tests: replace copyrighted lyrics with placeholder text

The align, transcribe, lyrics and notes tests quoted lines from published songs, and the UltraStar parser fixtures were complete community charts of commercial songs. Replace them with counting words and two hand-written charts that keep what the parser tests rely on: CRLF line endings, Latin-1 bytes, a decimal-comma BPM, freestyle/golden/rap notes, negative pitch and a two-number line break.
Merge pull request 'Keep matched words that fall inside a missing-lyrics span' (#13 ) from missing-notes into main
2026-10-09 17:17:02 -07:00 · 2026-10-09 16:54:17 -07:00 · 2026-10-09 16:52:56 -07:00 · 2026-10-09 16:43:24 -07:00 · 2026-10-09 14:36:29 -07:00 · 2026-10-06 13:44:44 -07:00
64 changed files with 3996 additions and 410 deletions
@@ -11,9 +11,9 @@ COPY pyproject.toml ./
 COPY src ./src
 RUN pip install --no-cache-dir .

-ENV KARAOKE_MAKER_CONFIG=/config/config.toml \
+ENV SONGTRACE_CONFIG=/config/config.toml \
    TORCH_HOME=/data/models \
    XDG_CACHE_HOME=/data/models/cache

 EXPOSE 8000
-CMD ["karaoke-maker", "serve"]
+CMD ["songtrace", "serve"]
@@ -0,0 +1,661 @@
+                    GNU AFFERO GENERAL PUBLIC LICENSE
+                       Version 3, 19 November 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 Affero General Public License is a free, copyleft license for
+software and other kinds of works, specifically designed to ensure
+cooperation with the community in the case of network server software.
+
+  The licenses for most software and other practical works are designed
+to take away your freedom to share and change the works.  By contrast,
+our General Public Licenses are 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.
+
+  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.
+
+  Developers that use our General Public Licenses protect your rights
+with two steps: (1) assert copyright on the software, and (2) offer
+you this License which gives you legal permission to copy, distribute
+and/or modify the software.
+
+  A secondary benefit of defending all users' freedom is that
+improvements made in alternate versions of the program, if they
+receive widespread use, become available for other developers to
+incorporate.  Many developers of free software are heartened and
+encouraged by the resulting cooperation.  However, in the case of
+software used on network servers, this result may fail to come about.
+The GNU General Public License permits making a modified version and
+letting the public access it on a server without ever releasing its
+source code to the public.
+
+  The GNU Affero General Public License is designed specifically to
+ensure that, in such cases, the modified source code becomes available
+to the community.  It requires the operator of a network server to
+provide the source code of the modified version running there to the
+users of that server.  Therefore, public use of a modified version, on
+a publicly accessible server, gives the public access to the source
+code of the modified version.
+
+  An older license, called the Affero General Public License and
+published by Affero, was designed to accomplish similar goals.  This is
+a different license, not a version of the Affero GPL, but Affero has
+released a new version of the Affero GPL which permits relicensing under
+this license.
+
+  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 Affero 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. Remote Network Interaction; Use with the GNU General Public License.
+
+  Notwithstanding any other provision of this License, if you modify the
+Program, your modified version must prominently offer all users
+interacting with it remotely through a computer network (if your version
+supports such interaction) an opportunity to receive the Corresponding
+Source of your version by providing access to the Corresponding Source
+from a network server at no charge, through some standard or customary
+means of facilitating copying of software.  This Corresponding Source
+shall include the Corresponding Source for any work covered by version 3
+of the GNU General Public License that is incorporated pursuant to the
+following paragraph.
+
+  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 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 work with which it is combined will remain governed by version
+3 of the GNU General Public License.
+
+  14. Revised Versions of this License.
+
+  The Free Software Foundation may publish revised and/or new versions of
+the GNU Affero 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 Affero 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 Affero 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 Affero 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 Affero 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 Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero 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 your software can interact with users remotely through a computer
+network, you should also make sure that it provides a way for users to
+get its source.  For example, if your program is a web application, its
+interface could display a "Source" link that leads users to an archive
+of the code.  There are many ways you could offer source, and different
+solutions will be better for different programs; see section 13 for the
+specific requirements.
+
+  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 AGPL, see
+<https://www.gnu.org/licenses/>.
@@ -1,17 +1,22 @@
-# karaoke-maker
+# songtrace
+
+Turn a song and its lyrics into a playable UltraStar chart.

 A CLI and minimal web UI that turns a song — title, artist, lyrics you supply,
 and a YouTube URL or local video file — into a playable
-[UltraStar](https://usdx.eu/format/)-format song folder for
-[Melody Mania](https://melodymania.net/), a commercial UltraStar Deluxe fork.
+[UltraStar](https://usdx.eu/format/)-format song folder. Also compatible with
+[Melody Mania](https://melodymania.org/).
+
+![Web UI](docs/images/webui.png)

 The pipeline:

 1. **acquire** — download audio/video via `yt-dlp` (or use a local video file)
 2. **separate** — split vocals/instrumental with Demucs
-3. **transcribe** — send the vocal stem to your own intranet OpenAI-compatible
-   Whisper server, as a sanity check that the supplied lyrics match the audio
-   and to locate vocal sections
+3. **transcribe** — send the vocal stem to your own intranet Whisper server
+   (OpenAI-compatible, or [onerahmet/openai-whisper-asr-webservice](https://github.com/ahmetoner/whisper-asr-webservice)),
+   as a sanity check that the supplied lyrics match the audio and to locate
+   vocal sections
 4. **align** — forced-align the *supplied* reference lyrics against the vocal
   stem with wav2vec2 (torchaudio's MMS_FA) to get word timings
 5. **pitch** — track pitch on the vocal stem with torchcrepe
@@ -26,16 +31,37 @@ output.
 **The output is a best-effort draft**, not a finished song. Every run ends by
 printing the low-confidence regions it found (e.g.
 `3 low-confidence notes near 0:42–0:47`) — open the result in an UltraStar
-editor such as [Yass](http://www.yass-hq.de/) or
+editor such as [Yass](https://www.yass-along.com/) or
 [USDX](https://usdx.eu/) and touch those spots up before playing.

+## Reviewing a song
+
+The web UI plays the finished song folder back against the review, so you can
+check the timing without loading it into a game first. Pick a song in the
+queue and the review panel gets the packaged video with a transport above the
+lyrics; as it plays, the line being sung is highlighted, words light up as
+they are reached, and a playhead sweeps each line against its notes. Click any
+line or word to play from there.
+
+Playback is modelled on the game: the `#MP3` is the audio and the clock, and
+the `#VIDEO` is muted background slaved to it. It also uses the same seconds
+the notes were generated from, so a word that lights up late in the browser is
+late in the game too.
+
+The packaged video has its audio track stripped for the same reason. It is the
+same recording as the `#MP3`, so a player that does not mute the video plays
+both at once, a few frames apart, which is heard as an echo. Songs built
+before this change still have it — check with `ffprobe <song>.mp4` for an
+audio stream, and rebuild with **Restart from** set to `package` to clear it
+without re-downloading.
+
 ## Configuration

 Config is TOML. Resolution order:

 1. `--config <path>` passed explicitly on the command line
-2. the `KARAOKE_MAKER_CONFIG` environment variable
-3. `~/.config/karaoke-maker/config.toml`
+2. the `SONGTRACE_CONFIG` environment variable
+3. `~/.config/songtrace/config.toml`

 Copy the example and fill in your Whisper server:

@@ -45,8 +71,17 @@ $EDITOR config.toml   # set whisper_url at minimum
 ```

 See `config.example.toml` for every available field (`whisper_url`,
-`whisper_api_key`, `whisper_model`, `output_dir`, `work_dir`, `creator`,
-`demucs_model`, `crepe_model`, `device`, `bind_host`, `port`).
+`whisper_api_key`, `whisper_model`, `whisper_api`, `output_dir`, `work_dir`,
+`creator`, `demucs_model`, `crepe_model`, `device`, `bind_host`, `port`,
+`ytdlp_clients`, `ytdlp_cookies`, `ytdlp_impersonate`,
+`ytdlp_pot_provider_url`, `ytdlp_version`).
+
+`whisper_api` selects the wire format: `"openai"` (default) speaks the
+OpenAI-compatible `POST {whisper_url}/audio/transcriptions` API; `"asr-webservice"`
+speaks [onerahmet/openai-whisper-asr-webservice](https://github.com/ahmetoner/whisper-asr-webservice)'s
+`POST {whisper_url}/asr` API instead (point `whisper_url` at that server's base
+URL with no `/v1` suffix — `whisper_model`/`whisper_api_key` are unused for
+this backend, since the model is fixed by how that server was started).

 ## Docker

@@ -63,7 +98,7 @@ container recreation.
 Run the CLI inside the container instead of the web UI:

 ```sh
-docker compose run --rm karaoke-maker karaoke-maker make \
+docker compose run --rm songtrace songtrace make \
  --title "Song Title" --artist "Artist Name" \
  --lyrics /data/lyrics.txt --youtube https://youtu.be/...
 ```
@@ -73,21 +108,45 @@ Toolkit](https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/)
 installed on the host; `docker-compose.yml` requests one GPU via the
 `nvidia` device driver.

+## Sharing the GPU
+
+`songtrace serve` is long-lived but only wants the GPU in bursts, so it hands
+it back after every stage rather than holding it for the life of the process.
+Three stages load a model — `separate` (Demucs), `align` (wav2vec2) and
+`pitch` (torchcrepe) — and each one's weights are dropped as soon as that
+stage finishes, followed by `torch.cuda.empty_cache()` to return the
+allocator's freed blocks to the driver instead of keeping them for reuse. This
+matters most between `separate` and `align`: `transcribe` sits between them
+waiting on your Whisper server, which can be minutes of a GPU nothing is
+using.
+
+Stages that gave memory back say so in the job log:
+
+```
+[separate] released 3704 MiB of GPU memory
+```
+
+The cost is reloading each model once per job instead of once per process,
+which is seconds against stages that run for minutes. What this does not
+release is the CUDA context itself, a few hundred MB that belongs to the
+process until it exits; freeing that would mean running each stage in a
+subprocess. If you need every last MB back, stop the container between runs.
+
 ## Bare-metal

 ```sh
 pip install -e .
-karaoke-maker make --help
+songtrace make --help
 ```

 ```
-karaoke-maker make --title TITLE --artist ARTIST --lyrics LYRICS_FILE \
+songtrace make --title TITLE --artist ARTIST --lyrics LYRICS_FILE \
  (--youtube URL | --video PATH) \
  [--genre GENRE] [--year YEAR] [--language LANGUAGE] [--cover COVER_IMAGE] \
  [--bpm BPM] [--force-from STAGE] [--config CONFIG]
 ```

-Run the web UI instead with `karaoke-maker serve [--config CONFIG]`.
+Run the web UI instead with `songtrace serve [--config CONFIG]`.

 ## Re-running a stage

@@ -98,9 +157,82 @@ stage and everything after it, reusing the cached output of everything
 before it:

