("Database appears to be corrupted".into()));
-        show_corruption_dialog.set(true);
-    } else {
-        search_error.set(Some(e));
-    }
-    last_search_time.set(Some(elapsed));
-    is_searching.set(false);
-}

crates/quicksearch-gui/src/search_tab.rs

@@ -0,0 +1,947 @@
+//! The Search tab: query strip, streaming results table, snippet
+//! preview, context menu, ignore-filter dialog, and syntax help.
+
+use std::time::Instant;
+
+use egui::text::{LayoutJob, TextFormat};
+use egui_extras::{Column, TableBuilder};
+use quicksearch_core::search::{SearchHit, SearchUpdate};
+use quicksearch_core::snippet::Snippet;
+
+use crate::format::{fmt_elapsed, fmt_mtime, human_size};
+use crate::platform;
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum SortKey {
+    Rank,
+    Name,
+    Path,
+    Size,
+    Modified,
+}
+
+pub struct IgnoreDialog {
+    pub source_name: String,
+    pub source_path: String,
+    pub pattern: String,
+    pub persist: bool,
+}
+
+/// What the tab asks the app to do after this frame.
+#[derive(Default)]
+pub struct SearchActions {
+    /// Re-run the search (query/fuzzy/session filters changed *now*, not
+    /// debounced — e.g. a chip was removed).
+    pub rerun: bool,
+    /// Persist an ignore pattern into the config.
+    pub persist_ignore: Option<String>,
+    /// The fuzzy toggle changed; remember it in the config.
+    pub save_fuzzy_default: Option<bool>,
+}
+
+pub struct SearchTab {
+    pub query: String,
+    pub fuzzy: bool,
+    /// Set on every edit; the app fires the search after the debounce.
+    pub pending_edit: Option<Instant>,
+    pub generation: u64,
+    pub results: Vec<SearchHit>,
+    /// The next search's hits, held back while the old table fades out;
+    /// swapped into `results` at zero opacity. Prevents the empty-refill
+    /// strobe while typing.
+    staging: Vec<SearchHit>,
+    staging_has_snippets: bool,
+    /// True from search start until the staged set has been swapped in.
+    swap_pending: bool,
+    /// Display permutation over `results`.
+    order: Vec<u32>,
+    sort: (SortKey, bool),
+    sort_dirty: bool,
+    pub selected: Option<u32>,
+    pub running: bool,
+    /// When the in-flight search was submitted.
+    search_started: Option<Instant>,
+    /// Wall time of the last completed search (all cascade passes).
+    elapsed: Option<std::time::Duration>,
+    pub limited: bool,
+    pub error: Option<String>,
+    pub session_ignores: Vec<String>,
+    pub ignore_dialog: Option<IgnoreDialog>,
+    pub help_open: bool,
+    has_snippets: bool,
+    focus_query: bool,
+    /// Query syntax-highlight segments, cached per text.
+    highlight: crate::query_highlight::HighlightCache,
+}
+
+impl SearchTab {
+    pub fn new(fuzzy_default: bool) -> SearchTab {
+        SearchTab {
+            query: String::new(),
+            fuzzy: fuzzy_default,
+            pending_edit: None,
+            generation: 0,
+            results: Vec::new(),
+            staging: Vec::new(),
+            staging_has_snippets: false,
+            swap_pending: false,
+            order: Vec::new(),
+            sort: (SortKey::Rank, true),
+            sort_dirty: false,
+            selected: None,
+            running: false,
+            search_started: None,
+            elapsed: None,
+            limited: false,
+            error: None,
+            session_ignores: Vec::new(),
+            ignore_dialog: None,
+            help_open: false,
+            has_snippets: false,
+            focus_query: true,
+            highlight: Default::default(),
+        }
+    }
+
+    /// Pre-fill the query and let the normal debounce path run it, so a
+    /// command-line query lands the user on results rather than an empty box.
+    pub fn seed(&mut self, query: String) {
+        self.query = query;
+        self.pending_edit = Some(Instant::now());
+    }
+
+    /// A new search was submitted under `generation`. The previous
+    /// results stay on screen (fading out); the new ones stage until the
+    /// fade reaches zero.
+    pub fn on_search_started(&mut self, generation: u64) {
+        self.generation = generation;
+        self.staging.clear();
+        self.staging_has_snippets = false;
+        self.swap_pending = true;
+        self.running = true;
+        self.search_started = Some(Instant::now());
+        self.elapsed = None;
+        self.limited = false;
+        self.error = None;
+    }
+
+    pub fn apply_update(&mut self, update: SearchUpdate, display_limit: usize) {
+        if update.generation() != self.generation {
+            return;
+        }
+        match update {
+            SearchUpdate::Started { .. } => {}
+            SearchUpdate::Hits { hits, .. } => {
+                if self.swap_pending {
+                    // Old results are still fading out; hold the new ones.
+                    for hit in hits {
+                        if self.staging.len() >= display_limit {
+                            self.limited = true;
+                            break;
+                        }
+                        self.staging_has_snippets |= hit.snippet.is_some();
+                        self.staging.push(hit);
+                    }
+                } else {
+                    // Post-swap stream: later cascade passes append live.
+                    for hit in hits {
+                        if self.results.len() >= display_limit {
+                            self.limited = true;
+                            break;
+                        }
+                        self.has_snippets |= hit.snippet.is_some();
+                        self.results.push(hit);
+                    }
+                    // Arrival order *is* rank order, so the default sort
+                    // needs no work; anything else re-sorts on the set.
+                    if self.sort != (SortKey::Rank, true) {
+                        self.sort_dirty = true;
+                    } else {
+                        self.order = (0..self.results.len() as u32).collect();
+                    }
+                }
+            }
+            SearchUpdate::Completed { limited, .. } => {
+                self.running = false;
+                self.elapsed = self.search_started.map(|t| t.elapsed());
+                self.limited |= limited;
+            }
+            SearchUpdate::Error { message, .. } => {
+                self.running = false;
+                self.elapsed = self.search_started.map(|t| t.elapsed());
+                self.error = Some(message);
+            }
+        }
+    }
+
+    pub fn result_count_label(&self) -> Option<String> {
+        if self.query.trim().is_empty() && self.results.is_empty() {
+            return None;
+        }
+        Some(if self.limited {
+            format!("{}+ results (truncated)", self.results.len())
+        } else {
+            format!("{} results", self.results.len())
+        })
+    }
+
+    fn resort(&mut self) {
+        let (key, ascending) = self.sort;
+        let selected_id = self
+            .selected
+            .and_then(|i| self.results.get(i as usize))
+            .map(|h| h.file_id);
+        self.order = (0..self.results.len() as u32).collect();
+        let results = &self.results;
+        self.order.sort_by(|&a, &b| {
+            let (a, b) = (&results[a as usize], &results[b as usize]);
+            let ord = match key {
+                SortKey::Rank => a.rank.partial_cmp(&b.rank).unwrap_or(std::cmp::Ordering::Equal),
+                SortKey::Name => a.name.cmp(&b.name),
+                SortKey::Path => a.path.cmp(&b.path),
+                SortKey::Size => a.size.cmp(&b.size),
+                SortKey::Modified => a.mtime.cmp(&b.mtime),
+            };
+            if ascending {
+                ord
+            } else {
+                ord.reverse()
+            }
+        });
+        // Selection follows the file, not the visual slot.
+        self.selected = selected_id.and_then(|id| {
+            self.results
+                .iter()
+                .position(|h| h.file_id == id)
+                .map(|i| i as u32)
+        });
+        self.sort_dirty = false;
+    }
+
+    /// A sortable column header: the whole cell is the click target, the
+    /// label is centered, and the sort indicator is a painter-drawn
+    /// triangle (the default egui fonts have no ▲/▼ glyphs — they render
+    /// as boxes).
+    fn sort_header(&mut self, ui: &mut egui::Ui, key: SortKey, label: &str) {
+        let (cur, asc) = self.sort;
+        let selected = cur == key;
+        let (rect, response) =
+            ui.allocate_exact_size(ui.available_size(), egui::Sense::click());
+        if ui.is_rect_visible(rect) {
+            if response.hovered() {
+                ui.painter()
+                    .rect_filled(rect, 2.0, ui.visuals().widgets.hovered.weak_bg_fill);
+            }
+            let font_id = egui::TextStyle::Body.resolve(ui.style());
+            let color = ui.visuals().strong_text_color();
+            let galley = ui
+                .painter()
+                .layout_no_wrap(label.to_string(), font_id, color);
+            let text_size = galley.size();
+            let arrow_space = if selected { 11.0 } else { 0.0 };
+            let text_pos = egui::pos2(
+                rect.center().x - (text_size.x + arrow_space) / 2.0,
+                rect.center().y - text_size.y / 2.0,
+            );
+            ui.painter().galley(text_pos, galley, color);
+            if selected {
+                let cx = text_pos.x + text_size.x + 7.0;
+                let cy = rect.center().y;
+                let (w, h) = (3.5, 3.0);
+                let points = if asc {
+                    vec![
+                        egui::pos2(cx, cy - h),
+                        egui::pos2(cx - w, cy + h),