 ```sh
-karaoke-maker make --title "..." --artist "..." --lyrics lyrics.txt \
+songtrace make --title "..." --artist "..." --lyrics lyrics.txt \
  --youtube https://youtu.be/... --force-from align
 ```

 Valid stages, in pipeline order: `acquire`, `separate`, `transcribe`,
 `align`, `pitch`, `notes`, `grid`, `package`.
+
+In the web UI this is what selecting a queue entry does: its title, artist,
+lyrics and the rest come back up in the form, so you can adjust them and hit
+REBUILD SONG. Only the stages your edit actually invalidates are rerun —
+fixing a line of lyrics realigns without re-downloading or re-separating.
+CREATE SONG, at the top of the queue, clears the page to start a fresh one.
+
+That planning works off the fields you changed, which cannot see a change in
+songtrace itself: after an upgrade, a rebuild of an older song finds every
+stage cached and does nothing. **Restart from** on the rebuild form is the way
+to redo a stage anyway — pick `package` to repackage without re-downloading.
+It can only widen what is rerun, so an edit that invalidates an earlier stage
+still wins.
+
+## YouTube download failures
+
+`HTTP Error 403: Forbidden` from the acquire stage means YouTube served the
+page but refused the media URL. In rough order of likelihood:
+
+1. **yt-dlp is stale.** YouTube changes its defences every few weeks, so
+   songtrace keeps yt-dlp current itself. Once a day, and again whenever a
+   download fails outright, it checks PyPI for a newer stable release,
+   installs it under `$XDG_CACHE_HOME/songtrace/` (`/data/models/cache` in the
+   container), and tries it before the copy installed with songtrace. That
+   copy is never modified and is always the fallback, so a release that breaks
+   compatibility costs a retry rather than the download. The job log names
+   the yt-dlp version behind each attempt.
+
+   `ytdlp_version` overrides this: `"bundled"` uses only the installed copy,
+   and a release number such as `"2026.08.19"` pins that release instead of
+   the newest.
+
+2. **Missing JS runtime.** yt-dlp needs `yt-dlp-ejs` plus a JavaScript engine
+   to solve YouTube's player challenges; without them it silently falls back to
+   formats that 403. The `[default,curl-cffi,deno]` extras in `pyproject.toml`
+   cover this, for both the installed copy and the ones fetched at runtime.
+   Check with `docker compose exec songtrace yt-dlp -v --simulate <url>` — the debug
+   header must show a `JS runtimes:` line that is not `none`.
+
+3. **The IP is rate-limited or flagged.** Datacenter and VPN ranges get
+   throttled hardest. Try from another network before changing config.
+
+4. **A PO token is required.** If the log says so explicitly, run
+   [bgutil-ytdlp-pot-provider](https://github.com/Brainicism/bgutil-ytdlp-pot-provider)
+   as a sidecar, install its plugin into the image, and point
+   `ytdlp_pot_provider_url` at it.
+
+`ytdlp_clients` retries the download against each listed player client in
+turn, so a client that YouTube has broken is skipped automatically.
+`ytdlp_cookies` (a Netscape cookie jar, for age-gated videos) and
+`ytdlp_impersonate` (a browser TLS fingerprint, e.g. `"chrome"`) are the other
+two levers. Do not use a cookie jar from an account you care about — YouTube
+bans for automated access.
+
+Failed downloads leave the work dir intact, so re-running the job resumes
+rather than starting over.
+
+## Regenerating the web UI screenshot
+
+`scripts/screenshot_webui.py` seeds the real app with a fixed set of jobs
+covering every status and screenshots it to `docs/images/webui.png`. It uses
+`ffmpeg` to give the selected song a playable folder so the shot includes the
+player; without `ffmpeg` on `PATH` it still runs, just without it.
+
+```sh
+pip install -e '.[dev]'
+playwright install chromium
+python scripts/screenshot_webui.py
+```
+
+## License
+
+[AGPLv3](LICENSE)
@@ -1,6 +1,8 @@
 whisper_url = "http://whisper.intranet:9000/v1"
 # whisper_api_key = ""
 # whisper_model = "whisper-1"
+# whisper_api = "openai"  # or "asr-webservice" for onerahmet/openai-whisper-asr-webservice
+                          # (set whisper_url without the /v1 suffix, e.g. "http://whisper.intranet:9000")
 output_dir = "/data/output"
 work_dir = "/data/work"
 creator = "brad"
@@ -9,3 +11,13 @@ creator = "brad"
 # device = "cuda"
 bind_host = "0.0.0.0"
 # port = 8000
+
+# YouTube downloads. Each player client is tried in order until one works;
+# "default" means whatever yt-dlp currently picks for itself.
+# ytdlp_clients = ["default", "tv"]
+# ytdlp_cookies = "/config/youtube-cookies.txt"  # Netscape cookie jar
+# ytdlp_impersonate = "chrome"                   # TLS fingerprint to present
+# ytdlp_pot_provider_url = "http://bgutil-provider:4416"
+# Newer yt-dlp releases are fetched from PyPI at runtime and tried before the
+# bundled one; "bundled" turns that off, a release number pins one.
+# ytdlp_version = "auto"
@@ -1,5 +1,5 @@
 services:
-  karaoke-maker:
+  songtrace:
    build: .
    ports:
      - "8000:8000"
@@ -1,10 +1,16 @@
 [project]
-name = "karaoke-maker"
+name = "songtrace"
 version = "0.1.0"
 description = "Generate UltraStar/Melody Mania karaoke songs from a video and reference lyrics"
 requires-python = ">=3.11"
+license = "AGPL-3.0-or-later"
 dependencies = [
-    "yt-dlp",
+    # yt-dlp declares no runtime deps of its own; the extras are what make
+    # YouTube work. "default" pulls yt-dlp-ejs (JS challenge solving) and the
+    # requests/urllib3 handlers, "deno" ships the JS runtime that executes it,
+    # and "curl-cffi" provides the --impersonate targets. Without them YouTube
+    # extraction degrades to HTTP 403 on the media URLs.
+    "yt-dlp[default,curl-cffi,deno]",
    "demucs",
    "torch",
    "torchaudio",
@@ -22,10 +28,10 @@ dependencies = [
 ]

 [project.optional-dependencies]
-dev = ["pytest"]
+dev = ["pytest", "playwright"]

 [project.scripts]
-karaoke-maker = "karaoke_maker.cli:main"
+songtrace = "songtrace.cli:main"

 [build-system]
 requires = ["setuptools>=68"]
@@ -35,7 +41,7 @@ build-backend = "setuptools.build_meta"
 where = ["src"]

 [tool.setuptools.package-data]
-karaoke_maker = ["static/*", "static/fonts/*"]
+songtrace = ["static/*", "static/fonts/*"]

 [tool.pytest.ini_options]
 testpaths = ["tests"]
@@ -0,0 +1,256 @@
+"""Regenerate docs/images/webui.png against the current web UI.
+
+Seeds the real app (via create_app's fake-runner seam, same as tests/test_web.py)
+with a fixed set of jobs covering every status, then screenshots it.
+
+One-time setup: pip install -e '.[dev]' && playwright install chromium
+Run: python scripts/screenshot_webui.py
+"""
+import json
+import shutil
+import subprocess
+import tempfile
+import threading
+import time
+from pathlib import Path
+
+import httpx
+import uvicorn
+from playwright.sync_api import sync_playwright
+
+from songtrace.config import Config
+from songtrace.pipeline import PipelineError, PipelineResult, song_slug
+from songtrace.web import create_app
+
+REPO_ROOT = Path(__file__).resolve().parent.parent
+OUT_IMAGE = REPO_ROOT / "docs" / "images" / "webui.png"
+STAGES = ["acquire", "separate", "transcribe", "align", "pitch", "notes", "grid", "package"]
+DEFAULT_LYRICS = "[verse]\n...\n[chorus]\n...\n"
+
+# public domain (16th century English folk song, author unknown)
+GREENSLEEVES_LYRICS = """\
+Alas, my love, you do me wrong
+To cast me off discourteously
+For I have loved you well and long
+Delighting in your company
+
+Greensleeves was all my joy
+Greensleeves was my delight
+Greensleeves was my heart of gold
+And who but my lady Greensleeves
+
+I have been ready at your hand
+To grant whatever you would crave
+I have both waged life and land
+Your love and good will for to have
+
+Greensleeves was all my joy
+Greensleeves was my delight
+Greensleeves was my heart of gold
+And who but my lady Greensleeves
+"""
+
+# order matters: the app has one worker thread, so each job runs in submission
+# order. "running" must be submitted after the done/failed ones complete and
+# before "queued", so it's the one still in progress when "queued" arrives.
+SONGS = [
+    ("Every Single Night", "Fiona Apple", "done",
+     {"warnings": ["1 low-confidence note near 2:41–2:44"]}),
+    ("This Year", "The Mountain Goats", "done", {"warnings": []}),
+    ("Pink Moon", "Nick Drake", "done",
+     {"warnings": ["2 low-confidence notes near 0:38–0:41",
+                    "1 low-confidence note near 1:22–1:24"]}),
+    ("Landslide", "Fleetwood Mac", "done",
+     {"warnings": ["1 low-confidence note near 0:52–0:55"]}),
+    ("Greensleeves", "Traditional", "done",
+     {"lyrics": GREENSLEEVES_LYRICS,
+      "warnings": ["1 low-confidence note near 0:45–0:47",
+                   "1 line dropped (not sung) near \"for i have loved\"",
+                   "missing lyrics near 2:10–2:18"],
+      "review": {
+          "lines": [
+              # Pitches are the real Greensleeves melody (E natural minor, as
+              # arranged by Gilbert DeBenedetti, "Level 4 - Flowing"), converted
+              # to UltraStar semitones (0 = C4 = MIDI 60): E5=16 F#5=18 G5=19
+              # A5=21 B4=11 D5=14.
+              {"index": 0, "text": "Alas, my love, you do me wrong",
+               "dropped": False,
+               "notes": [{"start": 10.0, "end": 10.4, "pitch": 16, "low_confidence": False},
+                         {"start": 10.4, "end": 10.6, "pitch": 18, "low_confidence": False},
+                         {"start": 10.6, "end": 11.2, "pitch": 19, "low_confidence": False},
+                         {"start": 11.2, "end": 11.5, "pitch": 21, "low_confidence": False},
+                         {"start": 11.5, "end": 11.7, "pitch": 19, "low_confidence": False},
+                         {"start": 11.7, "end": 11.9, "pitch": 18, "low_confidence": False},
+                         {"start": 11.9, "end": 12.9, "pitch": 14, "low_confidence": False}],
+               "words": [{"text": "Alas,", "start": 10.0, "end": 10.4},
+                         {"text": "my", "start": 10.4, "end": 10.6},
+                         {"text": "love,", "start": 10.6, "end": 11.2},
+                         {"text": "you", "start": 11.2, "end": 11.5},
+                         {"text": "do", "start": 11.5, "end": 11.7},
+                         {"text": "me", "start": 11.7, "end": 11.9},
+                         {"text": "wrong", "start": 11.9, "end": 12.9}]},
+              {"index": 1, "text": "To cast me off discourteously",
+               "dropped": False,
+               "notes": [{"start": 44.0, "end": 44.3, "pitch": 11, "low_confidence": False},
+                         {"start": 44.3, "end": 44.6, "pitch": 16, "low_confidence": False},
+                         {"start": 44.6, "end": 44.8, "pitch": 18, "low_confidence": False},
+                         {"start": 44.8, "end": 45.0, "pitch": 19, "low_confidence": False},
+                         {"start": 45.0, "end": 45.4, "pitch": 18, "low_confidence": True}],
+               "words": [{"text": "To", "start": 44.0, "end": 44.3},
+                         {"text": "cast", "start": 44.3, "end": 44.6},
+                         {"text": "me", "start": 44.6, "end": 44.8},
+                         {"text": "off", "start": 44.8, "end": 45.0},
+                         {"text": "discourteously", "start": 45.0, "end": 45.4}]},
+              {"index": 2, "text": "For I have loved you well and long",
+               "dropped": True, "notes": [], "words": []},
+          ],
+          "missing_spans": [{"start": 130.0, "end": 138.0}],
+      }}),
+    ("Such Great Heights", "The Postal Service", "failed",
+     {"error": "yt-dlp: unable to download video data: HTTP Error 403: Forbidden"}),
+    ("Time Escaping", "Big Thief", "running", {}),
+    ("Vampire in the Corner", "Magdalena Bay", "queued", {}),
+]
+SELECTED = "Greensleeves"
+REVIEW_HEADING = "REVIEW"
+
+running_event = threading.Event()
+
+
+def make_runner():
+    specs = {title: (kind, extra) for title, _, kind, extra in SONGS}
+
+    def runner(inp, cfg, log=print, force_from=None, stages=None):
+        kind, extra = specs[inp.title]
+        if kind == "done":
+            for s in STAGES:
+                log(f"[{s}] running")
+            if "review" in extra:
+                work = cfg.work_dir / song_slug(inp)
+                work.mkdir(parents=True, exist_ok=True)
+                notes = [
+                    {**n, "duration": n["end"] - n["start"], "text": "~",
+                     "line_index": line["index"]}
+                    for line in extra["review"]["lines"] for n in line["notes"]
+                ]
+                (work / "notes_sec.json").write_text(json.dumps({
+                    "notes": notes, "breaks": [], "warnings": extra["warnings"]}))
+                (work / "align_meta.json").write_text(json.dumps({
+                    "dropped_lines": [line["index"] for line in extra["review"]["lines"]
+                                      if line["dropped"]],
+                    "missing_spans": [[s["start"], s["end"]]
+                                      for s in extra["review"]["missing_spans"]],
+                }))
+                (work / "alignment.json").write_text(json.dumps([
+                    {"word": w["text"], "start": w["start"], "end": w["end"],
+                     "score": 0.9, "line_index": line["index"]}
+                    for line in extra["review"]["lines"] for w in line.get("words", [])
+                ]))
+            out = cfg.output_dir / f"{inp.artist} - {inp.title}"
+            out.mkdir(parents=True, exist_ok=True)
+            return PipelineResult(output_dir=out, warnings=extra["warnings"])
+        if kind == "failed":
+            log("[acquire] running")
+            raise PipelineError(extra["error"])
+        for s in ["acquire", "separate", "transcribe", "align"]:
+            log(f"[{s}] running")
+        running_event.wait()
+        return PipelineResult(output_dir=cfg.output_dir / "x", warnings=[])
+
+    return runner
+
+
+def start_server(tmp_path: Path) -> tuple[str, Config]:
+    cfg = Config(whisper_url="http://localhost:9000/v1", output_dir=tmp_path / "out",
+                 work_dir=tmp_path / "work", bind_host="127.0.0.1", port=8811)
+    app = create_app(cfg, runner=make_runner())
+    server = uvicorn.Server(uvicorn.Config(app, host=cfg.bind_host, port=cfg.port,
+                                            log_level="warning"))
+    threading.Thread(target=server.run, daemon=True).start()
+    return f"http://{cfg.bind_host}:{cfg.port}", cfg
+
+
+def install_media(cfg: Config) -> bool:
+    """Give the selected song a playable folder, so the shot shows the player.
+
+    A flat fill rather than a test pattern: the point is the transport and the
+    lyric tracking underneath it, not the picture. Skipped without ffmpeg.
+    """
+    if shutil.which("ffmpeg") is None:
+        print("ffmpeg not found; screenshotting without the player")
+        return False
+    artist, title = next((a, t) for t, a, _, _ in SONGS if t == SELECTED)
+    slug = f"{artist} - {title}"
+    out = cfg.output_dir / slug
+    out.mkdir(parents=True, exist_ok=True)
+    subprocess.run(
+        ["ffmpeg", "-y", "-loglevel", "error", "-f", "lavfi",
+         "-i", "color=c=0x1A1430:s=640x360:d=60", "-c:v", "libx264",
+         "-pix_fmt", "yuv420p", "-preset", "ultrafast", "-an",
+         str(out / f"{slug}.mp4")], check=True)
+    subprocess.run(
+        ["ffmpeg", "-y", "-loglevel", "error", "-f", "lavfi",
+         "-i", "anullsrc=r=44100:cl=stereo:d=60", "-b:a", "64k",
+         str(out / f"{slug}.mp3")], check=True)
+    (out / f"{slug}.txt").write_text(
+        f"#TITLE:{title}\n#ARTIST:{artist}\n#MP3:{slug}.mp3\n#VIDEO:{slug}.mp4\n"
+        f"#BPM:240\n#GAP:10000\n: 0 4 16 Alas,\nE\n", encoding="utf-8")
+    return True
+
+
+def seed_jobs(base_url: str) -> None:
+    with httpx.Client(timeout=10) as client:
+        for title, artist, kind, extra in SONGS:
+            resp = client.post(f"{base_url}/api/jobs", data={
+                "title": title, "artist": artist,
+                "lyrics": extra.get("lyrics", DEFAULT_LYRICS),
+                "youtube_url": "https://youtu.be/00000000000"})
+            resp.raise_for_status()
+            job_id = resp.json()["id"]
+            if kind == "queued":
+                continue
+            for _ in range(200):
+                job = client.get(f"{base_url}/api/jobs/{job_id}").json()
+                if kind == "running" and job["stage"] == "align":
+                    break
+                if job["status"] in ("done", "failed"):
+                    break
+                time.sleep(0.05)
+
+
+def capture(base_url: str, with_player: bool) -> None:
+    with sync_playwright() as p:
+        browser = p.chromium.launch()
+        page = browser.new_page(viewport={"width": 1280, "height": 1000})
+        page.goto(base_url, wait_until="networkidle")
+        page.get_by_text(SELECTED, exact=True).click()
+        page.wait_for_timeout(500)
+        if with_player:
+            # paused mid-line, so the shot carries the playhead and the sung /
+            # current word colouring rather than an untouched transport
+            page.evaluate("document.getElementById('player-media').currentTime = 11.6")
+            page.wait_for_timeout(400)
+            # scrolled to the review: the player pins to the top of #main and
+            # the tracked lines sit under it, which is the panel in use
+            page.evaluate("""() => {
+              const main = document.getElementById('main');
+              main.scrollTo(0, document.getElementById('review').offsetTop - 12);
+            }""")
+            page.wait_for_timeout(400)
+        OUT_IMAGE.parent.mkdir(parents=True, exist_ok=True)
+        page.screenshot(path=str(OUT_IMAGE))
+        browser.close()
+
+
+def main() -> None:
+    with tempfile.TemporaryDirectory() as tmp:
+        base_url, cfg = start_server(Path(tmp))
+        time.sleep(1)
+        seed_jobs(base_url)
+        with_player = install_media(cfg)
+        capture(base_url, with_player)
+    print(f"wrote {OUT_IMAGE}")
+
+
+if __name__ == "__main__":
+    main()
@@ -1,36 +0,0 @@
-import os
-import tomllib
-from dataclasses import dataclass, fields
-from pathlib import Path
-
-
-@dataclass
-class Config:
-    whisper_url: str
-    whisper_api_key: str | None = None
-    whisper_model: str = "whisper-1"
-    output_dir: Path = Path("output")
-    work_dir: Path = Path("work")
-    creator: str = "karaoke-maker"
-    demucs_model: str = "htdemucs_ft"
-    crepe_model: str = "full"
-    device: str = "cuda"
-    bind_host: str = "127.0.0.1"
-    port: int = 8000
-
-
-_PATH_FIELDS = {"output_dir", "work_dir"}
-
-
-def load_config(path: Path | None = None) -> Config:
-    if path is None:
-        env = os.environ.get("KARAOKE_MAKER_CONFIG")
-        path = Path(env) if env else Path.home() / ".config/karaoke-maker/config.toml"
-    data = tomllib.loads(Path(path).read_text())
-    known = {f.name for f in fields(Config)}
-    unknown = set(data) - known
-    if unknown:
-        raise ValueError(f"unknown config keys: {', '.join(sorted(unknown))}")
-    for key in _PATH_FIELDS & set(data):
-        data[key] = Path(data[key])
-    return Config(**data)
@@ -1,54 +0,0 @@
-import json
-import shutil
-from pathlib import Path
-
-from ..media import ffmpeg
-from ..pipeline import Ctx, PipelineError, Stage
-
-IMAGE_EXTENSIONS = {".jpg", ".jpeg", ".png", ".webp", ".avif", ".gif"}
-
-
-def _classify_downloads(dest: Path) -> tuple[Path, Path | None]:
-    files = list(dest.glob("source_video.*"))
-    thumbs = [p for p in files if p.suffix.lower() in IMAGE_EXTENSIONS]
-    videos = [p for p in files if p.suffix.lower() not in IMAGE_EXTENSIONS]
-    if not videos:
-        raise PipelineError("yt-dlp produced no video file")
-    video = max(videos, key=lambda p: p.stat().st_size)
-    return video, thumbs[0] if thumbs else None
-
-
-def download_youtube(url: str, dest: Path) -> tuple[Path, Path | None]:
-    import yt_dlp
-    opts = {
-        "outtmpl": str(dest / "source_video.%(ext)s"),
-        "format": "bestvideo*+bestaudio/best",
-        "writethumbnail": True,
-        "noplaylist": True,
-    }
-    with yt_dlp.YoutubeDL(opts) as ydl:
-        ydl.extract_info(url, download=True)
-    return _classify_downloads(dest)
-
-
-def _run(ctx: Ctx) -> None:
-    if bool(ctx.inp.youtube_url) == bool(ctx.inp.video_path):
-        raise PipelineError("provide exactly one of youtube_url / video_path")
-    thumb = None
-    if ctx.inp.youtube_url:
-        video, thumb = download_youtube(ctx.inp.youtube_url, ctx.work)
-    else:
-        video = ctx.work / f"source_video{ctx.inp.video_path.suffix}"
-        shutil.copyfile(ctx.inp.video_path, video)
-    ffmpeg("-i", str(video), "-vn", "-ac", "2", "-ar", "44100",
-           str(ctx.work / "audio.wav"))
-    meta = {"video_file": video.name, "thumbnail": thumb.name if thumb else None}
-    (ctx.work / "meta.json").write_text(json.dumps(meta))
-
-
-def acquire_stage() -> Stage:
-    return Stage(
-        name="acquire",
-        outputs=lambda ctx: [ctx.work / "meta.json", ctx.work / "audio.wav"],
-        run=_run,
-    )
@@ -1,43 +0,0 @@
-import difflib
-import json
-
-from ..lyrics import RefLyrics, alignment_token
-from ..pipeline import Ctx, Stage
-from ..whisper_client import TranscriptSegment, transcribe
-
-
-def vocal_sections(segments, min_gap: float = 6.0, pad: float = 1.0):
-    sections = []
-    for s in sorted(segments, key=lambda s: s.start):
-        if sections and s.start - sections[-1][1] < min_gap:
-            sections[-1][1] = max(sections[-1][1], s.end)
-        else:
-            sections.append([s.start, s.end])
-    return [(max(0.0, a - pad), b + pad) for a, b in sections]
-
-
-def _tokens(text: str) -> list[str]:
-    return [t for w in text.split() if (t := alignment_token(w))]
-
-
-def lyrics_similarity(segments, ref: RefLyrics) -> float:
-    heard = [t for s in segments for t in _tokens(s.text)]
-    expected = [t for w in ref.words if (t := alignment_token(w))]
-    return difflib.SequenceMatcher(a=heard, b=expected).ratio()
-
-
-def _run(ctx: Ctx) -> None:
-    segments = transcribe(ctx.cfg, ctx.work / "vocals.wav", ctx.inp.language)
-    (ctx.work / "transcript.json").write_text(json.dumps(
-        [{"start": s.start, "end": s.end, "text": s.text} for s in segments]))
-
-
-def transcribe_stage() -> Stage:
-    return Stage(name="transcribe",
-                 outputs=lambda ctx: [ctx.work / "transcript.json"],
-                 run=_run)
-
-
-def load_transcript(ctx: Ctx) -> list[TranscriptSegment]:
-    data = json.loads((ctx.work / "transcript.json").read_text())
-    return [TranscriptSegment(**d) for d in data]
@@ -1,39 +0,0 @@
-from dataclasses import dataclass
-from pathlib import Path
-
-import httpx
-
-from .config import Config
-from .languages import LANGUAGES
-
-
-@dataclass
-class TranscriptSegment:
-    start: float
-    end: float
-    text: str
-
-
-def transcribe(cfg: Config, audio_path: Path, language: str | None = None,
-               client: httpx.Client | None = None) -> list[TranscriptSegment]:
-    headers = {}
-    if cfg.whisper_api_key:
-        headers["Authorization"] = f"Bearer {cfg.whisper_api_key}"
-    data = {"model": cfg.whisper_model, "response_format": "verbose_json"}
-    if language:
-        entry = LANGUAGES.get(language.lower())
-        data["language"] = entry.iso if entry else language
-    own = client is None
-    client = client or httpx.Client(timeout=600)
-    try:
-        with audio_path.open("rb") as f:
-            resp = client.post(
-                f"{cfg.whisper_url.rstrip('/')}/audio/transcriptions",
-                headers=headers, data=data,
-                files={"file": (audio_path.name, f, "audio/wav")})
-        resp.raise_for_status()
-        return [TranscriptSegment(s["start"], s["end"], s["text"].strip())
-                for s in resp.json().get("segments", [])]
-    finally:
-        if own:
-            client.close()
@@ -0,0 +1,14 @@
+from pathlib import Path
+
+import soundfile as sf
+import torch
+
+
+def load_wav(path: Path) -> tuple[torch.Tensor, int]:
+    """Load a WAV file as a (channels, samples) float32 tensor.
+
+    Reads via soundfile rather than torchaudio.load, whose default backend
+    now requires the separate torchcodec package.
+    """
+    data, sr = sf.read(str(path), dtype="float32", always_2d=True)
+    return torch.from_numpy(data.T).contiguous(), sr
@@ -16,7 +16,7 @@ def _positive_float(value: str) -> float:


 def build_parser() -> argparse.ArgumentParser:
-    parser = argparse.ArgumentParser(prog="karaoke-maker")
+    parser = argparse.ArgumentParser(prog="songtrace")
    parser.add_argument("--version", action="store_true")
    sub = parser.add_subparsers(dest="command")

@@ -72,7 +72,7 @@ def _cmd_make(args) -> int:
 def main(argv: list[str] | None = None) -> int:
    args = build_parser().parse_args(argv)
    if args.version:
-        print(f"karaoke-maker {version('karaoke-maker')}")
+        print(f"songtrace {version('songtrace')}")
        return 0
    if args.command == "make":
        return _cmd_make(args)
@@ -0,0 +1,46 @@
+import os
+import tomllib
+from dataclasses import dataclass, field, fields
+from pathlib import Path
+
+
+@dataclass
+class Config:
+    whisper_url: str
+    whisper_api_key: str | None = None
+    whisper_model: str = "whisper-1"
+    whisper_api: str = "openai"
+    output_dir: Path = Path("output")
+    work_dir: Path = Path("work")
+    creator: str = "songtrace"
+    demucs_model: str = "htdemucs_ft"
+    crepe_model: str = "full"
+    device: str = "cuda"
+    bind_host: str = "127.0.0.1"
+    port: int = 8000
+    # YouTube player clients to try, in order, until one yields a download.
+    # "default" means yt-dlp's own choice; see its extractor-args docs.
+    ytdlp_clients: list[str] = field(default_factory=lambda: ["default", "tv"])
+    ytdlp_cookies: Path | None = None
+    ytdlp_impersonate: str | None = None
+    ytdlp_pot_provider_url: str | None = None
+    # "auto" keeps yt-dlp current from PyPI, "bundled" uses only the copy
+    # installed with songtrace, and a release such as "2026.08.19" pins that.
+    ytdlp_version: str = "auto"
+
+
+_PATH_FIELDS = {"output_dir", "work_dir", "ytdlp_cookies"}
+
+
+def load_config(path: Path | None = None) -> Config:
+    if path is None:
+        env = os.environ.get("SONGTRACE_CONFIG")
+        path = Path(env) if env else Path.home() / ".config/songtrace/config.toml"
+    data = tomllib.loads(Path(path).read_text())
+    known = {f.name for f in fields(Config)}
+    unknown = set(data) - known
+    if unknown:
+        raise ValueError(f"unknown config keys: {', '.join(sorted(unknown))}")
+    for key in _PATH_FIELDS & set(data):
+        data[key] = Path(data[key])
+    return Config(**data)
@@ -0,0 +1,47 @@
+"""Handing the GPU back between stages.
+
+The GPU this runs on is shared, and `songtrace serve` outlives every job it
+runs, so a stage that caches its weights holds VRAM until the process exits —
+including while a later stage is waiting on the Whisper server, which can be
+minutes of a GPU nothing is using. Two things have to be undone: the caches
+themselves, and PyTorch's allocator, which keeps freed blocks for reuse rather
+than returning them to the driver.
+
+A stage that caches a model registers how to drop it with @on_release, so this
+module needs to know nothing about the pipeline. Registration happens at import
+time, which run_pipeline has already done by the time it calls release().
+"""
+import gc
+import sys
+from typing import Callable
+
+_callbacks: list[Callable[[], None]] = []
+
+
+def on_release(fn: Callable[[], None]) -> Callable[[], None]:
+    """Register a callback that drops a cached model."""
+    _callbacks.append(fn)
+    return fn
+
+
+def release() -> int:
+    """Drop every cached model and return its VRAM to the driver.
+
+    Answers the bytes handed back, which is the only part of this a caller can
+    show someone. Safe whenever: a stage that never loaded anything has nothing
+    to drop, and the allocator is left alone unless this process actually
+    initialised CUDA.
+    """
+    for callback in _callbacks:
+        callback()
+    # weights reach the allocator's free list only once the last reference
+    # goes, and a module graph is cyclic enough to outlive refcounting
+    gc.collect()
+    torch = sys.modules.get("torch")
+    if torch is None or not torch.cuda.is_available():
+        return 0
+    if not torch.cuda.is_initialized():
+        return 0
+    before = torch.cuda.memory_reserved()
+    torch.cuda.empty_cache()
+    return max(before - torch.cuda.memory_reserved(), 0)
@@ -23,6 +23,7 @@ class TimedNote:
    pitch: int            # UltraStar semitone (0 = C4 = MIDI 60)
    text: str             # display text incl. leading space / "~"
    low_confidence: bool
+    line_index: int


 def hz_to_midi_float(hz: float) -> float:
@@ -78,7 +79,8 @@ def syllabify(word: str, language: str | None) -> list[str]:
    return parts or [word]


-def build_notes(rows: list[dict], times, f0, conf, language: str | None
+def build_notes(rows: list[dict], times, f0, conf, language: str | None,
+                dropped_lines=(), missing_spans=(), ref_lines=None
                ) -> tuple[list[TimedNote], list[float], list[str]]:
    notes: list[TimedNote] = []
    breaks: list[float] = []
@@ -92,7 +94,8 @@ def build_notes(rows: list[dict], times, f0, conf, language: str | None
        if not segs:
            pitch = notes[-1].pitch if notes else 0
            notes.append(_note(row["start"], row["end"] - row["start"], pitch,
-                               _text(row["word"], 0, line_initial), True))
+                               _text(row["word"], 0, line_initial), True,
+                               row["line_index"]))
            continue
        low = row["score"] < LOW_SCORE
        sylls = syllabify(row["word"], language)
@@ -100,14 +103,16 @@ def build_notes(rows: list[dict], times, f0, conf, language: str | None
            segs = _subdivide(segs, len(sylls))
        for i, (s, e, midi) in enumerate(segs):