+                        egui::pos2(cx + w, cy + h),
+                    ]
+                } else {
+                    vec![
+                        egui::pos2(cx, cy + h),
+                        egui::pos2(cx - w, cy - h),
+                        egui::pos2(cx + w, cy - h),
+                    ]
+                };
+                ui.painter().add(egui::Shape::convex_polygon(
+                    points,
+                    color,
+                    egui::Stroke::NONE,
+                ));
+            }
+        }
+        if response.clicked() {
+            self.sort = if selected { (key, !asc) } else { (key, true) };
+            self.sort_dirty = true;
+        }
+    }
+
+    pub fn ui(&mut self, ui: &mut egui::Ui) -> SearchActions {
+        let mut actions = SearchActions::default();
+
+        // --- Query strip -------------------------------------------------
+        ui.horizontal(|ui| {
+            let show_elapsed =
+                !self.running && self.elapsed.is_some() && !self.query.trim().is_empty();
+            let slot_room = if self.running {
+                24.0
+            } else if show_elapsed {
+                60.0
+            } else {
+                0.0
+            };
+            let width = ui.available_width() - 170.0 - slot_room;
+            let highlight = &mut self.highlight;
+            let mut layouter = move |ui: &egui::Ui, buf: &dyn egui::TextBuffer, _wrap: f32| {
+                crate::query_highlight::galley(ui, highlight, buf.as_str())
+            };
+            let response = ui.add(
+                egui::TextEdit::singleline(&mut self.query)
+                    .desired_width(width.max(120.0))
+                    .hint_text("Search names and contents…  (type:Document regex:… budget*)")
+                    .layouter(&mut layouter),
+            );
+            if self.focus_query {
+                response.request_focus();
+                self.focus_query = false;
+            }
+            if response.changed() {
+                self.pending_edit = Some(Instant::now());
+            }
+            // One slot right of the box: spinner while searching, then the
+            // total wall time of all cascade passes once it lands.
+            if self.running {
+                ui.add(egui::Spinner::new().size(16.0));
+            } else if show_elapsed {
+                if let Some(elapsed) = self.elapsed {
+                    ui.label(egui::RichText::new(fmt_elapsed(elapsed)).small().weak())
+                        .on_hover_text("Time to run all search passes");
+                }
+            }
+            if ui
+                .checkbox(&mut self.fuzzy, "Fuzzy")
+                .on_hover_text("Also run fuzzy filename and full-text passes (slower)")
+                .changed()
+            {
+                actions.save_fuzzy_default = Some(self.fuzzy);
+                actions.rerun = true;
+            }
+            if ui.button("?").on_hover_text("Query syntax help").clicked() {
+                self.help_open = !self.help_open;
+            }
+        });
+
+        // Session ignore chips.
+        if !self.session_ignores.is_empty() {
+            ui.horizontal_wrapped(|ui| {
+                ui.label(egui::RichText::new("Ignoring:").small().weak());
+                let mut remove: Option<usize> = None;
+                for (i, pattern) in self.session_ignores.iter().enumerate() {
+                    if ui
+                        .small_button(format!("{} ✕", pattern))
+                        .on_hover_text("Remove this session filter")
+                        .clicked()
+                    {
+                        remove = Some(i);
+                    }
+                }
+                if let Some(i) = remove {
+                    self.session_ignores.remove(i);
+                    actions.rerun = true;
+                }
+            });
+        }
+
+        // Notices.
+        if let Some(err) = &self.error {
+            ui.colored_label(ui.visuals().error_fg_color, err);
+        } else if self.limited {
+            ui.label(
+                egui::RichText::new(format!(
+                    "Showing first {} matches; refine the query (limit configurable in Options).",
+                    self.results.len()
+                ))
+                .small()
+                .weak(),
+            );
+        } else if !self.running
+            && !self.swap_pending
+            && self.results.is_empty()
+            && !self.query.trim().is_empty()
+            && self.error.is_none()
+        {
+            ui.label(egui::RichText::new("No results.").small().weak());
+        }
+
+        // Result-set transitions pulse instead of strobing: the old table
+        // fades out over 0.25 s while the new hits stage, the sets swap at
+        // zero opacity, and the new table fades back in over 0.25 s.
+        // `animate_value_with_time` keeps requesting repaints until the
+        // value settles.
+        let fade_target = if self.swap_pending { 0.0 } else { 1.0 };
+        let fade = ui.ctx().animate_value_with_time(
+            egui::Id::new("qs-results-fade"),
+            fade_target,
+            0.25,
+        );
+        if self.swap_pending && fade <= 0.01 {
+            self.results = std::mem::take(&mut self.staging);
+            self.has_snippets = self.staging_has_snippets;
+            self.selected = None;
+            self.swap_pending = false;
+            if self.sort == (SortKey::Rank, true) {
+                self.order = (0..self.results.len() as u32).collect();
+            } else {
+                self.sort_dirty = true;
+            }
+        }
+
+        if self.sort_dirty {
+            self.resort();
+        }
+
+        // Fade covers the table and the preview strip below it; the modal
+        // windows and notices render at full opacity on their own layers.
+        ui.set_opacity(fade);
+
+        // --- Results table ------------------------------------------------
+        // Reserve room for the selected-row snippet preview strip. Only
+        // content matches get one — a filename match's "snippet" is the
+        // name, already on screen.
+        let preview_snippet: Option<Snippet> = self
+            .selected
+            .and_then(|i| self.results.get(i as usize))
+            .filter(|h| matches!(h.stage, 5 | 6 | 8))
+            .and_then(|h| h.snippet.clone());
+        let preview_height = if preview_snippet.is_some() { 44.0 } else { 0.0 };
+        let table_height = (ui.available_height() - preview_height).max(60.0);
+
+        let text_height = egui::TextStyle::Body.resolve(ui.style()).size + 4.0;
+        let mut open_ignore_dialog: Option<usize> = None;
+
+        ui.push_id("results", |ui| {
+            let mut table = TableBuilder::new(ui)
+                .striped(true)
+                .resizable(true)
+                .sense(egui::Sense::click())
+                .max_scroll_height(table_height)
+                .min_scrolled_height(60.0)
+                .column(Column::initial(220.0).at_least(80.0).clip(true)) // name
+                .column(Column::remainder().at_least(120.0).clip(true)); // path
+            if self.has_snippets {
+                table = table.column(Column::remainder().at_least(120.0).clip(true));
+            }
+            table = table
+                .column(Column::exact(72.0)) // size
+                .column(Column::exact(110.0)) // modified
+                .column(Column::exact(52.0)); // rank
+
+            table
+                .header(text_height + 4.0, |mut header| {
+                    header.col(|ui| self.sort_header(ui, SortKey::Name, "Name"));
+                    header.col(|ui| self.sort_header(ui, SortKey::Path, "Path"));
+                    if self.has_snippets {
+                        header.col(|ui| {
+                            ui.with_layout(
+                                egui::Layout::centered_and_justified(
+                                    egui::Direction::LeftToRight,
+                                ),
+                                |ui| {
+                                    ui.label(egui::RichText::new("Match").strong());
+                                },
+                            );
+                        });
+                    }
+                    header.col(|ui| self.sort_header(ui, SortKey::Size, "Size"));
+                    header.col(|ui| self.sort_header(ui, SortKey::Modified, "Modified"));
+                    header.col(|ui| self.sort_header(ui, SortKey::Rank, "Rank"));
+                })
+                .body(|body| {
+                    let order = self.order.clone();
+                    body.rows(text_height, order.len(), |mut row| {
+                        let result_ix = order[row.index()] as usize;
+                        let hit = &self.results[result_ix];
+                        row.set_selected(self.selected == Some(result_ix as u32));
+
+                        row.col(|ui| {
+                            ui.label(&hit.name);
+                        });
+                        row.col(|ui| {
+                            ui.label(egui::RichText::new(&hit.path).weak());
+                        });
+                        if self.has_snippets {
+                            let snippet = hit.snippet.clone();
+                            // Name and path matches show a whole field, so
+                            // they render bracketed: [matched field].
+                            let whole_field =
+                                hit.stage <= 4 || hit.stage == 7 || hit.stage >= 9;
+                            row.col(|ui| {
+                                if let Some(snip) = &snippet {
+                                    let width = ui.available_width();
+                                    let job = centered_match_job(ui, snip, width, whole_field);
+                                    let response = ui
+                                        .with_layout(
+                                            egui::Layout::centered_and_justified(
+                                                egui::Direction::LeftToRight,
+                                            ),
+                                            |ui| ui.label(job),
+                                        )
+                                        .inner;
+                                    if !snip.ranges.is_empty() {
+                                        let hover = snip.clone();
+                                        response.on_hover_ui(|ui| {
+                                            ui.set_max_width(520.0);
+                                            let job = snippet_job(ui, &hover, 10);
+                                            ui.label(job);
+                                        });
+                                    }
+                                }
+                            });
+                        }
+                        row.col(|ui| {
+                            ui.with_layout(
+                                egui::Layout::centered_and_justified(
+                                    egui::Direction::LeftToRight,
+                                ),
+                                |ui| {
+                                    ui.label(human_size(hit.size));
+                                },
+                            );
+                        });
+                        row.col(|ui| {
+                            let color = recency_color(ui, hit.mtime);
+                            ui.with_layout(
+                                egui::Layout::centered_and_justified(
+                                    egui::Direction::LeftToRight,
+                                ),
+                                |ui| {
+                                    ui.label(
+                                        egui::RichText::new(fmt_mtime(hit.mtime)).color(color),
+                                    );
+                                },
+                            );
+                        });
+                        row.col(|ui| {
+                            ui.with_layout(
+                                egui::Layout::centered_and_justified(
+                                    egui::Direction::LeftToRight,
+                                ),
+                                |ui| {
+                                    ui.label(
+                                        egui::RichText::new(format!(" {:.2} ", hit.rank))
+                                            .background_color(rank_tier_color(hit.stage))
+                                            .color(egui::Color32::from_rgb(32, 32, 32)),
+                                    );
+                                },
+                            );
+                        });
+
+                        let response = row.response();
+                        if response.clicked() {
+                            self.selected = Some(result_ix as u32);
+                        }
+                        if response.double_clicked() {
+                            platform::open_file(&self.results[result_ix].path);
+                        }
+                        response.context_menu(|ui| {
+                            let path = self.results[result_ix].path.clone();
+                            if ui.button("Open").clicked() {
+                                platform::open_file(&path);
+                                ui.close();
+                            }
+                            if ui.button("Open containing folder").clicked() {
+                                platform::reveal_in_folder(&path);
+                                ui.close();
+                            }
+                            ui.separator();
+                            if ui.button("Build ignore filter…").clicked() {
+                                open_ignore_dialog = Some(result_ix);
+                                ui.close();
+                            }
+                        });
+                    });
+                });
+        });
+
+        if let Some(ix) = open_ignore_dialog {
+            let hit = &self.results[ix];
+            self.ignore_dialog = Some(IgnoreDialog {
+                source_name: hit.name.clone(),
+                source_path: hit.path.clone(),
+                pattern: hit.name.clone(),
+                persist: false,
+            });
+        }
+
+        // Selected-row preview strip: the full snippet, wrapped.
+        if let Some(snip) = &preview_snippet {
+            ui.separator();
+            let job = snippet_job(ui, snip, 2);
+            ui.label(job);
+        }
+
+        self.ignore_dialog_ui(ui.ctx(), &mut actions);
+        self.help_window_ui(ui.ctx());
+        actions
+    }
+
+    fn ignore_dialog_ui(&mut self, ctx: &egui::Context, actions: &mut SearchActions) {
+        let Some(dialog) = &mut self.ignore_dialog else {
+            return;
+        };
+        let mut apply = false;
+        let mut cancel = false;
+        egui::Window::new("Ignore filter")
+            .collapsible(false)
+            .resizable(false)
+            .anchor(egui::Align2::CENTER_CENTER, [0.0, 0.0])
+            .show(ctx, |ui| {
+                ui.label(format!("From: {}", dialog.source_path));
+                ui.add_space(4.0);
+                ui.horizontal(|ui| {
+                    if ui.button("This name").clicked() {
+                        dialog.pattern = dialog.source_name.clone();
+                    }
+                    if let Some(ext) = std::path::Path::new(&dialog.source_name)
+                        .extension()
+                        .and_then(|e| e.to_str())
+                    {
+                        if ui.button(format!("*.{}", ext)).clicked() {
+                            dialog.pattern = format!("*.{}", ext);
+                        }
+                    }
+                    if let Some(parent) = std::path::Path::new(&dialog.source_path)
+                        .parent()
+                        .and_then(|p| p.to_str())
+                    {
+                        if ui.button("This directory").clicked() {
+                            dialog.pattern = format!("{}/*", parent);
+                        }
+                    }
+                });
+                ui.add(
+                    egui::TextEdit::singleline(&mut dialog.pattern)
+                        .desired_width(360.0)
+                        .hint_text("glob pattern"),
+                );
+                ui.checkbox(&mut dialog.persist, "Persist to config");
+                ui.label(
+                    egui::RichText::new(
+                        "Session filters hide results immediately. Persisted filters also \
+                         exclude files from the index at the next reindex.",
+                    )
+                    .small()
+                    .weak(),
+                );
+                ui.horizontal(|ui| {
+                    if ui.button("Apply").clicked() {
+                        apply = true;
+                    }
+                    if ui.button("Cancel").clicked() {
+                        cancel = true;
+                    }
+                });
+            });
+        if apply {
+            let dialog = self.ignore_dialog.take().unwrap();
+            let pattern = dialog.pattern.trim().to_string();
+            if !pattern.is_empty() {
+                if !self.session_ignores.contains(&pattern) {
+                    self.session_ignores.push(pattern.clone());
+                }
+                if dialog.persist {
+                    actions.persist_ignore = Some(pattern);
+                }
+                actions.rerun = true;
+            }
+        } else if cancel {
+            self.ignore_dialog = None;
+        }
+    }
+
+    fn help_window_ui(&mut self, ctx: &egui::Context) {
+        let mut open = self.help_open;
+        egui::Window::new("Query syntax")
+            .open(&mut open)
+            .resizable(false)
+            .default_width(540.0)
+            .show(ctx, |ui| {
+                ui.label(
+                    "Everything that is not a filter is matched as one phrase, in order. \
+                     Filters combine freely with the search text.",
+                );
+                ui.add_space(6.0);
+                egui::Grid::new("query-syntax-table")
+                    .num_columns(2)
+                    .spacing([18.0, 5.0])
+                    .striped(true)
+                    .show(ui, |ui| {
+                        let row = |ui: &mut egui::Ui, syntax: &str, meaning: &str| {
+                            ui.monospace(syntax);
+                            ui.label(meaning);
+                            ui.end_row();
+                        };
+                        row(
+                            ui,
+                            "budget report",
+                            "names, contents, and paths containing \"budget report\"",
+                        );
+                        row(
+                            ui,
+                            "\"exact phrase\"",
+                            "quotes keep spaces, stars, and filter-like words literal",
+                        );
+                        row(
+                            ui,
+                            "bud*report",
+                            "* matches any run of characters (within a line); \
+                             also works in name: values",
+                        );
+                        row(
+                            ui,