            text = _text(sylls[i], i, line_initial) if i < len(sylls) else "~"
-            notes.append(_note(s, e - s, round(midi) - 60, text, low))
+            notes.append(_note(s, e - s, round(midi) - 60, text, low,
+                               row["line_index"]))
    _deoverlap(notes)
-    return notes, breaks, _warnings(notes)
+    return notes, breaks, _warnings(notes, dropped_lines, missing_spans, ref_lines or [])


 def _note(start: float, duration: float, pitch: int, text: str,
-          low: bool) -> TimedNote:
-    return TimedNote(float(start), float(duration), int(pitch), text, low)
+          low: bool, line_index: int) -> TimedNote:
+    return TimedNote(float(start), float(duration), int(pitch), text, low,
+                     int(line_index))


 def _deoverlap(notes: list[TimedNote]) -> None:
@@ -153,7 +158,8 @@ def _fmt(t: float) -> str:
    return f"{int(t // 60)}:{int(t % 60):02d}"


-def _warnings(notes: list[TimedNote]) -> list[str]:
+def _warnings(notes: list[TimedNote], dropped_lines=(), missing_spans=(),
+             ref_lines=()) -> list[str]:
    warnings, run = [], []
    for n in list(notes) + [None]:
        if n is not None and n.low_confidence:
@@ -164,4 +170,21 @@ def _warnings(notes: list[TimedNote]) -> list[str]:
                f"{len(run)} low-confidence note{'' if len(run) == 1 else 's'} "
                f"near {_fmt(first.start)}–{_fmt(last.start + last.duration)}")
            run = []
+    for group in _consecutive_runs(dropped_lines):
+        n = len(group)
+        text = " ".join(ref_lines[group[0]]) if group[0] < len(ref_lines) else ""
+        warnings.append(
+            f"{n} line{'' if n == 1 else 's'} dropped (not sung) near \"{text}\"")
+    for a, b in missing_spans:
+        warnings.append(f"missing lyrics near {_fmt(a)}–{_fmt(b)}")
    return warnings
+
+
+def _consecutive_runs(indices) -> list[list[int]]:
+    runs = []
+    for i in sorted(indices):
+        if runs and i == runs[-1][-1] + 1:
+            runs[-1].append(i)
+        else:
+            runs.append([i])
+    return runs
@@ -3,6 +3,7 @@ from dataclasses import dataclass, field
 from pathlib import Path
 from typing import Callable

+from . import models
 from .config import Config

 class PipelineError(Exception):
@@ -45,6 +46,12 @@ def song_slug(inp: SongInput) -> str:
    name = f"{inp.artist} - {inp.title}"
    return name.replace("/", "").replace("\0", "")

+def read_warnings(work: Path) -> list[str]:
+    notes_file = work / "notes_sec.json"
+    if not notes_file.exists():
+        return []
+    return json.loads(notes_file.read_text()).get("warnings", [])
+
 def _real_stages() -> list[Stage]:
    from . import stages
    return stages.all_stages()
@@ -68,9 +75,14 @@ def run_pipeline(inp: SongInput, cfg: Config, log=print,
            log(f"[{stage.name}] cached, skipping")
            continue
        log(f"[{stage.name}] running")
-        stage.run(ctx)
-    warnings = []
-    notes_file = work / "notes_sec.json"
-    if notes_file.exists():
-        warnings = json.loads(notes_file.read_text()).get("warnings", [])
-    return PipelineResult(output_dir=cfg.output_dir / song_slug(inp), warnings=warnings)
+        # per stage rather than per job: each model is used by one stage, so
+        # holding it past that only keeps VRAM from whatever else wants it
+        try:
+            stage.run(ctx)
+        finally:
+            # logged only when it actually gave something back, so the line is
+            # evidence rather than noise on a box where this matters
+            if freed := models.release():
+                log(f"[{stage.name}] released {freed / 2**20:.0f} MiB of GPU memory")
+    return PipelineResult(output_dir=cfg.output_dir / song_slug(inp),
+                          warnings=read_warnings(work))
@@ -38,6 +38,17 @@ def plan(old: SongInput, new: SongInput) -> str | None:
    return min((FIELD_STAGE[f] for f in changed), key=STAGE_ORDER.index)


+def earliest(*stages: str | None) -> str | None:
+    """The earliest of the stages given, ignoring None ("nothing forced").
+
+    An operator's explicit choice is combined with plan()'s rather than
+    replacing it: the planner's stage is the latest that is still correct for
+    the edit, so a choice may only ever redo more.
+    """
+    named = [s for s in stages if s]
+    return min(named, key=STAGE_ORDER.index) if named else None
+
+
 def relocate_work(old: SongInput, new: SongInput, cfg: Config) -> bool:
    """Move the cached work dir when title/artist changed its slug.

@@ -0,0 +1,93 @@
+import json
+import shutil
+from pathlib import Path
+
+from .. import ytdlp
+from ..config import Config
+from ..media import ffmpeg
+from ..pipeline import Ctx, PipelineError, Stage
+
+IMAGE_EXTENSIONS = {".jpg", ".jpeg", ".png", ".webp", ".avif", ".gif"}
+# yt-dlp's in-progress scratch files, which a failed attempt leaves behind.
+PARTIAL_EXTENSIONS = {".part", ".ytdl", ".temp"}
+
+
+def _classify_downloads(dest: Path) -> tuple[Path, Path | None]:
+    files = [p for p in dest.glob("source_video.*")
+             if p.suffix.lower() not in PARTIAL_EXTENSIONS]
+    thumbs = [p for p in files if p.suffix.lower() in IMAGE_EXTENSIONS]
+    videos = [p for p in files if p.suffix.lower() not in IMAGE_EXTENSIONS]
+    if not videos:
+        raise PipelineError("yt-dlp produced no video file")
+    video = max(videos, key=lambda p: p.stat().st_size)
+    return video, thumbs[0] if thumbs else None
+
+
+def _ydl_args(cfg: Config, dest: Path, client: str, url: str) -> list[str]:
+    args = [
+        "--output", str(dest / "source_video.%(ext)s"),
+        "--format", "bestvideo*+bestaudio/best",
+        "--write-thumbnail", "--no-playlist", "--no-progress",
+        "--retries", "10", "--fragment-retries", "10",
+        "--extractor-retries", "5",
+    ]
+    if client != "default":
+        args += ["--extractor-args", f"youtube:player_client={client}"]
+    if cfg.ytdlp_pot_provider_url:
+        args += ["--extractor-args",
+                 f"youtubepot-bgutilhttp:base_url={cfg.ytdlp_pot_provider_url}"]
+    if cfg.ytdlp_cookies:
+        args += ["--cookies", str(cfg.ytdlp_cookies)]
+    if cfg.ytdlp_impersonate:
+        args += ["--impersonate", cfg.ytdlp_impersonate]
+    return [*args, "--", url]
+
+
+def download_youtube(url: str, dest: Path, cfg: Config,
+                     log=print) -> tuple[Path, Path | None]:
+    clients = list(cfg.ytdlp_clients) or ["default"]
+    tried: list[str] = []
+    last: Exception | None = None
+    # If every copy fails, look for a release newer than the daily check found.
+    for refresh in (False, True):
+        for ydl in ytdlp.candidates(cfg, log, refresh):
+            if ydl.version in tried:
+                continue
+            tried.append(ydl.version)
+            for client in clients:
+                try:
+                    ydl.run(_ydl_args(cfg, dest, client, url), log)
+                    return _classify_downloads(dest)
+                except PipelineError as e:
+                    last = e
+                    log(f"yt-dlp {ydl.version} player_client={client} "
+                        f"failed: {e}")
+    raise PipelineError(
+        f"yt-dlp {', '.join(tried)} could not download {url} after trying "
+        f"player_client={', '.join(clients)}. Last error: {last}. "
+        "See the yt-dlp options in config.toml (cookies, impersonation, "
+        "PO token provider).")
+
+
+def _run(ctx: Ctx) -> None:
+    if bool(ctx.inp.youtube_url) == bool(ctx.inp.video_path):
+        raise PipelineError("provide exactly one of youtube_url / video_path")
+    thumb = None
+    if ctx.inp.youtube_url:
+        video, thumb = download_youtube(ctx.inp.youtube_url, ctx.work,
+                                       ctx.cfg, ctx.log)
+    else:
+        video = ctx.work / f"source_video{ctx.inp.video_path.suffix}"
+        shutil.copyfile(ctx.inp.video_path, video)
+    ffmpeg("-i", str(video), "-vn", "-ac", "2", "-ar", "44100",
+           str(ctx.work / "audio.wav"))
+    meta = {"video_file": video.name, "thumbnail": thumb.name if thumb else None}
+    (ctx.work / "meta.json").write_text(json.dumps(meta))
+
+
+def acquire_stage() -> Stage:
+    return Stage(
+        name="acquire",
+        outputs=lambda ctx: [ctx.work / "meta.json", ctx.work / "audio.wav"],
+        run=_run,
+    )
@@ -1,12 +1,15 @@
 import json
 from pathlib import Path

-from ..lyrics import RefLyrics, alignment_token, normalize_lyrics
+from ..lyrics import alignment_token, normalize_lyrics
+from ..models import on_release
 from ..pipeline import Ctx, PipelineError, Stage
-from .transcribe import load_transcript, lyrics_similarity, vocal_sections
+from .transcribe import classify_lyrics_match, load_transcript, lyrics_similarity, vocal_sections

 SIMILARITY_THRESHOLD = 0.4

+# a stage aligns section by section, so the model is cached across the calls
+# within one run and dropped at the end of it
 _model_cache = {}


@@ -17,13 +20,20 @@ def _get_model(device: str):
    return _model_cache[device]


+@on_release
+def unload_model() -> None:
+    _model_cache.clear()
+
+
 def force_align_section(vocals_path: Path, start: float, end: float,
                        tokens: list[str], device: str) -> list[tuple[float, float, float]]:
    import torch
    import torchaudio
    from torchaudio.pipelines import MMS_FA as bundle

-    waveform, sr = torchaudio.load(str(vocals_path))
+    from ..audio_io import load_wav
+
+    waveform, sr = load_wav(vocals_path)
    waveform = waveform.mean(0, keepdim=True)
    lo, hi = int(start * sr), min(int(end * sr), waveform.shape[1])
    waveform = torchaudio.functional.resample(
@@ -40,16 +50,15 @@ def force_align_section(vocals_path: Path, start: float, end: float,
             sum(t.score for t in s) / len(s)) for s in spans]


-def _assign_lines_to_sections(ref: RefLyrics, sections) -> list[list[int]]:
-    """Split lyric lines across vocal sections proportionally to section duration."""
+def _assign_lines_to_sections(line_indices: list[int], sections) -> list[list[int]]:
+    """Split matched lyric line indices across vocal sections proportionally to section duration."""
    total = sum(b - a for a, b in sections) or 1.0
-    n_lines = len(ref.lines)
+    n = len(line_indices)
    groups, elapsed, start = [], 0.0, 0
    for i, (a, b) in enumerate(sections):
        elapsed += b - a
-        stop = n_lines if i == len(sections) - 1 else min(
-            n_lines, round(n_lines * elapsed / total))
-        groups.append(list(range(start, max(stop, start))))
+        stop = n if i == len(sections) - 1 else min(n, round(n * elapsed / total))
+        groups.append(line_indices[start:max(stop, start)])
        start = max(stop, start)
    return groups

@@ -97,7 +106,7 @@ def _align_section(vocals: Path, sec: tuple[float, float], words: list[tuple[str

 def _run(ctx: Ctx) -> None:
    ref = normalize_lyrics(ctx.inp.lyrics_text)
-    segments = load_transcript(ctx)
+    segments = sorted(load_transcript(ctx), key=lambda s: s.start)
    sim = lyrics_similarity(segments, ref)
    if sim < SIMILARITY_THRESHOLD:
        raise PipelineError(
@@ -105,15 +114,21 @@ def _run(ctx: Ctx) -> None:
            "wrong lyrics or wrong recording?")
    # normalised so an inverted span can never collapse a section onto one instant
    sections = [(min(a, b), max(a, b)) for a, b in vocal_sections(segments)]
+    dropped_lines, missing_spans = classify_lyrics_match(segments, ref)
+    matched = [i for i in range(len(ref.lines)) if i not in set(dropped_lines)]
    vocals = ctx.work / "vocals.wav"
    rows = []
-    for sec, line_idxs in zip(sections, _assign_lines_to_sections(ref, sections)):
+    for sec, line_idxs in zip(sections, _assign_lines_to_sections(matched, sections)):
        words = [(w, li) for li in line_idxs for w in ref.lines[li]]
        if words:
            rows.extend(_align_section(vocals, sec, words, ctx.inp.language,
                                       ctx.cfg.device))
    _enforce_monotonic(rows)
    (ctx.work / "alignment.json").write_text(json.dumps(rows))
+    (ctx.work / "align_meta.json").write_text(json.dumps({
+        "dropped_lines": dropped_lines,
+        "missing_spans": [list(s) for s in missing_spans],
+    }))


 def align_stage() -> Stage:
@@ -3,6 +3,7 @@ from dataclasses import asdict

 import numpy as np

+from ..lyrics import normalize_lyrics
 from ..notes import build_notes
 from ..pipeline import Ctx, Stage

@@ -10,8 +11,15 @@ from ..pipeline import Ctx, Stage
 def _run(ctx: Ctx) -> None:
    rows = json.loads((ctx.work / "alignment.json").read_text())
    data = np.load(ctx.work / "pitch.npz")
+    meta_path = ctx.work / "align_meta.json"
+    meta = json.loads(meta_path.read_text()) if meta_path.exists() else {}
+    dropped_lines = meta.get("dropped_lines", [])
+    missing_spans = [tuple(s) for s in meta.get("missing_spans", [])]
+    ref_lines = normalize_lyrics(ctx.inp.lyrics_text).lines
    notes, breaks, warnings = build_notes(
-        rows, data["times"], data["f0"], data["conf"], ctx.inp.language)
+        rows, data["times"], data["f0"], data["conf"], ctx.inp.language,
+        dropped_lines=dropped_lines, missing_spans=missing_spans,
+        ref_lines=ref_lines)
    (ctx.work / "notes_sec.json").write_text(json.dumps({
        "notes": [asdict(n) for n in notes],
        "breaks": breaks,
@@ -51,11 +51,15 @@ def _run(ctx: Ctx) -> None:
    meta = json.loads((ctx.work / "meta.json").read_text())
    src_video = ctx.work / meta["video_file"]
    fmt, codec = ffprobe_codec(src_video)
+    # The video is background only: the song's audio is #MP3, and the two are
+    # the same recording. A player that does not mute the video plays both at
+    # once, which is heard as an echo, so the audio track is dropped here.
    if (fmt, codec) in GOOD_VIDEO:
-        shutil.copyfile(src_video, out / f"{slug}{src_video.suffix}")
+        ffmpeg("-i", str(src_video), "-c", "copy", "-an",
+               str(out / f"{slug}{src_video.suffix}"))
    else:
        ffmpeg("-i", str(src_video), "-c:v", "libx264", "-crf", "20",
-               "-preset", "medium", "-c:a", "aac", str(out / f"{slug}.mp4"))
+               "-preset", "medium", "-an", str(out / f"{slug}.mp4"))
        _set_header(txt, "VIDEO", f"{slug}.mp4")
    cover_src = ctx.inp.cover_path
    if cover_src is None and meta.get("thumbnail"):
@@ -1,15 +1,30 @@
+import sys
 from pathlib import Path

 import numpy as np

+from ..models import on_release
 from ..pipeline import Ctx, Stage


+@on_release
+def unload_model() -> None:
+    """torchcrepe caches its model as attributes on its own infer function."""
+    torchcrepe = sys.modules.get("torchcrepe")
+    if torchcrepe is None:
+        return
+    for attr in ("model", "capacity"):
+        if hasattr(torchcrepe.infer, attr):
+            delattr(torchcrepe.infer, attr)
+
+
 def track_pitch(vocals_path: Path, model: str, device: str):
    import torchaudio
    import torchcrepe

-    waveform, sr = torchaudio.load(str(vocals_path))
+    from ..audio_io import load_wav
+
+    waveform, sr = load_wav(vocals_path)