+                            "regex:\"(foo|bar)\\d+\"",
+                            "regular expression, matched against names, contents, \
+                             and paths; case-insensitive — use (?-i:…) to override; \
+                             quote patterns containing spaces",
+                        );
+                        row(
+                            ui,
+                            "type:Document",
+                            "file class: Audio, Image, Video, Document, Text, \
+                             Archive, Spreadsheet, Presentation, Folder",
+                        );
+                        row(
+                            ui,
+                            "modified:>=2024-01-01",
+                            "modification date (yyyy-mm-dd); also <, <=, > and =",
+                        );
+                        row(
+                            ui,
+                            "path:/home/me/docs",
+                            "only results in that folder and its subfolders; \
+                             quote paths containing spaces",
+                        );
+                        row(ui, "mime:application/pdf", "exact MIME type");
+                        row(
+                            ui,
+                            "name:report",
+                            "filename contains, applied as an unranked filter",
+                        );
+                    });
+                ui.add_space(6.0);
+                ui.label("Example:");
+                ui.monospace("type:Document modified:>=2024-01-01 quarterly budget");
+                ui.add_space(6.0);
+                ui.label(
+                    egui::RichText::new(
+                        "Ranking: exact filename matches, then filename substrings, then \
+                         full-text matches (ordered by occurrences), then fuzzy matches \
+                         when enabled, and finally matches on the rest of the file path.",
+                    )
+                    .small()
+                    .weak(),
+                );
+                ui.label(
+                    egui::RichText::new(
+                        "The complete reference, including ranking details and the \
+                         fuzzy edit budget, is the \"Query syntax\" section of \
+                         README.md in the QuickSearch folder.",
+                    )
+                    .small()
+                    .weak(),
+                );
+            });
+        self.help_open = open;
+    }
+}
+
+struct SnippetFormats {
+    normal: TextFormat,
+    highlight: TextFormat,
+    weak: TextFormat,
+}
+
+fn snippet_formats(ui: &egui::Ui) -> SnippetFormats {
+    let font_id = egui::TextStyle::Body.resolve(ui.style());
+    SnippetFormats {
+        normal: TextFormat {
+            font_id: font_id.clone(),
+            color: ui.visuals().text_color(),
+            ..Default::default()
+        },
+        highlight: TextFormat {
+            font_id: font_id.clone(),
+            color: ui.visuals().strong_text_color(),
+            background: ui.visuals().selection.bg_fill.gamma_multiply(0.4),
+            ..Default::default()
+        },
+        weak: TextFormat {
+            font_id,
+            color: ui.visuals().weak_text_color(),
+            ..Default::default()
+        },
+    }
+}
+
+/// Build a highlighted snippet LayoutJob from byte ranges, wrapped to at
+/// most `max_rows`. Cheap enough to run per visible row per frame.
+fn snippet_job(ui: &egui::Ui, snip: &Snippet, max_rows: usize) -> LayoutJob {
+    let fmt = snippet_formats(ui);
+    let mut job = LayoutJob::default();
+    job.wrap.max_rows = max_rows;
+    if max_rows == 1 {
+        job.wrap.break_anywhere = true;
+    }
+    if snip.truncated_start {
+        job.append("… ", 0.0, fmt.weak.clone());
+    }
+    let mut cursor = 0;
+    for &(start, end) in &snip.ranges {
+        if start > cursor {
+            job.append(&snip.window[cursor..start], 0.0, fmt.normal.clone());
+        }
+        job.append(&snip.window[start..end], 0.0, fmt.highlight.clone());
+        cursor = end;
+    }
+    if cursor < snip.window.len() {
+        job.append(&snip.window[cursor..], 0.0, fmt.normal.clone());
+    }
+    if snip.truncated_end {
+        job.append(" …", 0.0, fmt.weak);
+    }
+    job
+}
+
+/// The Match column cell: one line with the (first) matched span centered
+/// and an equal amount of context on both sides, trimmed to what fits the
+/// column width. Matches on a whole field — a filename or a path — are
+/// wrapped in brackets: `[name]`.
+fn centered_match_job(
+    ui: &egui::Ui,
+    snip: &Snippet,
+    width_px: f32,
+    whole_field: bool,
+) -> LayoutJob {
+    let fmt = snippet_formats(ui);
+    let font_id = egui::TextStyle::Body.resolve(ui.style());
+    let char_width = ui.fonts(|f| f.glyph_width(&font_id, '0')).max(1.0);
+    let mut budget = ((width_px / char_width) as usize).saturating_sub(2).max(8);
+    if whole_field {
+        budget = budget.saturating_sub(2); // room for the brackets
+    }
+
+    // Newlines force line breaks even in a one-row LayoutJob, wrecking the
+    // centered single-line cell. Flatten them to spaces — a byte-for-byte
+    // ASCII replacement, so the match ranges stay valid. The mouseover
+    // renders the original window untouched.
+    let flattened = snip.window.replace(['\n', '\r', '\t'], " ");
+    let window = flattened.as_str();
+    let (start, end) = match snip.ranges.first().copied() {
+        Some((a, b)) => {
+            let match_chars = window[a..b].chars().count();
+            let side = budget.saturating_sub(match_chars) / 2;
+            let before = &window[..a];
+            let after = &window[b..];
+            let before_count = before.chars().count();
+            let after_count = after.chars().count();
+            // Equal context on both sides; leftover budget from a short
+            // side flows to the other.
+            let take_before = (side + side.saturating_sub(after_count)).min(before_count);
+            let take_after = (side + side.saturating_sub(before_count)).min(after_count);
+            let start = if take_before == 0 {
+                a
+            } else {
+                before
+                    .char_indices()
+                    .nth_back(take_before - 1)
+                    .map(|(i, _)| i)
+                    .unwrap_or(0)
+            };
+            let end = b + after
+                .char_indices()
+                .nth(take_after)
+                .map(|(i, _)| i)
+                .unwrap_or(after.len());
+            (start, end)
+        }
+        None => {
+            // No ranges (shouldn't happen for match cells) — head trim.
+            let end = window
+                .char_indices()
+                .nth(budget)
+                .map(|(i, _)| i)
+                .unwrap_or(window.len());
+            (0, end)
+        }
+    };
+
+    let mut job = LayoutJob::default();
+    job.wrap.max_rows = 1;
+    job.wrap.break_anywhere = true;
+    if whole_field {
+        job.append("[", 0.0, fmt.weak.clone());
+    }
+    if start > 0 || snip.truncated_start {
+        job.append("…", 0.0, fmt.weak.clone());
+    }
+    let mut cursor = start;
+    for &(a, b) in &snip.ranges {
+        let (a, b) = (a.max(start), b.min(end));
+        if a >= b || a >= end {
+            continue;
+        }
+        if a > cursor {
+            job.append(&window[cursor..a], 0.0, fmt.normal.clone());
+        }
+        job.append(&window[a..b], 0.0, fmt.highlight.clone());
+        cursor = b;
+    }
+    if cursor < end {
+        job.append(&window[cursor..end], 0.0, fmt.normal.clone());
+    }
+    if end < window.len() || snip.truncated_end {
+        job.append("…", 0.0, fmt.weak.clone());
+    }
+    if whole_field {
+        job.append("]", 0.0, fmt.weak);
+    }
+    job
+}
+
+/// Tier-list chip color per cascade stage — lower rank, higher tier:
+/// S-red for exact case-sensitive filename matches down through the
+/// pastel ramp to purple for fuzzy full-text and on to the grey path
+/// tiers. Dark text on these pastels stays readable in both themes.
+fn rank_tier_color(stage: u8) -> egui::Color32 {
+    match stage {
+        1 => egui::Color32::from_rgb(255, 127, 127), // S
+        2 => egui::Color32::from_rgb(255, 191, 127), // A
+        3 => egui::Color32::from_rgb(255, 223, 127), // B
+        4 => egui::Color32::from_rgb(255, 255, 127), // C
+        5 => egui::Color32::from_rgb(191, 255, 127), // D
+        6 => egui::Color32::from_rgb(127, 255, 127), // E
+        7 => egui::Color32::from_rgb(127, 191, 255), // F
+        8 => egui::Color32::from_rgb(191, 127, 255), // G
+        9 => egui::Color32::from_rgb(223, 159, 255), // H — path, exact case
+        10 => egui::Color32::from_rgb(239, 191, 239), // I — path, any case
+        _ => egui::Color32::from_rgb(199, 199, 199), // J — fuzzy path
+    }
+}
+
+/// Timestamp color: fresh files get a green tint that fades into the weak
+/// text color over ~2 years on a log scale.
+fn recency_color(ui: &egui::Ui, mtime: i64) -> egui::Color32 {
+    let now = std::time::SystemTime::now()
+        .duration_since(std::time::UNIX_EPOCH)
+        .map(|d| d.as_secs() as i64)
+        .unwrap_or(0);
+    let age_hours = ((now - mtime).max(0) as f32 / 3600.0).max(1.0);
+    const HORIZON_HOURS: f32 = 24.0 * 365.0 * 2.0;
+    let t = (age_hours.ln() / HORIZON_HOURS.ln()).clamp(0.0, 1.0);
+    let fresh = egui::Color32::from_rgb(87, 187, 122);
+    let old = ui.visuals().weak_text_color();
+    let lerp = |a: u8, b: u8| (a as f32 + (b as f32 - a as f32) * t).round() as u8;
+    egui::Color32::from_rgb(
+        lerp(fresh.r(), old.r()),
+        lerp(fresh.g(), old.g()),
+        lerp(fresh.b(), old.b()),
+    )
+}

crates/quicksearch-gui/src/tracker.rs

@@ -0,0 +1,144 @@
+//! Indexing-rate estimation for the status displays.
+//!
+//! The old tracker sampled the counter every poll tick but pruned to a
+//! 1-second window, so anything slower than ~1 file/sec measured a
+//! genuine zero and displayed "0.0 files/sec" despite progress. This one
+//! records a point only when the counter *changes*, keeps up to 60 s of
+//! history but never fewer than two points (so slow rates stay
+//! computable), and measures against `now` so the estimate decays during
+//! stalls instead of freezing at the last burst.
+
+use std::time::{Duration, Instant};
+
+const HISTORY: Duration = Duration::from_secs(60);
+
+pub struct SpeedTracker {
+    /// (when, counter value) — appended only on counter change.
+    points: Vec<(Instant, usize)>,
+}
+
+impl SpeedTracker {
+    pub fn new() -> SpeedTracker {
+        SpeedTracker { points: Vec::new() }
+    }
+
+    /// Reset between phases (each phase restarts its counter).
+    pub fn reset(&mut self) {
+        self.points.clear();
+    }
+
+    pub fn record(&mut self, files_processed: usize) {
+        self.record_at(Instant::now(), files_processed);
+    }
+
+    fn record_at(&mut self, now: Instant, files_processed: usize) {
+        match self.points.last() {
+            Some(&(_, last)) if last == files_processed => return,
+            // Counter went backwards — a new phase started without an
+            // explicit reset.
+            Some(&(_, last)) if files_processed < last => self.points.clear(),
+            _ => {}
+        }
+        self.points.push((now, files_processed));
+        // Prune old points, but always keep at least two so a slow but
+        // steady rate never becomes unmeasurable.
+        while self.points.len() > 2
+            && now.duration_since(self.points[0].0) > HISTORY
+        {
+            self.points.remove(0);
+        }
+    }
+
+    /// Estimated files/sec, measured from the oldest retained progress
+    /// point to *now*. `None` until two data points exist.
+    pub fn files_per_sec(&self) -> Option<f64> {
+        self.files_per_sec_at(Instant::now())
+    }
+
+    fn files_per_sec_at(&self, now: Instant) -> Option<f64> {
+        let (t0, c0) = *self.points.first()?;
+        let (_, c1) = *self.points.last()?;
+        if self.points.len() < 2 {
+            return None;
+        }
+        let span = now.duration_since(t0).as_secs_f64();
+        if span <= 0.0 {
+            return None;
+        }
+        Some((c1 - c0) as f64 / span)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn needs_two_points() {
+        let mut t = SpeedTracker::new();
+        let base = Instant::now();
+        assert!(t.files_per_sec_at(base).is_none());
+        t.record_at(base, 10);
+        assert!(t.files_per_sec_at(base).is_none());
+    }
+
+    #[test]
+    fn slow_rate_is_measurable_not_zero() {
+        // One file every 2.5 s — the old 1 s window reported 0.0 here.
+        let mut t = SpeedTracker::new();
+        let base = Instant::now();
+        for i in 0..4 {
+            t.record_at(base + Duration::from_millis(2500 * i), 10 + i as usize);
+        }
+        let rate = t.files_per_sec_at(base + Duration::from_millis(7500)).unwrap();
+        assert!((rate - 0.4).abs() < 0.01, "expected ~0.4/s, got {}", rate);
+    }
+
+    #[test]
+    fn unchanged_counter_adds_no_points() {
+        let mut t = SpeedTracker::new();
+        let base = Instant::now();
+        for i in 0..100 {
+            t.record_at(base + Duration::from_millis(50 * i), 42);
+        }
+        assert_eq!(t.points.len(), 1, "only the first observation recorded");
+    }
+
+    #[test]
+    fn stall_decays_toward_zero() {
+        let mut t = SpeedTracker::new();
+        let base = Instant::now();
+        t.record_at(base, 0);
+        t.record_at(base + Duration::from_secs(1), 100); // 100/s burst
+        let just_after = t.files_per_sec_at(base + Duration::from_secs(1)).unwrap();
+        let stalled = t.files_per_sec_at(base + Duration::from_secs(20)).unwrap();
+        assert!(just_after > 90.0);
+        assert!(stalled < 6.0, "estimate must decay during a stall: {}", stalled);
+    }
+
+    #[test]
+    fn pruning_keeps_at_least_two_points() {
+        let mut t = SpeedTracker::new();
+        let base = Instant::now();
+        t.record_at(base, 1);
+        t.record_at(base + Duration::from_secs(30), 2);
+        // Far beyond the history window; both points are older than 60 s
+        // relative to this record.
+        t.record_at(base + Duration::from_secs(300), 3);
+        assert!(t.points.len() >= 2);
+        assert!(t
+            .files_per_sec_at(base + Duration::from_secs(300))
+            .is_some());
+    }
+
+    #[test]
+    fn counter_regression_resets() {
+        let mut t = SpeedTracker::new();
+        let base = Instant::now();
+        t.record_at(base, 500);
+        t.record_at(base + Duration::from_secs(1), 600);
+        // New phase restarts from a small number.
+        t.record_at(base + Duration::from_secs(2), 3);
+        assert_eq!(t.points.len(), 1);
+    }
+}