    waveform = waveform.mean(0, keepdim=True)
    waveform = torchaudio.functional.resample(waveform, sr, 16000)
    f0, periodicity = torchcrepe.predict(
@@ -0,0 +1,99 @@
+import difflib
+import json
+
+from ..lyrics import RefLyrics, alignment_token
+from ..pipeline import Ctx, Stage
+from ..whisper_client import TranscriptSegment, transcribe
+
+
+def vocal_sections(segments, min_gap: float = 6.0, pad: float = 1.0):
+    sections = []
+    for s in sorted(segments, key=lambda s: s.start):
+        if sections and s.start - sections[-1][1] < min_gap:
+            sections[-1][1] = max(sections[-1][1], s.end)
+        else:
+            sections.append([s.start, s.end])
+    return [(max(0.0, a - pad), b + pad) for a, b in sections]
+
+
+def _tokens(text: str) -> list[str]:
+    return [t for w in text.split() if (t := alignment_token(w))]
+
+
+def match_lyrics(segments, ref: RefLyrics):
+    heard, seg_of = [], []
+    for si, s in enumerate(segments):
+        for t in _tokens(s.text):
+            heard.append(t)
+            seg_of.append(si)
+    pairs = [(t, li) for w, li in zip(ref.words, ref.word_line_index)
+             if (t := alignment_token(w))]
+    expected = [t for t, _ in pairs]
+    line_of = [li for _, li in pairs]
+    matcher = difflib.SequenceMatcher(a=heard, b=expected)
+    return matcher, seg_of, line_of
+
+
+def lyrics_similarity(segments, ref: RefLyrics) -> float:
+    matcher, _, _ = match_lyrics(segments, ref)
+    return matcher.ratio()
+
+
+def classify_lyrics_match(segments, ref: RefLyrics, pad: float = 1.0
+                          ) -> tuple[list[int], list[tuple[float, float]]]:
+    """Split reference lines into matched vs. dropped, and flag sung spans no
+    reference line explains.
+
+    dropped_lines: reference line indices covered only by 'insert' opcodes
+    (present in the lyrics, never heard). missing_spans: the padded, merged
+    bounds of the heard segments behind 'delete' opcodes (heard, but matching
+    no reference word). A segment can also hold matched lines, so a span
+    marks where to look, not words to discard.
+    """
+    matcher, seg_of, line_of = match_lyrics(segments, ref)
+    covered = [False] * len(line_of)
+    for tag, _, _, bj1, bj2 in matcher.get_opcodes():
+        if tag in ("equal", "replace"):
+            for j in range(bj1, bj2):
+                covered[j] = True
+    by_line = {}
+    for j, li in enumerate(line_of):
+        by_line.setdefault(li, []).append(covered[j])
+    dropped_lines = sorted(li for li, cov in by_line.items() if not any(cov))
+
+    spans = []
+    for tag, ai1, ai2, _, _ in matcher.get_opcodes():
+        if tag != "delete":
+            continue
+        sis = [seg_of[i] for i in range(ai1, ai2)]
+        start = max(0.0, min(segments[si].start for si in sis) - pad)
+        end = max(segments[si].end for si in sis) + pad
+        spans.append((start, end))
+    return dropped_lines, _merge_spans(spans)
+
+
+def _merge_spans(spans: list[tuple[float, float]]) -> list[tuple[float, float]]:
+    merged: list[list[float]] = []
+    for a, b in sorted(spans):
+        if merged and a <= merged[-1][1]:
+            merged[-1][1] = max(merged[-1][1], b)
+        else:
+            merged.append([a, b])
+    return [tuple(s) for s in merged]
+
+
+def _run(ctx: Ctx) -> None:
+    segments = transcribe(ctx.cfg, ctx.work / "vocals.wav", ctx.inp.language)
+    (ctx.work / "transcript.json").write_text(json.dumps(
+        [{"start": s.start, "end": s.end, "text": s.text} for s in segments]))
+
+
+def transcribe_stage() -> Stage:
+    return Stage(name="transcribe",
+                 outputs=lambda ctx: [ctx.work / "transcript.json"],
+                 run=_run)
+
+
+def load_transcript(ctx: Ctx) -> list[TranscriptSegment]:
+    data = json.loads((ctx.work / "transcript.json").read_text())
+    return [TranscriptSegment(**d) for d in data]
@@ -4,7 +4,7 @@
 <meta charset="utf-8">
 <meta name="viewport" content="width=device-width, initial-scale=1">
 <meta name="color-scheme" content="dark">
-<title>karaoke-maker</title>
+<title>songtrace</title>
 <style>
@font-face {
  font-family: "Chakra Petch";
@@ -111,6 +111,7 @@ body {
 }

 #sidebar .eyebrow, #jobs .empty-note { padding: 0 18px; }
+#create-song { display: block; width: calc(100% - 36px); margin: 0 18px 20px; padding: 9px 12px; }

 #jobs { display: flex; flex-direction: column; gap: 1px; margin-top: 8px; }

@@ -194,8 +195,13 @@ body {
 }

 #song-form { display: grid; gap: 18px; }
+.song-form-grid { display: grid; grid-template-columns: 1fr 1fr; gap: 18px; }
+.song-form-main { display: grid; gap: 18px; align-content: start; }
+.song-form-lyrics textarea { height: 100%; flex: 1 1 auto; }
 .pair { display: grid; grid-template-columns: 1fr 1fr; gap: 18px; }
 .field { display: flex; flex-direction: column; gap: 6px; }
+/* [hidden] is only display:none, which the rule above would otherwise win */
+.field[hidden], .keep-note[hidden] { display: none; }
 .field > span {
  font-family: var(--font-display);
  font-weight: 600;
@@ -220,7 +226,6 @@ input:not([type="file"]):hover { border-bottom-color: var(--text-muted); }
 input:not([type="file"]):focus { border-bottom-color: var(--aqua); }

 textarea {
-  min-height: 8.5rem;
  padding: 11px 13px;
  line-height: 1.6;
  background: var(--panel);
@@ -229,6 +234,7 @@ textarea {
 }
 textarea:hover { border-color: var(--text-muted); }
 textarea:focus { border-color: var(--aqua); }
+.song-form-lyrics textarea { min-height: 8.5rem; }

 input[type="file"] {
  font-size: 0.82rem;
@@ -304,6 +310,7 @@ button { cursor: pointer; }
  transition: background 120ms ease;
 }
 .primary:hover { background: #BE7BF9; }
+.primary:disabled { background: var(--control); color: var(--text-muted); cursor: not-allowed; }

 #form-error { margin: 12px 0 0; min-height: 1.2em; font-size: 0.85rem; color: var(--fail); }

@@ -316,20 +323,47 @@ button { cursor: pointer; }
 }
 #review h3 span { color: var(--text-muted); font-weight: 400; }

-#warnings { display: flex; flex-direction: column; gap: 1px; }
-.warn-row {
-  display: grid;
-  grid-template-columns: auto auto 1fr;
-  align-items: center;
-  gap: 10px;
-  padding: 10px 12px;
-  background: var(--panel);
+/* sticks to the top of #main so the video stays in frame as the lyrics scroll */
+#player { position: sticky; top: 0; z-index: 2; margin: 0 0 14px; padding: 6px 0 10px; background: var(--bg); }
+#player-video {
+  display: block;
+  width: 100%;
+  max-height: 30vh;
+  background: #000;
  border: 1px solid var(--border);
+  border-radius: 2px;
+}
+#player-video[hidden] { display: none; }
+/* the native transport renders light otherwise, against everything around it */
+#player-media { display: block; width: 100%; margin-top: 8px; color-scheme: dark; }
+
+#lyrics-review { display: flex; flex-direction: column; gap: 10px; }
+.review-line { padding: 8px 12px; background: var(--panel); border: 1px solid var(--border); }
+.review-line-active { background: var(--panel-raised); border-color: var(--aqua); }
+.review-word-sung { color: var(--text-muted); }
+.review-word-active { color: var(--aqua); font-weight: 600; }
+.review-playhead { position: absolute; top: 0; bottom: 0; width: 2px; margin-left: -1px; background: var(--aqua); }
+.review-seek { cursor: pointer; }
+.review-line-text { display: block; font-size: 0.9rem; }
+.review-line-dropped { opacity: 0.55; }
+.review-line-dropped .review-line-text { text-decoration: line-through; }
+.review-line-note { display: block; margin-top: 4px; font-size: 0.72rem; color: var(--text-muted); }
+.review-line-timeline { position: relative; margin-top: 4px; }
+.review-lane { position: absolute; top: 0; bottom: 0; }
+.review-lane-alt { background: var(--panel-raised); }
+.review-word-row { position: relative; height: 18px; }
+.review-word { position: absolute; top: 0; white-space: nowrap; font-size: 0.9rem; }
+.review-roll { position: relative; display: block; width: 100%; height: 20px; margin-top: 4px; }
+.review-note { fill: var(--aqua); }
+.review-note-low { fill: var(--warn); }
+.review-missing {
+  margin: 0;
+  padding: 6px 12px;
+  font-size: 0.78rem;
+  color: var(--warn);
+  background: var(--panel);
+  border: 1px dashed var(--border);
 }
-.warn-row + .warn-row { margin-top: 6px; }
-.warn-icon { color: var(--warn); }
-.warn-count { font-size: 0.85rem; }
-.warn-range { justify-self: end; font-family: var(--font-mono); font-size: 0.82rem; color: var(--text-muted); }

 .review-error { margin: 0 0 14px; font-size: 0.85rem; color: var(--fail); }

@@ -343,7 +377,7 @@ button { cursor: pointer; }
  border: 1px solid var(--border);
 }
 .output-path code { flex: 1; overflow-wrap: anywhere; font-family: var(--font-mono); font-size: 0.82rem; }
-#copy-path, .action, #cancel-rebuild {
+#copy-path, .action {
  flex: none;
  padding: 5px 12px;
  font-family: var(--font-display);
@@ -356,10 +390,11 @@ button { cursor: pointer; }
  border: 1px solid var(--control);
  border-radius: 2px;
 }
-#copy-path:hover, .action:hover, #cancel-rebuild:hover { border-color: var(--text-muted); }
+#copy-path:hover, .action:hover { border-color: var(--text-muted); }

 .form-actions { display: flex; align-items: center; gap: 14px; }
 .keep-note { margin: -8px 0 0; font-size: 0.78rem; color: var(--aqua); }
+#form-lock { margin: 0; font-size: 0.78rem; color: var(--text-muted); }
 #review-actions { display: flex; gap: 10px; margin-top: 14px; }
 #review-lineage { margin: -10px 0 14px; }
 .lineage {
@@ -464,6 +499,11 @@ button { cursor: pointer; }
  #stage-track { margin-left: 0; flex-wrap: wrap; }
 }

+@media (max-width: 720px) {
+  .song-form-grid { grid-template-columns: 1fr; }
+  .song-form-lyrics textarea { height: auto; }
+}
+
@media (prefers-reduced-motion: reduce) {
  *, *::before, *::after { animation: none !important; transition: none !important; }
 }
@@ -473,10 +513,11 @@ button { cursor: pointer; }

 <aside id="sidebar">
  <div class="brand">
-    <h1 class="wordmark"><span>KARAOKE</span> <span class="second">MAKER</span></h1>
+    <h1 class="wordmark"><span>SONG</span> <span class="second">TRACE</span></h1>
    <div class="kana" aria-hidden="true">カラオケ</div>
  </div>
  <div>
+    <button type="button" id="create-song" class="primary">CREATE SONG</button>
    <h2 class="eyebrow">QUEUE</h2>
    <div id="jobs"></div>
  </div>
@@ -486,32 +527,40 @@ button { cursor: pointer; }
  <section id="new-song">
    <h2>NEW SONG</h2>
    <form id="song-form">
-      <div class="pair">
-        <label class="field"><span>Title</span><input name="title" required></label>
-        <label class="field"><span>Artist</span><input name="artist" required></label>
-      </div>
-      <label class="field"><span>Lyrics</span><textarea name="lyrics" required></textarea></label>
-      <details>
-        <summary>Advanced</summary>
-        <div class="stack">
+      <div class="song-form-grid">
+        <div class="song-form-main">
+          <div class="pair">
+            <label class="field"><span>Title</span><input name="title" required></label>
+            <label class="field"><span>Artist</span><input name="artist" required></label>
+          </div>
          <label class="field"><span>YouTube URL</span><input name="youtube_url"></label>
          <label class="field"><span>Video file</span><input type="file" name="video"></label>
          <p class="keep-note" data-keep="video" hidden></p>
-          <div class="pair">
-            <label class="field"><span>Genre</span><input name="genre"></label>
-            <label class="field"><span>Year</span><input name="year" type="number"></label>
-          </div>
-          <div class="pair">
-            <label class="field"><span>Language</span><input name="language"></label>
-            <label class="field"><span>BPM override</span><input name="bpm" type="number" step="any"></label>
-          </div>
-          <label class="field"><span>Cover image</span><input type="file" name="cover"></label>
-          <p class="keep-note" data-keep="cover" hidden></p>
+          <label class="field" id="restart-field" hidden>
+            <span>Restart from</span>
+            <select name="force_from"></select>
+          </label>
+          <details>
+            <summary>Advanced</summary>
+            <div class="stack">
+              <div class="pair">
+                <label class="field"><span>Genre</span><input name="genre"></label>
+                <label class="field"><span>Year</span><input name="year" type="number"></label>
+              </div>
+              <div class="pair">
+                <label class="field"><span>Language</span><input name="language"></label>
+                <label class="field"><span>BPM override</span><input name="bpm" type="number" step="any"></label>
+              </div>
+              <label class="field"><span>Cover image</span><input type="file" name="cover"></label>
+              <p class="keep-note" data-keep="cover" hidden></p>
+            </div>
+          </details>
        </div>
-      </details>
+        <label class="field song-form-lyrics"><span>Lyrics</span><textarea name="lyrics" required></textarea></label>
+      </div>
      <div class="form-actions">
        <button type="submit" class="primary">CREATE SONG</button>
-        <button type="button" id="cancel-rebuild" hidden>CANCEL</button>
+        <p id="form-lock" hidden></p>
      </div>
    </form>
    <p id="form-error" role="alert"></p>
@@ -522,7 +571,11 @@ button { cursor: pointer; }
    <h3 id="review-title"></h3>
    <p id="review-lineage" hidden></p>
    <p id="review-error" class="review-error" hidden></p>
-    <div id="warnings"><p class="empty-note">no songs queued yet — nothing to review</p></div>
+    <div id="player" hidden>
+      <video id="player-video" muted playsinline preload="metadata"></video>
+      <audio id="player-media" controls preload="metadata"></audio>
+    </div>
+    <div id="lyrics-review"><p class="empty-note">no songs queued yet — nothing to review</p></div>
    <div id="output-path"></div>
    <div id="review-actions"></div>
    <details>
@@ -552,17 +605,20 @@ const nowTitle = document.getElementById("now-title");
 const stageTrack = document.getElementById("stage-track");
 const reviewActions = document.getElementById("review-actions");
 const reviewLineage = document.getElementById("review-lineage");
-const cancelRebuild = document.getElementById("cancel-rebuild");
+const createSong = document.getElementById("create-song");
+const formLock = document.getElementById("form-lock");
 const formHeading = document.querySelector("#new-song h2");
 const submitButton = form.querySelector("button[type=submit]");
 let rebuildOf = null;
+let formDirty = false;
+// set by CREATE SONG: the page is a blank sheet on purpose, so the poll must
+// not quietly select the newest song back into it
+let composing = false;
 let JOBS = [];
 let selectedId = null;
 let STAGES = [];
 let pollTimer = null;

-const WARN_RE = /^(\d+)\s+low-confidence notes?\s+near\s+(\d+:\d{2})–(\d+:\d{2})/;
-
 const STATE = {
  queued:  { glyph: "○", cls: "st-queued"  },
  running: { glyph: "◈", cls: "st-running" },