packaging/build-deb.sh

@@ -0,0 +1,224 @@
+#!/usr/bin/env bash
+#
+# Build a Debian package for QuickSearch.
+#
+#   ./packaging/build-deb.sh                 build and package
+#   ./packaging/build-deb.sh --no-build      package an existing release binary
+#   ./packaging/build-deb.sh --no-strip      keep debug symbols (25 MB vs 20 MB)
+#   ./packaging/build-deb.sh -o /tmp/out     write the .deb somewhere else
+#
+# Environment: DEB_REVISION (default 1), DEB_MAINTAINER, SOURCE_DATE_EPOCH.
+#
+# Deliberately does not use cargo-deb, debhelper, fakeroot or an SVG rasteriser:
+# dpkg-deb and desktop-file-utils are the only tools required, and both are part
+# of a standard Debian or Ubuntu install.
+
+set -euo pipefail
+# Directories created along the way must be 0755, not whatever the caller's
+# umask happens to be, or the package ships group-writable directories.
+umask 022
+
+readonly PKG=quicksearch
+# The GUI binary and the console-subsystem terminal binary. Both ship: on Unix
+# `quicksearch` does both jobs, but the README and the shared man page name
+# `quicksearch-cli` too, so it has to exist wherever the docs are installed.
+readonly BINARIES=(quicksearch quicksearch-cli)
+readonly REPO_ROOT="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")/.." && pwd)"
+readonly ICON_SRC="$REPO_ROOT/crates/quicksearch-gui/assets/icons"
+readonly ICON_SVG="$ICON_SRC/quicksearch_icon.svg"
+
+do_build=1
+do_strip=1
+out_dir="$REPO_ROOT/dist"
+
+die() { printf 'build-deb: %s\n' "$*" >&2; exit 1; }
+say() { printf '\033[1m==>\033[0m %s\n' "$*"; }
+
+while [ $# -gt 0 ]; do
+    case "$1" in
+        --no-build) do_build=0 ;;
+        --no-strip) do_strip=0 ;;
+        -o|--output-dir) shift; [ $# -gt 0 ] || die "--output-dir needs a path"; out_dir="$1" ;;
+        # Print the header comment block, however long it grows.
+        -h|--help) awk 'NR > 1 { if ($0 !~ /^#/) exit; sub(/^# ?/, ""); print }' "${BASH_SOURCE[0]}"; exit 0 ;;
+        *) die "unknown option: $1 (try --help)" ;;
+    esac
+    shift
+done
+
+for tool in dpkg-deb dpkg desktop-file-validate objdump gzip; do
+    command -v "$tool" >/dev/null 2>&1 || die "missing required tool: $tool"
+done
+[ "$do_strip" -eq 0 ] || command -v strip >/dev/null 2>&1 || die "missing strip (install binutils, or pass --no-strip)"
+
+# Version comes from [workspace.package] so the package can never drift from the
+# crate version.
+version="$(sed -n '/^\[workspace\.package\]/,/^\[/{ s/^version[[:space:]]*=[[:space:]]*"\([^"]*\)".*/\1/p }' "$REPO_ROOT/Cargo.toml")"
+[ -n "$version" ] || die "could not read version from Cargo.toml"
+
+revision="${DEB_REVISION:-1}"
+maintainer="${DEB_MAINTAINER:-Jeremy <jeremy@karsttech.com>}"
+arch="$(dpkg --print-architecture)"
+deb_version="${version}-${revision}"
+stage="$out_dir/${PKG}_${deb_version}_${arch}"
+deb="$out_dir/${PKG}_${deb_version}_${arch}.deb"
+
+# ---------------------------------------------------------------- build ----
+
+if [ "$do_build" -eq 1 ]; then
+    say "Building quicksearch $version (release)"
+    ( cd "$REPO_ROOT" && cargo build --release -p quicksearch-gui )
+fi
+
+for bin in "${BINARIES[@]}"; do
+    [ -x "$REPO_ROOT/target/release/$bin" ] \
+        || die "no release binary at target/release/$bin (drop --no-build?)"
+done
+# Both binaries link the same crates, so either gives the same glibc floor.
+readonly primary_binary="$REPO_ROOT/target/release/$PKG"
+[ -f "$ICON_SVG" ] || die "no icon at $ICON_SVG"
+
+say "Validating desktop entry"
+desktop-file-validate "$REPO_ROOT/packaging/$PKG.desktop"
+
+# --------------------------------------------------------------- stage -----
+
+say "Staging $stage"
+rm -rf "$stage"
+mkdir -p "$stage"
+
+for bin in "${BINARIES[@]}"; do
+    install -Dm755 "$REPO_ROOT/target/release/$bin"     "$stage/usr/bin/$bin"
+done
+install -Dm644 "$REPO_ROOT/packaging/$PKG.desktop"      "$stage/usr/share/applications/$PKG.desktop"
+install -Dm644 "$ICON_SVG"                              "$stage/usr/share/icons/hicolor/scalable/apps/$PKG.svg"
+install -Dm644 "$REPO_ROOT/packaging/copyright"         "$stage/usr/share/doc/$PKG/copyright"
+install -Dm644 "$REPO_ROOT/config_example.toml"         "$stage/usr/share/doc/$PKG/config_example.toml"
+install -Dm644 "$REPO_ROOT/README.md"                   "$stage/usr/share/doc/$PKG/README.md"
+
+# The PNGs are committed next to the quicksearch_icon.svg they came from, so
+# this script only copies them and an ordinary `cargo build` needs no image
+# tooling. To re-render them after editing the SVG, build a throwaway crate
+# (outside this workspace, to keep it out of Cargo.lock) depending on
+# `resvg = { version = "0.45", default-features = false }` and, for each size N,
+# parse with usvg::Tree::from_data, make a tiny_skia::Pixmap::new(N, N), call
+# resvg::render with Transform::from_scale(N/240.0, N/240.0) and save_png. The
+# SVG contains no <text>, so no font support is needed.
+#
+# Sizes follow whatever is committed, so adding one needs no script change.
+shopt -s nullglob
+icons=("$ICON_SRC"/$PKG-*.png)
+shopt -u nullglob
+[ "${#icons[@]}" -gt 0 ] || die "no icons in $ICON_SRC"
+for png in "${icons[@]}"; do
+    n="$(basename "$png" .png)"; n="${n#$PKG-}"
+    # The glob would also catch a non-size name like quicksearch-cli.png.
+    case "$n" in ''|*[!0-9]*) die "unexpected icon name: $(basename "$png")" ;; esac
+    install -Dm644 "$png" "$stage/usr/share/icons/hicolor/${n}x${n}/apps/$PKG.png"
+done
+say "Installed ${#icons[@]} icon sizes plus the scalable SVG"
+
+# Debian wants man pages and the changelog compressed, with no gzip timestamp so
+# repeat builds are byte-identical. quicksearch-cli.1 is a one-line .so stub
+# pointing at quicksearch.1, which documents both binaries.
+install -dm755 "$stage/usr/share/man/man1"
+for page in "${BINARIES[@]}"; do
+    gzip -9nc "$REPO_ROOT/packaging/$page.1" > "$stage/usr/share/man/man1/$page.1.gz"
+    chmod 644 "$stage/usr/share/man/man1/$page.1.gz"
+done
+
+if [ -n "${SOURCE_DATE_EPOCH:-}" ]; then
+    changelog_date="$(date -R -u -d "@$SOURCE_DATE_EPOCH")"
+else
+    changelog_date="$(date -R)"
+fi
+gzip -9nc <<EOF > "$stage/usr/share/doc/$PKG/changelog.Debian.gz"
+$PKG ($deb_version) unstable; urgency=medium
+
+  * Package build of $PKG $version.
+
+ -- $maintainer  $changelog_date
+EOF
+chmod 644 "$stage/usr/share/doc/$PKG/changelog.Debian.gz"
+
+# No config.toml is installed anywhere. Config::config_path() treats a
+# config.toml sitting next to the executable as portable mode and lets it
+# override the per-user config outright, so one in /usr/bin would hijack every
+# account on the machine. Config::load_from creates ~/.config/quicksearch/
+# config.toml on first run instead.
+
+if [ "$do_strip" -eq 1 ]; then
+    for bin in "${BINARIES[@]}"; do
+        before="$(du -h "$stage/usr/bin/$bin" | cut -f1)"
+        strip --strip-unneeded "$stage/usr/bin/$bin"
+        say "Stripped $bin: $before -> $(du -h "$stage/usr/bin/$bin" | cut -f1)"
+    done
+fi
+
+installed_size="$(du -ks "$stage" | cut -f1)"
+
+# ------------------------------------------------------------- control -----
+
+# The dynamic section only names libc, libgcc, libm and libbz2: winit and glutin
+# dlopen the entire display stack, so dpkg-shlibdeps cannot see it and the list
+# below is maintained by hand. Re-derive it with
+#   objdump -p target/release/quicksearch | grep NEEDED
+#   strings -a target/release/quicksearch | grep -oE 'lib[A-Za-z0-9_+-]+\.so(\.[0-9]+)*' | sort -u
+# and map each soname to a package with `dpkg -S`.
+glibc_min="$(objdump -T "$primary_binary" | sed -n 's/.*GLIBC_\([0-9][0-9.]*\).*/\1/p' | sort -V | tail -1)"
+[ -n "$glibc_min" ] || die "could not determine the glibc version requirement"
+
+depends="libc6 (>= ${glibc_min}), libgcc-s1 (>= 3.0), libbz2-1.0"
+depends="$depends, libx11-6, libxcb1, libxcursor1, libxi6, libxrender1"
+depends="$depends, libxkbcommon0, libxkbcommon-x11-0"
+depends="$depends, libwayland-client0, libwayland-egl1"
+depends="$depends, libegl1, libgl1"
+# Required by policy for anything installing into the hicolor theme; it also
+# provides the dpkg trigger that refreshes the icon cache on install.
+depends="$depends, hicolor-icon-theme"
+
+# desktop-file-utils owns the /usr/share/applications trigger; dbus-bin provides
+# the dbus-send used by "reveal in folder"; xdg-utils provides the xdg-open
+# fallback; xdg-desktop-portal backs the native folder picker. None are needed
+# to search, so none are hard dependencies.
+recommends="desktop-file-utils, xdg-utils, dbus-bin, xdg-desktop-portal"
+
+install -dm755 "$stage/DEBIAN"
+cat > "$stage/DEBIAN/control" <<EOF
+Package: $PKG
+Version: $deb_version
+Section: utils
+Priority: optional
+Architecture: $arch
+Maintainer: $maintainer
+Installed-Size: $installed_size
+Depends: $depends
+Recommends: $recommends
+Homepage: https://code.karsttech.com/jeremy/quick_search
+Description: fast full-text search across your files
+ QuickSearch keeps a SQLite/FTS5 index of the directories you choose and
+ searches them by both filename and file content. It extracts text from
+ documents, PDFs, archives and office files, watches the indexed paths for
+ changes and reindexes in the background while the application is open.
+ .
+ The same binary doubles as a terminal search tool: "quicksearch <terms>"
+ prints ranked results and exits without starting the indexer, the file
+ watcher or any background thread.
+EOF
+
+# No postinst/postrm: hicolor-icon-theme and desktop-file-utils declare
+# interest-noawait on /usr/share/icons/hicolor and /usr/share/applications, so
+# dpkg refreshes both caches on install and removal by itself.
+
+# --------------------------------------------------------------- build -----
+
+say "Building $deb"
+dpkg-deb --root-owner-group --build "$stage" "$deb" >/dev/null
+rm -rf "$stage"
+
+echo
+dpkg-deb --info "$deb"
+echo
+say "Done: $deb"
+echo "    install with:  sudo apt install $deb"
+echo "    inspect with:  dpkg-deb --contents $deb"

packaging/copyright

@@ -0,0 +1,61 @@
+Format: https://www.debian.org/doc/packaging-manuals/copyright-format/1.0/
+Upstream-Name: quicksearch
+Upstream-Contact: Jeremy <jeremy@karsttech.com>
+Source: https://code.karsttech.com/jeremy/quick_search.git
+
+Files: *
+Copyright: 2025-2026 Jeremy <jeremy@karsttech.com>
+License: GPL-3.0-or-later
+
+Files: crates/quicksearch-gui/assets/icons/*
+Copyright: 2026 Jeremy <jeremy@karsttech.com>
+License: GPL-3.0-or-later
+
+License: GPL-3.0-or-later
+ This program is free software: you can redistribute it and/or modify
+ it under the terms of the GNU General Public License as published by
+ the Free Software Foundation, either version 3 of the License, or
+ (at your option) any later version.
+ .
+ This program is distributed in the hope that it will be useful,
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ GNU General Public License for more details.
+ .
+ You should have received a copy of the GNU General Public License
+ along with this program.  If not, see <https://www.gnu.org/licenses/>.
+ .
+ On Debian systems the full text of the GNU General Public License
+ version 3 can be found in /usr/share/common-licenses/GPL-3.
+
+Comment:
+ /usr/bin/quicksearch and /usr/bin/quicksearch-cli are statically linked Rust
+ binaries built from the same crates. Their third-party components are not
+ shipped as separate files, so they are summarised here
+ rather than given individual stanzas. Every one of the 512 crates in
+ Cargo.lock is under a permissive licence, all of which are compatible with
+ GPL-3.0-or-later:
+ .
+   * MIT and/or Apache-2.0 (the large majority)
+   * Unicode-3.0, BSD-2-Clause, BSD-3-Clause, ISC, Zlib, 0BSD,
+     BSL-1.0, CC0-1.0, Unlicense
+ .
+ There are no GPL-only, LGPL-only, MPL or proprietary dependencies. Note
+ that a dozen crates (winit, glutin, ab_glyph, owned_ttf_parser and
+ related) are Apache-2.0 only; Apache-2.0 is compatible with GPL version 3
+ but not with GPL version 2, which is why this package is
+ GPL-3.0-or-later rather than GPL-2.0-or-later.
+ .
+ Bundled C sources compiled into the binary: SQLite (public domain, via
+ rusqlite), Zstandard (BSD-3-Clause arm of its BSD/GPL-2.0 dual licence,
+ via zstd-sys) and bzip2 (BSD-style, via bzip2-sys).
+ .
+ The binary also embeds egui's default fonts (crate epaint_default_fonts,
+ "(MIT OR Apache-2.0) AND OFL-1.1 AND Ubuntu-font-1.0"). The FSF regards
+ OFL-1.1 and the Ubuntu Font Licence as free but GPL-incompatible; they are
+ included here as font data rather than as linked program code, which is
+ the same basis on which Debian ships these fonts and other egui-based
+ applications.
+ .
+ Run `cargo metadata --all-features` against the source tree to reproduce
+ the per-crate licence list.