@@ -592,10 +648,16 @@ form.addEventListener("submit", async (e) => {
  if (!resp.ok) {
    const body = await resp.json().catch(() => ({}));
    formError.textContent = body.detail || `error: ${resp.status}`;
-    return;  // before the reset below, so a rejected submission is never retyped
+    return;
  }
  selectedId = (await resp.json()).id;
-  exitRebuild();
+  composing = false;
+  // the fields stay as submitted: the song just sent is the obvious starting
+  // point for the next one, and nothing has to be retyped to correct it. The
+  // queue's CREATE SONG button is what empties the form.
+  rebuildOf = null;
+  formDirty = false;
+  setFormMode(null, null);
  refreshJobs();
 });

@@ -620,7 +682,25 @@ for (const input of form.querySelectorAll('input[type="file"]')) {
  });
 }

-cancelRebuild.addEventListener("click", exitRebuild);
+// selecting another queue entry reloads the form, so track whether there is
+// anything of the operator's in there worth warning about first
+form.addEventListener("input", () => { formDirty = true; });
+
+// Always available, and the form is usually holding a selected song by now, so
+// this clears the whole page down to a blank sheet — fields, review, player and
+// queue selection — and brings the form back into view from wherever playback
+// scrolled the panel to. Leaving another song's review under an empty form
+// reads as though the two belong together.
+createSong.addEventListener("click", () => {
+  if (formDirty && !confirm("Discard your unsaved edits to this form?")) return;
+  composing = true;
+  selectedId = null;
+  for (const card of jobsDiv.querySelectorAll(".job")) card.removeAttribute("aria-current");
+  exitRebuild();
+  clearDetail("nothing selected");
+  document.getElementById("new-song").scrollIntoView({ block: "start", behavior: "smooth" });
+  form.elements.title.focus();
+});

 function jobCard(job) {
  const card = document.createElement("button");
@@ -695,6 +775,11 @@ function renderJobs(jobs) {
 async function loadStages() {
  const resp = await fetch("/api/stages");
  STAGES = (await resp.json()).stages;
+  // a rebuild redoes only what the edit invalidated, which cannot see a change
+  // in songtrace itself — this is how you redo a stage anyway
+  const select = form.elements.force_from;
+  select.replaceChildren(new Option("only what changed", ""),
+                         ...STAGES.map(s => new Option(s, s)));
 }

 function renderStageTrack(job) {
@@ -754,19 +839,28 @@ async function refreshJobs() {
  const resp = await fetch("/api/jobs");
  const jobs = await resp.json();
  JOBS = jobs;
-  if (selectedId === null && jobs.length) selectedId = jobs[0].id;
+  // deliberately not via selectJob: the auto-selection on load fills the review
+  // panel but leaves the form blank, so landing on the page is still the way to
+  // start a new song. Clicking a queue entry is what loads its inputs.
+  if (!composing && selectedId === null && jobs.length) selectedId = jobs[0].id;
  renderJobs(jobs);
  renderNowBar(jobs);
  if (selectedId) refreshDetail();
+  else clearDetail(composing ? "nothing selected" : undefined);
  schedulePoll(jobs);
 }

 async function selectJob(id) {
+  // the form already holds this job; reloading it would drop edits in progress
+  const reload = id !== rebuildOf;
+  if (reload && formDirty && !confirm("Discard your unsaved edits to this form?")) return;
+  composing = false;
  selectedId = id;
  for (const card of jobsDiv.querySelectorAll(".job")) {
    if (card.dataset.id === id) card.setAttribute("aria-current", "true");
    else card.removeAttribute("aria-current");
  }
+  if (reload) await loadInputs(id, JOBS.find(j => j.id === id)?.status);
  refreshDetail();
 }

@@ -777,44 +871,327 @@ function emptyNote(text) {
  return p;
 }

-function renderWarnings(job) {
-  const box = document.getElementById("warnings");
+const SVG_NS = "http://www.w3.org/2000/svg";
+
+function fmtTime(t) {
+  const s = Math.round(t);
+  return `${Math.floor(s / 60)}:${String(s % 60).padStart(2, "0")}`;
+}
+
+function missingSpanNode(span) {
+  const p = document.createElement("p");
+  p.className = "review-missing";
+  p.textContent = `♪ unmatched singing, ${fmtTime(span.start)}–${fmtTime(span.end)}`;
+  return p;
+}
+
+function pianoRoll(notes, minPitch, span, t0, t1) {
+  const svg = document.createElementNS(SVG_NS, "svg");
+  svg.setAttribute("class", "review-roll");
+  svg.setAttribute("viewBox", "0 0 100 20");
+  svg.setAttribute("preserveAspectRatio", "none");
+  const dur = Math.max(t1 - t0, 0.001);
+  for (const n of notes) {
+    const rect = document.createElementNS(SVG_NS, "rect");
+    const x = ((n.start - t0) / dur) * 100;
+    const w = Math.max(((n.end - n.start) / dur) * 100, 0.5);
+    const y = 18 - ((n.pitch - minPitch) / span) * 16;
+    rect.setAttribute("x", x.toFixed(2));
+    rect.setAttribute("width", w.toFixed(2));
+    rect.setAttribute("y", y.toFixed(2));
+    rect.setAttribute("height", "2");
+    rect.setAttribute("class", n.low_confidence ? "review-note review-note-low" : "review-note");
+    svg.append(rect);
+  }
+  return svg;
+}
+
+// Lays words out along a shared time axis (like a karaoke lyric line) so
+// each word's notes render directly beneath it instead of as one flat strip.
+function lineTimeline(words, notes, minPitch, span) {
+  const t0 = words[0].start;
+  const t1 = words[words.length - 1].end;
+  const dur = Math.max(t1 - t0, 0.001);
+  const box = document.createElement("div");
+  box.className = "review-line-timeline";
+  words.forEach((w, i) => {
+    const lane = document.createElement("div");
+    lane.className = "review-lane" + (i % 2 ? " review-lane-alt" : "");
+    lane.style.left = `${((w.start - t0) / dur) * 100}%`;
+    lane.style.width = `${Math.max(((w.end - w.start) / dur) * 100, 0.5)}%`;
+    box.append(lane);
+  });
+  const wordsRow = document.createElement("div");
+  wordsRow.className = "review-word-row";
+  for (const w of words) {
+    const label = document.createElement("span");
+    label.className = "review-word";
+    label.style.left = `${((w.start - t0) / dur) * 100}%`;
+    label.textContent = w.text;
+    wordsRow.append(label);
+  }
+  box.append(wordsRow);
+  if (notes.length) box.append(pianoRoll(notes, minPitch, span, t0, t1));
+  const playhead = document.createElement("div");
+  playhead.className = "review-playhead";
+  playhead.hidden = true;
+  box.append(playhead);
+  return box;
+}
+
+function lineNode(line, minPitch, span) {
+  const row = document.createElement("div");
+  row.className = "review-line" + (line.dropped ? " review-line-dropped" : "");
+  if (line.dropped) {
+    const text = document.createElement("span");
+    text.className = "review-line-text";
+    text.textContent = line.text;
+    row.append(text);
+    const note = document.createElement("span");
+    note.className = "review-line-note";
+    note.textContent = "not sung — dropped";
+    row.append(note);
+    return row;
+  }
+  if (line.words.length) {
+    row.append(lineTimeline(line.words, line.notes, minPitch, span));
+    return row;
+  }
+  const text = document.createElement("span");
+  text.className = "review-line-text";
+  text.textContent = line.text;
+  row.append(text);
+  if (line.notes.length) {
+    const t0 = line.notes[0].start;
+    const t1 = Math.max(...line.notes.map(n => n.end));
+    row.append(pianoRoll(line.notes, minPitch, span, t0, t1));
+  }
+  return row;
+}
+
+// What playback needs from a rendered line: the time range it was laid out
+// over, the word labels along it, and the bar that sweeps across it.
+function lineEntry(row, line) {
+  const laid = line.words.length ? line.words : line.notes;
+  if (!laid.length || line.dropped) return null;
+  const labels = [...row.querySelectorAll(".review-word")];
+  return {
+    row,
+    t0: laid[0].start,
+    t1: Math.max(...laid.map(x => x.end)),
+    words: line.words.map((w, i) => ({ el: labels[i], start: w.start, end: w.end })),
+    playhead: row.querySelector(".review-playhead"),
+  };
+}
+
+function buildLyricsReview(data) {
+  const allNotes = data.lines.flatMap(l => l.notes);
+  const pitches = allNotes.map(n => n.pitch);
+  const minPitch = pitches.length ? Math.min(...pitches) : 0;
+  const maxPitch = pitches.length ? Math.max(...pitches) : 0;
+  const span = Math.max(1, maxPitch - minPitch);
+  const spans = [...data.missing_spans].sort((a, b) => a.start - b.start);
+  const nodes = [];
+  const entries = [];
+  let spanIdx = 0;
+  for (const line of data.lines) {
+    const lineStart = line.notes.length ? Math.min(...line.notes.map(n => n.start)) : null;
+    while (spanIdx < spans.length && lineStart !== null && spans[spanIdx].start < lineStart) {
+      nodes.push(missingSpanNode(spans[spanIdx]));
+      spanIdx++;
+    }
+    const row = lineNode(line, minPitch, span);
+    const entry = lineEntry(row, line);
+    if (entry) entries.push(entry);
+    nodes.push(row);
+  }
+  while (spanIdx < spans.length) {
+    nodes.push(missingSpanNode(spans[spanIdx]));
+    spanIdx++;
+  }
+  return { nodes, entries };
+}
+
+// The song folder's own media, played against the same seconds the review was
+// built from, so the highlighted line is the one you are hearing.
+//
+// Modelled on the game: #MP3 is the audio and the clock, the video is muted
+// background slaved to it. Playing the video's own audio alongside the mp3 is
+// the same recording twice, a few frames apart, which is heard as an echo.
+const player = document.getElementById("player");
+const media = document.getElementById("player-media");
+const video = document.getElementById("player-video");
+const DRIFT = 0.25;  // seconds before the picture is worth re-seeking
+
+function slaveVideo(force) {
+  if (video.hidden || !video.getAttribute("src")) return;
+  if (force || Math.abs(video.currentTime - media.currentTime) > DRIFT) {
+    video.currentTime = media.currentTime;
+  }
+}
+
+for (const event of ["seeked", "ratechange"]) {
+  media.addEventListener(event, () => {
+    video.playbackRate = media.playbackRate;
+    slaveVideo(true);
+  });
+}
+media.addEventListener("pause", () => video.pause());
+let timeline = [];
+let activeEntry = null;
+let syncFrame = null;
+
+// the first line that has not finished yet: the one being sung, or the one
+// coming up while the track is between lines
+function entryAt(t) {
+  return timeline.find(e => t <= e.t1) ?? timeline[timeline.length - 1] ?? null;
+}
+
+function paintEntry(entry, t) {
+  for (const w of entry.words) {
+    w.el.classList.toggle("review-word-sung", t >= w.end);
+    w.el.classList.toggle("review-word-active", t >= w.start && t < w.end);
+  }
+  if (!entry.playhead) return;
+  const inside = t >= entry.t0 && t <= entry.t1;
+  entry.playhead.hidden = !inside;
+  if (inside) {
+    const frac = (t - entry.t0) / Math.max(entry.t1 - entry.t0, 0.001);
+    entry.playhead.style.left = `${frac * 100}%`;
+  }
+}
+
+function clearEntry(entry) {
+  entry.row.classList.remove("review-line-active");
+  if (entry.playhead) entry.playhead.hidden = true;
+  for (const w of entry.words) {
+    w.el.classList.remove("review-word-sung", "review-word-active");
+  }
+}
+
+function syncPlayback() {
+  const t = media.currentTime;
+  const entry = entryAt(t);
+  if (entry !== activeEntry) {
+    if (activeEntry) clearEntry(activeEntry);
+    activeEntry = entry;
+    if (entry) {
+      entry.row.classList.add("review-line-active");
+      focusActive();
+    }
+  }
+  if (entry) paintEntry(entry, t);
+}
+
+// centred so there is context either side, like a scrolling lyric sheet; only
+// while running, so it does not fight a reader who has scrolled away to look
+// at something else
+function focusActive() {
+  if (!activeEntry || media.paused) return;
+  activeEntry.row.scrollIntoView({ block: "center", behavior: "smooth" });
+}
+
+function pumpPlayback() {
+  syncPlayback();
+  slaveVideo(false);
+  syncFrame = media.paused ? null : requestAnimationFrame(pumpPlayback);
+}
+
+// timeupdate fires about four times a second, too coarse for a playhead
+media.addEventListener("play", () => {
+  if (syncFrame === null) syncFrame = requestAnimationFrame(pumpPlayback);
+  if (!video.hidden && video.getAttribute("src")) video.play().catch(() => {});
+  focusActive();  // the line may already be active from a seek while paused
+});
+for (const event of ["pause", "seeked", "loadedmetadata"]) {
+  media.addEventListener(event, syncPlayback);
+}
+
+function bindTimeline(entries) {
+  if (activeEntry) clearEntry(activeEntry);
+  activeEntry = null;
+  timeline = entries;
+  for (const entry of entries) {
+    entry.row.classList.add("review-seek");
+    entry.row.title = "click to play from here";
+    entry.row.addEventListener("click", (e) => {
+      const word = entry.words.find(w => w.el === e.target);
+      media.currentTime = word ? word.start : entry.t0;
+      if (media.paused) syncPlayback();
+    });
+  }
+}
+
+// Returns true when the folder gave us something to play. #MP3 is required —
+// it is the audio the game uses, so it is the one the review has to track.
+function setPlayerSource(job, available) {
+  const src = available?.audio ? `/api/jobs/${job.id}/media/audio` : "";
+  player.hidden = !src;
+  if (media.dataset.src === src) return !!src;
+  media.dataset.src = src;
+  if (src) media.src = src;
+  else { media.removeAttribute("src"); media.load(); }
+
+  video.hidden = !(src && available.video);
+  video.poster = available?.cover ? `/api/jobs/${job.id}/media/cover` : "";
+  if (!video.hidden) video.src = `/api/jobs/${job.id}/media/video`;
+  else { video.removeAttribute("src"); video.load(); }
+  return !!src;
+}
+
+function resetPlayer() {
+  // the 3s poll reaches here for every running job; reloading a cleared
+  // element each time would be pure churn
+  if (player.hidden && !timeline.length) return;
+  bindTimeline([]);
+  player.hidden = true;
+  media.pause();
+  video.pause();
+  media.dataset.src = "";
+  for (const el of [media, video]) {
+    el.removeAttribute("src");
+    el.removeAttribute("poster");
+    el.load();
+  }
+}
+
+async function renderLyricsReview(job) {
+  const box = document.getElementById("lyrics-review");
  if (job.status === "failed") {
+    delete box.dataset.jobId;
+    resetPlayer();
    box.replaceChildren(emptyNote("job failed — nothing to review"));
    return;
  }
  if (job.status !== "done") {
+    delete box.dataset.jobId;
+    resetPlayer();
    box.replaceChildren(emptyNote("waiting for this song to finish…"));
    return;
  }
-  if (!job.warnings.length) {
-    box.replaceChildren(emptyNote("no low-confidence regions — nothing to review"));
+  if (box.dataset.jobId === job.id) return;
+  const resp = await fetch(`/api/jobs/${job.id}/review`);
+  if (job.id !== selectedId) return;
+  if (!resp.ok) {
+    resetPlayer();
+    box.replaceChildren(emptyNote("no review data available"));
    return;
  }
-  const rows = job.warnings.map(w => {
-    const m = WARN_RE.exec(w);
-    const row = document.createElement("div");
-    row.className = "warn-row";
-    if (m) {
-      row.setAttribute("data-count", m[1]);
-      row.setAttribute("data-start", m[2]);