packaging/quicksearch-cli.1

@@ -0,0 +1,3 @@
+.\" quicksearch-cli is documented alongside quicksearch; this stub makes
+.\" `man quicksearch-cli` resolve to that page.
+.so man1/quicksearch.1

packaging/quicksearch.1

@@ -0,0 +1,136 @@
+.TH QUICKSEARCH 1 "2026-08-02" "quicksearch 0.1.0" "User Commands"
+.SH NAME
+quicksearch, quicksearch\-cli \- fast full\-text search across your files
+.SH SYNOPSIS
+.B quicksearch
+.br
+.B quicksearch
+.RI [ FLAGS ]
+.IR "query terms" ...
+.br
+.B quicksearch\-cli
+.RI [ FLAGS ]
+.IR "query terms" ...
+.SH DESCRIPTION
+.B quicksearch
+maintains a SQLite/FTS5 index of the paths listed in its configuration and
+searches it by filename and file content.
+.PP
+Invoked with no positional arguments it opens the desktop application, which
+owns the indexer: it walks the configured roots, watches them for changes and
+periodically reindexes.
+.PP
+Invoked with one or more query terms it runs the same ranked search cascade the
+application uses, prints the results and exits. Terminal mode never starts the
+indexer, the filesystem watcher or any background thread, so a one\-shot query
+neither leaves anything running nor competes for inotify watches with a running
+application. It requires an index that already exists; run the application once
+to build one.
+.PP
+.B quicksearch\-cli
+is terminal search and nothing else, and accepts the same flags and query
+syntax. On this platform it is equivalent to giving
+.B quicksearch
+a query, and exists so that scripts and documentation have one name that never
+opens a window. It matters on Windows, where the two cannot be one executable:
+a GUI built as a console application flashes a console window at every launch,
+and a console tool built as a GUI application cannot write back to the shell
+that invoked it. Given no query it prints usage and exits 2 rather than falling
+back to the application.
+.SH OPTIONS
+.TP
+.B \-\-fuzzy
+Also run the fuzzy filename and full\-text passes, which tolerate spelling
+differences at the cost of speed. The edit distance comes from
+.I [search].fuzzy_max_edits
+in the configuration.
+.TP
+.BI \-\-limit " N"
+Return at most
+.I N
+results. Defaults to
+.I [search].display_limit
+in the configuration.
+.B \-\-limit=N
+is also accepted.
+.TP
+.B \-\-long
+Print rank, size, modification time and a matching snippet for each hit instead
+of bare paths. Highlights the match in bold when stdout is a terminal.
+.TP
+.BR \-h ", " \-\-help
+Print usage and exit.
+.PP
+An unrecognised option given without any query terms is passed through and the
+application is opened, since it may be an option for the windowing backend.
+.SH QUERY SYNTAX
+Plain words form a single phrase. Filters may be combined with it:
+.TP
+.B type:\fIName\fR
+Match a file class, for example
+.IR type:Document ", " type:Image ", " type:Audio .
+.TP
+.B modified:\fIexpr\fR
+Compare against the modification date, for example
+.IR modified:>=2024-01-01 .
+.TP
+.B path:\fI/dir\fR
+Restrict results to a directory.
+.TP
+.B mime:\fItype\fR
+Match a MIME type, for example
+.IR mime:application/pdf .
+.TP
+.B name:\fIfragment\fR
+Match a fragment of the filename.
+.SH FILES
+.TP
+.I ~/.config/quicksearch/config.toml
+Configuration. Created with defaults on first run. See
+.I /usr/share/doc/quicksearch/config_example.toml
+for the annotated reference.
+.TP
+.I ~/.local/share/quicksearch/index.sqlite
+The index. The location is set by
+.I [paths].database_path
+in the configuration.
+.TP
+.I ./config.toml
+A configuration file placed next to the
+.B quicksearch
+executable selects portable mode and overrides the per\-user configuration
+entirely. The packaged build deliberately installs no such file.
+.PP
+.I XDG_CONFIG_HOME
+and
+.I XDG_DATA_HOME
+are honoured when set to absolute paths.
+.SH EXIT STATUS
+.TP
+.B 0
+The search completed, or
+.B \-\-help
+was requested.
+.TP
+.B 1
+The application failed to start.
+.TP
+.B 2
+The configuration could not be read, the index could not be opened, the query
+was rejected, or
+.B quicksearch\-cli
+was given no query.
+.SH EXAMPLES
+.TP
+Search for a phrase:
+.B quicksearch quarterly revenue report
+.TP
+Recent PDFs under a directory, with details:
+.B quicksearch \-\-long type:Document mime:application/pdf path:/home/me/docs modified:>=2026-01-01
+.TP
+Tolerate misspellings and cap the output:
+.B quicksearch \-\-fuzzy \-\-limit 10 recieve
+.SH SEE ALSO
+.I /usr/share/doc/quicksearch/README.md
+.SH AUTHOR
+Jeremy <jeremy@karsttech.com>

packaging/quicksearch.desktop

@@ -0,0 +1,19 @@
+[Desktop Entry]
+Type=Application
+Version=1.0
+Name=QuickSearch
+GenericName=File Search
+Comment=Fast full-text search across your files
+Exec=quicksearch
+TryExec=quicksearch
+Icon=quicksearch
+Terminal=false
+# One main category only (Utility), so the app shows up exactly once in the
+# menu; Filesystem is the additional category that describes what it does.
+Categories=Utility;Filesystem;
+Keywords=search;find;index;full-text;files;content;
+StartupNotify=true
+# Matches ViewportBuilder::with_app_id("quicksearch") in the GUI, which becomes
+# the Wayland app id and the X11 WM_CLASS. Without the match the desktop shows a
+# generic window icon.
+StartupWMClass=quicksearch

run.bat

@@ -1 +1,11 @@
-cargo run -p quicksearch-gui
+@echo off
+rem Build and launch the QuickSearch GUI, mirroring run.sh.
+rem Terminal search is a separate binary on Windows, because the GUI is built
+rem as a window-subsystem app and cannot write to the calling shell:
+rem     target\release\quicksearch-cli.exe --help
+setlocal
+cargo build --release -p quicksearch-gui
+if errorlevel 1 exit /b 1
+rem %~dp0 is this script's own directory (with a trailing backslash), so the
+rem launch does not depend on the current working directory.
+"%~dp0target\release\quicksearch.exe" %*

run.sh

@@ -1,2 +1,7 @@
+#!/usr/bin/env sh
+# Build and launch the QuickSearch GUI. On Unix the same binary doubles as the
+# terminal search tool: ./target/release/quicksearch --help
+# (On Windows that role belongs to quicksearch-cli.exe — see run.bat.)
+set -e
 cargo build --release -p quicksearch-gui
-./target/release/quicksearch
+exec ./target/release/quicksearch "$@"

setup.sh

@@ -1,23 +0,0 @@
-#!/usr/bin/env bash
-# One-time system setup for building and running the Dioxus-based
-# `quicksearch-gui` binary on Debian / Ubuntu. The core library
-# (`quicksearch-core`) has no native system deps and doesn't need this.
-#
-# Usage:
-#     ./setup.sh                # installs everything
-#
-# Adds `pkg-config` and `build-essential` explicitly because the Dioxus
-# build scripts (glib-sys, gdk-sys, gio-sys, javascriptcoregtk-sys,
-# webkit2gtk-sys) shell out to `pkg-config` and will refuse to build
-# without it. Some minimal Ubuntu flavors and container images don't
-# install pkg-config transitively.
-set -e
-
-sudo apt update
-sudo apt install -y \
-    pkg-config \
-    build-essential \
-    libsoup-3.0-dev \
-    libjavascriptcoregtk-4.1-dev \
-    libwebkit2gtk-4.1-dev \
-    libxdo-dev