-      row.setAttribute("data-end", m[3]);
-      const icon = document.createElement("span");
-      icon.className = "warn-icon";
-      icon.textContent = "⚠";
-      const count = document.createElement("span");
-      count.className = "warn-count";
-      count.textContent = `${m[1]} note${m[1] === "1" ? "" : "s"}`;
-      const range = document.createElement("span");
-      range.className = "warn-range";
-      range.textContent = `${m[2]}–${m[3]}`;
-      row.append(icon, count, range);
-    } else {
-      row.textContent = w;
-    }
-    return row;
-  });
-  box.replaceChildren(...rows);
+  const data = await resp.json();
+  if (job.id !== selectedId) return;
+  box.dataset.jobId = job.id;
+  if (!data.lines.length) {
+    resetPlayer();
+    box.replaceChildren(emptyNote("no lyrics to review"));
+    return;
+  }
+  const { nodes, entries } = buildLyricsReview(data);
+  box.replaceChildren(...nodes);
+  // with nothing to play, a highlighted line and a click-to-seek cursor would
+  // both be promising something the song folder cannot deliver
+  const playable = setPlayerSource(job, data.media);
+  bindTimeline(playable ? entries : []);
+  if (playable) syncPlayback();
 }

 async function copyPath(path, button) {
@@ -874,25 +1251,63 @@ function renderOutputPath(job) {
  outputPathDiv.replaceChildren(wrap);
 }

-function setFormMode(inp) {
+function clearDetail(why = "no songs queued yet") {
+  reviewTitle.replaceChildren();
+  reviewError.hidden = true;
+  reviewError.textContent = "";
+  resetPlayer();
+  // the review box, not an element that no longer exists: getting this wrong
+  // threw partway through and left the rest of the panel showing the old song
+  const box = document.getElementById("lyrics-review");
+  delete box.dataset.jobId;
+  box.replaceChildren(emptyNote(`${why} — nothing to review`));
+  outputPathDiv.replaceChildren();
+  reviewActions.replaceChildren();
+  delete reviewActions.dataset.jobId;
+  reviewLineage.replaceChildren();
+  reviewLineage.hidden = true;
+  delete reviewLineage.dataset.jobId;
+  detail.textContent = `${why} — nothing to log`;
+}
+
+// a rebuild is refused with 409 until the job settles, so the form says so
+// rather than letting the submission fail
+function setSubmitLock(status) {
+  const locked = rebuildOf !== null && (status === "queued" || status === "running");
+  submitButton.disabled = locked;
+  formLock.textContent = locked ? "editable once this run finishes" : "";
+  formLock.hidden = !locked;
+}
+
+function setFormMode(inp, status) {
  formHeading.textContent = rebuildOf ? "REBUILD" : "NEW SONG";
  submitButton.textContent = rebuildOf ? "REBUILD SONG" : "CREATE SONG";
-  cancelRebuild.hidden = !rebuildOf;
+  setSubmitLock(status);
+  // only a rebuild has stages to restart from; a new song always runs them all
+  document.getElementById("restart-field").hidden = !rebuildOf;
+  if (!rebuildOf) form.elements.force_from.value = "";
  for (const [field, key] of [["video", "video_filename"], ["cover", "cover_filename"]]) {
    const note = form.querySelector(`[data-keep="${field}"]`);
    const kept = rebuildOf && inp ? inp[key] : null;
    note.textContent = kept ? `keeping ${kept} — attach a file to replace it` : "";
    note.hidden = !kept;
  }
+  // the advanced fields carry into a rebuild whether or not the section is open,
+  // so anything actually set must not sit hidden behind a closed summary
+  form.querySelector("details").open = !!inp?.cover_filename ||
+    ["genre", "year", "language", "bpm"].some(n => form.elements[n].value);
 }

 function exitRebuild() {
  rebuildOf = null;
  form.reset();
-  setFormMode(null);
+  formDirty = false;
+  setFormMode(null, null);
 }

-async function startRebuild(id) {
+// selecting a queue entry brings its inputs back up, so a song can be adjusted
+// and rebuilt without being retyped
+async function loadInputs(id, status) {
  const resp = await fetch(`/api/jobs/${id}/input`);
  if (!resp.ok) return;
  const inp = await resp.json();
@@ -903,9 +1318,9 @@ async function startRebuild(id) {
    if (el && value !== null) el.value = value;
  }
  rebuildOf = id;
-  setFormMode(inp);
+  formDirty = false;
+  setFormMode(inp, status);
  formError.textContent = "";
-  form.elements.title.focus();
 }

 async function rerunJob(id) {
@@ -920,21 +1335,60 @@ async function rerunJob(id) {
  refreshJobs();
 }

+async function rebuildFullJob(id) {
+  formError.textContent = "";
+  const resp = await fetch(`/api/jobs/${id}/rebuild-full`, { method: "POST" });
+  if (!resp.ok) {
+    const body = await resp.json().catch(() => ({}));
+    formError.textContent = body.detail || `error: ${resp.status}`;
+    return;
+  }
+  selectedId = (await resp.json()).id;
+  refreshJobs();
+}
+
+async function deleteJob(id) {
+  if (!confirm("Delete this song? This removes all its files and cannot be undone.")) return;
+  formError.textContent = "";
+  const resp = await fetch(`/api/jobs/${id}`, { method: "DELETE" });
+  if (!resp.ok) {
+    const body = await resp.json().catch(() => ({}));
+    formError.textContent = body.detail || `error: ${resp.status}`;
+    return;
+  }
+  if (selectedId === id) selectedId = null;
+  // the form is holding a job that no longer exists; a rebuild would 404
+  if (rebuildOf === id) exitRebuild();
+  refreshJobs();
+}
+
+const ACTION_LABELS = {
+  rerun: "RERUN",
+  "rebuild-full": "REBUILD FROM SCRATCH",
+  delete: "DELETE",
+};
+
+const ACTION_HANDLERS = {
+  rerun: rerunJob,
+  "rebuild-full": rebuildFullJob,
+  delete: deleteJob,
+};
+
 function actionButton(action, jobId) {
  const button = document.createElement("button");
  button.type = "button";
  button.className = "action";
  button.dataset.action = action;
-  button.textContent = action === "rerun" ? "RERUN" : "EDIT & REBUILD";
-  button.addEventListener("click",
-    () => action === "rerun" ? rerunJob(jobId) : startRebuild(jobId));
+  button.textContent = ACTION_LABELS[action];
+  button.addEventListener("click", () => ACTION_HANDLERS[action](jobId));
  return button;
 }

 function renderActions(job) {
-  // no controls mid-run: both endpoints answer 409 until the job settles
-  const wanted = job.status === "failed" ? ["rerun", "edit"]
-               : job.status === "done" ? ["edit"] : [];
+  // no controls mid-run: every endpoint answers 409 until the job settles
+  const doneActions = ["rebuild-full", "delete"];
+  const wanted = job.status === "failed" ? ["rerun", ...doneActions]
+               : job.status === "done" ? doneActions : [];
  const shown = [...reviewActions.children].map(b => b.dataset.action);
  // rebuilding an unchanged row would detach a focused action on every poll
  if (reviewActions.dataset.jobId === job.id && shown.join() === wanted.join()) return;
@@ -969,7 +1423,8 @@ async function refreshDetail() {
  reviewTitle.replaceChildren(`${job.title} `, artist);
  reviewError.hidden = !job.error;
  reviewError.textContent = job.error ?? "";
-  renderWarnings(job);
+  if (rebuildOf === job.id) setSubmitLock(job.status);
+  renderLyricsReview(job);
  renderOutputPath(job);
  renderLineage(job);
  renderActions(job);
@@ -1,3 +1,4 @@
+import json
 import math
 import queue
 import re
@@ -5,7 +6,7 @@ import shutil
 import threading
 import time
 import uuid
-from dataclasses import dataclass, field
+from dataclasses import asdict, dataclass, field
 from pathlib import Path

 from fastapi import FastAPI, File, Form, HTTPException, UploadFile
@@ -15,11 +16,15 @@ from fastapi.staticfiles import StaticFiles

 from . import rebuild as rebuild_policy
 from .config import Config
-from .pipeline import (STAGE_ORDER, PipelineError, SongInput, run_pipeline,
-                       song_slug)
+from .lyrics import normalize_lyrics
+from .pipeline import (STAGE_ORDER, PipelineError, SongInput, read_warnings,
+                       run_pipeline, song_slug)
+from .ultrastar import get_header, parse_ultrastar

 STATIC = Path(__file__).parent / "static"
 STAGE_LINE = re.compile(r"^\[(\w+)\] (?:running|cached, skipping)")
+# the song folder names its own media, exactly as the game reads it
+MEDIA_HEADERS = {"audio": "MP3", "video": "VIDEO", "cover": "COVER"}


@dataclass
@@ -35,6 +40,7 @@ class Job:
    error: str | None = None
    inp: SongInput | None = None
    rebuilt_from: str | None = None
+    force_stage: str | None = None
    created_at: float = field(default_factory=time.time)
    started_at: float | None = None
    finished_at: float | None = None
@@ -91,6 +97,35 @@ def _cleanup_uploads(cfg: Config, job_id: str, video_path: Path | None) -> None:
            pass


+def _discover_jobs(cfg: Config) -> list[Job]:
+    discovered = []
+    if not cfg.work_dir.exists():
+        return discovered
+    for sidecar in cfg.work_dir.glob("*/input.json"):
+        try:
+            data = json.loads(sidecar.read_text())
+            for key in ("video_path", "cover_path"):
+                if data.get(key) is not None:
+                    data[key] = Path(data[key])
+            inp = SongInput(**data)
+            slug = song_slug(inp)
+            out = cfg.output_dir / slug
+            done = out.exists() and any(out.iterdir())
+            mtime = sidecar.stat().st_mtime
+            warnings = read_warnings(sidecar.parent) if done else []
+        except (OSError, ValueError, TypeError):
+            continue
+        discovered.append(Job(
+            id=uuid.uuid4().hex, title=inp.title, artist=inp.artist, inp=inp,
+            status="done" if done else "failed",
+            output_dir=str(out) if done else None,
+            warnings=warnings,
+            error=None if done else "interrupted before completion",
+            created_at=mtime, finished_at=mtime))
+    discovered.sort(key=lambda j: j.created_at)
+    return discovered
+
+
 def create_app(cfg: Config, runner=run_pipeline) -> FastAPI:
    app = FastAPI()
    app.mount("/static", StaticFiles(directory=STATIC), name="static")
@@ -99,17 +134,24 @@ def create_app(cfg: Config, runner=run_pipeline) -> FastAPI:
    lock = threading.Lock()
    todo: queue.Queue[str] = queue.Queue()

+    for job in _discover_jobs(cfg):
+        jobs[job.id] = job
+        order.append(job.id)
+
    def _prepare_rebuild(job: Job, log) -> str | None:
        """Relocate cached work and report the stage this rebuild restarts from."""
        if not job.rebuilt_from:
-            return None
+            return job.force_stage
        with lock:
            parent = jobs.get(job.rebuilt_from)
        if parent is None:
-            return None
-        force_from = rebuild_policy.plan(parent.inp, job.inp)
+            return job.force_stage
+        planned = rebuild_policy.plan(parent.inp, job.inp)
        changed = rebuild_policy.changed_fields(parent.inp, job.inp)
        rebuild_policy.relocate_work(parent.inp, job.inp, cfg)
+        # a stage the operator asked for can only widen what plan() decided,
+        # which is how "redo the packaging" reaches work nothing invalidated
+        force_from = rebuild_policy.earliest(planned, job.force_stage)
        log(f"[rebuild] restarting from {force_from or 'the first incomplete stage'}"
            f" ({', '.join(changed) if changed else 'nothing'} changed)")
        return force_from
@@ -128,6 +170,15 @@ def create_app(cfg: Config, runner=run_pipeline) -> FastAPI:
        shutil.rmtree(old, ignore_errors=True)
        log(f"[rebuild] removed superseded {old.name}")

+    def _write_sidecar(job: Job) -> None:
+        work = cfg.work_dir / song_slug(job.inp)
+        work.mkdir(parents=True, exist_ok=True)
+        data = asdict(job.inp)
+        for key in ("video_path", "cover_path"):
+            if data[key] is not None:
+                data[key] = str(data[key])
+        (work / "input.json").write_text(json.dumps(data))
+
    def worker():
        while True:
            job_id = todo.get()
@@ -144,6 +195,7 @@ def create_app(cfg: Config, runner=run_pipeline) -> FastAPI:

            try:
                force_from = _prepare_rebuild(job, log)
+                _write_sidecar(job)
                result = runner(job.inp, cfg, log=log, force_from=force_from)
                with lock:
                    job.output_dir = str(result.output_dir)
@@ -164,10 +216,10 @@ def create_app(cfg: Config, runner=run_pipeline) -> FastAPI:

    threading.Thread(target=worker, daemon=True).start()

-    def _enqueue(job_id: str, inp: SongInput,
-                 rebuilt_from: str | None = None) -> dict:
+    def _enqueue(job_id: str, inp: SongInput, rebuilt_from: str | None = None,
+                 force_stage: str | None = None) -> dict:
        job = Job(id=job_id, title=inp.title, artist=inp.artist, inp=inp,
-                  rebuilt_from=rebuilt_from)
+                  rebuilt_from=rebuilt_from, force_stage=force_stage)
        with lock:
            jobs[job_id] = job
            order.append(job_id)
@@ -196,6 +248,7 @@ def create_app(cfg: Config, runner=run_pipeline) -> FastAPI:
                      cover: UploadFile | None = File(None)):
        youtube_url, genre, language = (youtube_url or None, genre or None,
                                        language or None)
+        lyrics = lyrics.replace("\r\n", "\n")
        has_video = bool(video and video.filename)
        if bool(youtube_url) == has_video:
            raise HTTPException(400, "provide exactly one of youtube_url / video")
@@ -216,6 +269,12 @@ def create_app(cfg: Config, runner=run_pipeline) -> FastAPI:
        parent = _finished_job(job_id)
        return _enqueue(uuid.uuid4().hex, parent.inp, rebuilt_from=job_id)

+    @app.post("/api/jobs/{job_id}/rebuild-full", status_code=201)
+    def rebuild_full(job_id: str):
+        parent = _finished_job(job_id)
+        return _enqueue(uuid.uuid4().hex, parent.inp, rebuilt_from=job_id,
+                        force_stage=STAGE_ORDER[0])
+
    def _require_source(parent_inp: SongInput, inp: SongInput,
                        force_from: str | None) -> None:
        """Refuse a rebuild that will rerun acquire against an upload we no longer have."""
@@ -235,13 +294,20 @@ def create_app(cfg: Config, runner=run_pipeline) -> FastAPI:
                      lyrics: str = Form(...), youtube_url: str | None = Form(None),
                      genre: str | None = Form(None), year: int | None = Form(None),
                      language: str | None = Form(None), bpm: float | None = Form(None),
+                      force_from: str | None = Form(None),
                      video: UploadFile | None = File(None),
                      cover: UploadFile | None = File(None)):
        parent = _finished_job(job_id)
+        force_from = force_from or None
+        if force_from is not None and force_from not in STAGE_ORDER:
+            raise HTTPException(400, f"unknown stage {force_from!r}")
        # a cleared field arrives as "" from some clients; "" and None must not
        # look like different inputs to plan()
        youtube_url, genre, language = (youtube_url or None, genre or None,
                                        language or None)
+        # a browser submits a textarea with CRLF; without this every rebuild
+        # from the page looks like a lyrics edit and redoes the alignment
+        lyrics = lyrics.replace("\r\n", "\n")
        has_video = bool(video and video.filename)
        if youtube_url and has_video:
            raise HTTPException(400, "provide exactly one of youtube_url / video")
@@ -259,8 +325,9 @@ def create_app(cfg: Config, runner=run_pipeline) -> FastAPI:
                        youtube_url=youtube_url, video_path=video_path,
                        genre=genre, year=year, language=language,
                        cover_path=cover_path, bpm_override=bpm)
-        _require_source(parent.inp, inp, rebuild_policy.plan(parent.inp, inp))
-        return _enqueue(new_id, inp, rebuilt_from=job_id)
+        _require_source(parent.inp, inp, rebuild_policy.earliest(
+            rebuild_policy.plan(parent.inp, inp), force_from))
+        return _enqueue(new_id, inp, rebuilt_from=job_id, force_stage=force_from)

    @app.get("/api/stages")
    def list_stages():
@@ -303,6 +370,87 @@ def create_app(cfg: Config, runner=run_pipeline) -> FastAPI:
                "video_filename": inp.video_path.name if inp.video_path else None,
                "cover_filename": inp.cover_path.name if inp.cover_path else None}

+    @app.delete("/api/jobs/{job_id}", status_code=204)
+    def delete_job(job_id: str):
+        with lock:
+            job = jobs.get(job_id)
+            if job is None:
+                raise HTTPException(404)
+            if job.status in ("queued", "running"):
+                raise HTTPException(409, "job is still in progress")
+            slug = song_slug(job.inp)
+            if any(other.status in ("queued", "running") and song_slug(other.inp) == slug
+                   for other_id, other in jobs.items() if other_id != job_id):
+                raise HTTPException(409, "another job for this song is in progress")
+            jobs.pop(job_id, None)
+            if job_id in order:
+                order.remove(job_id)
+        shutil.rmtree(cfg.work_dir / slug, ignore_errors=True)
+        shutil.rmtree(cfg.output_dir / slug, ignore_errors=True)
+
+    def _resolve_media(job: Job, header: str) -> Path | None:
+        """The file a song folder's #MP3/#VIDEO/#COVER header points at."""
+        slug = song_slug(job.inp)
+        out = (cfg.output_dir / slug).resolve()
+        txt = out / f"{slug}.txt"
+        if not txt.exists():
+            return None
+        name = get_header(parse_ultrastar(txt.read_text(encoding="utf-8")), header)
+        if not name:
+            return None
+        # a header names a file in the song folder and nothing else
+        path = (out / name).resolve()
+        if path.parent != out or not path.is_file():
+            return None
+        return path
+
+    @app.get("/api/jobs/{job_id}/media/{kind}")
+    def get_media(job_id: str, kind: str):
+        header = MEDIA_HEADERS.get(kind)
+        if header is None:
+            raise HTTPException(404)
+        job = _finished_job(job_id)
+        if job.status != "done":
+            raise HTTPException(404)
+        path = _resolve_media(job, header)
+        if path is None:
+            raise HTTPException(404)
+        return FileResponse(path)
+
+    @app.get("/api/jobs/{job_id}/review")
+    def get_review(job_id: str):
+        job = _finished_job(job_id)
+        if job.status != "done":
+            raise HTTPException(404)
+        work = cfg.work_dir / song_slug(job.inp)
+        notes_file = work / "notes_sec.json"
+        if not notes_file.exists():
+            raise HTTPException(404)
+        notes_data = json.loads(notes_file.read_text())
+        meta_file = work / "align_meta.json"
+        meta = json.loads(meta_file.read_text()) if meta_file.exists() else {}
+        dropped = set(meta.get("dropped_lines", []))
+        missing_spans = [{"start": s, "end": e}
+                         for s, e in meta.get("missing_spans", [])]
+        align_file = work / "alignment.json"
+        align_rows = json.loads(align_file.read_text()) if align_file.exists() else []
+        words_by_line: dict[int, list] = {}
+        for r in align_rows:
+            words_by_line.setdefault(r["line_index"], []).append(
+                {"text": r["word"], "start": r["start"], "end": r["end"]})
+        ref = normalize_lyrics(job.inp.lyrics_text)
+        by_line: dict[int, list] = {}
+        for n in notes_data["notes"]:
+            by_line.setdefault(n["line_index"], []).append({
+                "start": n["start"], "end": n["start"] + n["duration"],
+                "pitch": n["pitch"], "low_confidence": n["low_confidence"]})
+        lines = [{"index": i, "text": " ".join(words), "dropped": i in dropped,
+                  "notes": by_line.get(i, []), "words": words_by_line.get(i, [])}
+                 for i, words in enumerate(ref.lines)]
+        media = {kind: _resolve_media(job, header) is not None
+                 for kind, header in MEDIA_HEADERS.items()}
+        return {"lines": lines, "missing_spans": missing_spans, "media": media}
+
    return app


@@ -0,0 +1,56 @@
+from dataclasses import dataclass
+from pathlib import Path
+
+import httpx
+
+from .config import Config
+from .languages import LANGUAGES
+
+
+@dataclass
+class TranscriptSegment:
+    start: float
+    end: float
+    text: str
+
+
+def transcribe(cfg: Config, audio_path: Path, language: str | None = None,
+               client: httpx.Client | None = None) -> list[TranscriptSegment]:
+    headers = {}
+    if cfg.whisper_api_key:
+        headers["Authorization"] = f"Bearer {cfg.whisper_api_key}"
+    iso_language = None
+    if language:
+        entry = LANGUAGES.get(language.lower())
+        iso_language = entry.iso if entry else language
+
+    own = client is None
+    client = client or httpx.Client(timeout=600)
+    try:
+        with audio_path.open("rb") as f:
+            if cfg.whisper_api == "asr-webservice":
+                # onerahmet/openai-whisper-asr-webservice: fixed /asr endpoint,
+                # file field is "audio_file", and language/output are query
+                # params rather than form fields - there is no "model" param,
+                # the model is fixed by how that server was started.
+                params = {"output": "json"}
+                if iso_language:
+                    params["language"] = iso_language
+                resp = client.post(
+                    f"{cfg.whisper_url.rstrip('/')}/asr",
+                    headers=headers, params=params,
+                    files={"audio_file": (audio_path.name, f, "audio/wav")})
+            else:
+                data = {"model": cfg.whisper_model, "response_format": "verbose_json"}
+                if iso_language:
+                    data["language"] = iso_language
+                resp = client.post(
+                    f"{cfg.whisper_url.rstrip('/')}/audio/transcriptions",
+                    headers=headers, data=data,
+                    files={"file": (audio_path.name, f, "audio/wav")})
+        resp.raise_for_status()
+        return [TranscriptSegment(s["start"], s["end"], s["text"].strip())
+                for s in resp.json().get("segments", [])]
+    finally:
+        if own:
+            client.close()
@@ -0,0 +1,185 @@
+"""The one place songtrace touches yt-dlp.
+
+YouTube breaks old yt-dlp releases faster than songtrace is released, so
+besides the copy installed with songtrace (the bundled one), newer stable
+releases are fetched from PyPI into a cache dir at runtime and tried first,
+with the bundled copy as the fallback. yt-dlp runs in a subprocess and is
+driven through its command line, its most stable interface, so a cached
+release never mixes with anything songtrace has imported.
+"""
+import importlib.metadata
+import os
+import re
+import shutil
+import subprocess
+import sys
+import tempfile
+import threading
+import time
+from dataclasses import dataclass
+from pathlib import Path
+
+import httpx
+
+from .config import Config
+from .pipeline import PipelineError
+
+# The extras are what make YouTube work; see pyproject.toml.
+REQUIREMENT = "yt-dlp[default,curl-cffi,deno]"
+CHECK_INTERVAL = 24 * 3600
+_lock = threading.Lock()
+
+
+class YtDlpError(PipelineError):
+    pass
+
+
+def _key(version: str) -> tuple[int, ...]:
+    return tuple(int(p) for p in version.split("."))
+
+
+def cache_dir() -> Path:
+    base = os.environ.get("XDG_CACHE_HOME") or Path.home() / ".cache"
+    # Installs hold compiled wheels, so they are only good for one Python.
+    return Path(base) / "songtrace" / f"yt-dlp-py{sys.version_info[0]}{sys.version_info[1]}"
+
+
+@dataclass(frozen=True)
+class YtDlp:
+    version: str
+    path: Path | None = None  # None for the bundled copy
+
+    def _env(self) -> dict[str, str]:
+        env = dict(os.environ)
+        if self.path:
+            env["PYTHONPATH"] = os.pathsep.join(
+                filter(None, [str(self.path), env.get("PYTHONPATH")]))
+            env["PATH"] = os.pathsep.join(
+                [str(self.path / "bin"), env.get("PATH", "")])
+        return env
+
+    def run(self, args: list[str], log) -> None:
+        """Runs yt-dlp with `args`, passing each line it prints to `log`."""
+        proc = subprocess.Popen(
+            [sys.executable, "-P", "-m", "yt_dlp", "--ignore-config", *args],
+            stdout=subprocess.PIPE, stderr=subprocess.STDOUT, text=True,
+            errors="replace", env=self._env())
+        last = error = ""
+        for line in proc.stdout:
+            last = line.rstrip()
+            log(last)
+            if last.startswith("ERROR:"):
+                error = last
+        if proc.wait() == 0:
+            return
+        if proc.returncode == 2:
+            raise YtDlpError(f"yt-dlp {self.version} rejected songtrace's "
+                             f"options, so it may be incompatible: {last}")
+        raise YtDlpError(error or f"yt-dlp exited with status {proc.returncode}")
+
+
+def bundled() -> YtDlp:
+    return YtDlp(importlib.metadata.version("yt-dlp"))
+
+
+def _latest_release() -> str:
+    r = httpx.get("https://pypi.org/pypi/yt-dlp/json", timeout=15,
+                  follow_redirects=True)
+    r.raise_for_status()
+    return r.json()["info"]["version"]
+
+
+def _pip_install(target: Path, version: str) -> None:
+    proc = subprocess.run(
+        [sys.executable, "-m", "pip", "install", "--quiet", "--no-input",
+         "--disable-pip-version-check", "--target", str(target),
+         f"{REQUIREMENT}=={version}"],
+        capture_output=True, text=True)
+    if proc.returncode:
+        raise YtDlpError(f"pip could not install yt-dlp {version}: "
+                         f"{proc.stderr.strip()[-500:]}")
+
+
+def _install(version: str, log) -> YtDlp:
+    """Installs `version` into the cache unless it is already there.
+
+    The install is checked to run before it is moved into place, so a
+    directory in the cache is always a usable yt-dlp."""
+    dest = cache_dir() / version
+    if dest.is_dir():
+        return YtDlp(version, dest)
+    dest.parent.mkdir(parents=True, exist_ok=True)
+    log(f"installing yt-dlp {version}")
+    tmp = Path(tempfile.mkdtemp(dir=dest.parent, prefix=".tmp-"))
+    try:
+        _pip_install(tmp, version)
+        proc = subprocess.run(
+            [sys.executable, "-P", "-m", "yt_dlp", "--version"],
+            capture_output=True, text=True, env=YtDlp(version, tmp)._env())
+        got = proc.stdout.strip()
+        if proc.returncode or _key(got) != _key(version):
+            raise YtDlpError(f"yt-dlp {version} failed its check after install: "
+                             f"{(got or proc.stderr.strip())[-500:]}")
+        try:
+            tmp.rename(dest)
+        except OSError:
+            if not dest.is_dir():
+                raise
+    finally:
+        shutil.rmtree(tmp, ignore_errors=True)
+    return YtDlp(version, dest)
+
+
+def _cached() -> list[YtDlp]:
+    root = cache_dir()
+    if not root.is_dir():
+        return []
+    found = [YtDlp(p.name, p) for p in root.iterdir()
+             if p.is_dir() and re.fullmatch(r"\d+(\.\d+)*", p.name)]
+    return sorted(found, key=lambda y: _key(y.version))
+
+
+def _newest(base: YtDlp, refresh: bool, log) -> YtDlp | None:
+    """The newest cached release newer than `base`, checking PyPI for a newer
+    one first if that has not been done for a day or `refresh` is set."""
+    marker = cache_dir() / ".checked"
+    if (refresh or not marker.exists()
+            or time.time() - marker.stat().st_mtime > CHECK_INTERVAL):
+        # Recorded even on failure, so a broken release is not reinstalled
+        # for every job; a failed download refreshes regardless.
+        marker.parent.mkdir(parents=True, exist_ok=True)
+        marker.touch()
+        try:
+            latest = _latest_release()
+            if _key(latest) > _key(base.version):
+                _install(latest, log)
+        except Exception as e:
+            log(f"could not update yt-dlp: {e}")
+    newer = [y for y in _cached() if _key(y.version) > _key(base.version)]
+    for old in newer[:-1]:
+        shutil.rmtree(old.path, ignore_errors=True)
+    return newer[-1] if newer else None
+
+
+def candidates(cfg: Config, log, refresh: bool = False) -> list[YtDlp]:
+    """The yt-dlp copies to try, in order: the newest release (or the one
+    `cfg.ytdlp_version` pins), then the bundled copy."""
+    base = bundled()
+    want = cfg.ytdlp_version
+    if want == "bundled":
+        return [base]
+    with _lock:
+        if want == "auto":
+            first = _newest(base, refresh, log)
+        elif not re.fullmatch(r"\d+(\.\d+)*", want):
+            raise YtDlpError(f"ytdlp_version must be \"auto\", \"bundled\" or a "
+                             f"release such as \"2026.08.19\", not {want!r}")
+        elif _key(want) == _key(base.version):
+            first = None
+        else:
+            try:
+                first = _install(want, log)
+            except Exception as e:
+                log(f"could not install the pinned yt-dlp {want}: {e}")
+                first = None
+    return [first, base] if first else [base]
@@ -0,0 +1,36 @@
+#TITLE:Placeholder Chart
+#ARTIST:Test Artist
+#LANGUAGE:English
+#YEAR:2000
+#CREATOR:songtrace tests
+#GENRE:Test
+#COVER:Test Artist - Placeholder Chart [CO].jpg
+#MP3:Test Artist - Placeholder Chart.mp3
+#VIDEO:Test Artist - Placeholder Chart.mp4
+#BPM:217,98
+#GAP:4880
+#MEDLEYSTARTBEAT:113
+#MEDLEYENDBEAT:240
+: 0 3 16 ah
+: 5 3 20  ah
+: 11 3 23  la
+: 17 3 16  la
+: 22 28 15  ah
+: 51 2 16  doo
+: 54 31 16 ~
+- 95  
+F 113 3 18 la
+F 117 3 21  la
+F 121 3 21  la
+* 125 3 21  doo
+* 129 12 19 ~
+- 150
+: 155 4 12 Na
+R 160 4 0  na
+G 165 4 0  na
+: 170 6 14  la
+- 190
+: 194 3 16 La
+: 198 3 18  la
+* 202 10 20  ah
+E
@@ -0,0 +1,21 @@
+#TITLE:Café Placeholder
+#ARTIST:Tést Artist
+#LANGUAGE:English
+#GENRE:Test
+#YEAR:2000
+#MP3:Tést Artist - Café Placeholder.mp3
+#CREATOR:songtrace tests
+#BPM:256
+#GAP:49673
+: 0 4 2 Doo
+: 6 4 4  doo
+: 12 8 5  laé
+- 28 30
+: 32 4 0 Na
+: 38 4 -2  na
+* 44 12 -9  ah
+- 64
+F 70 4 3 La
+: 76 6 -5  la
+: 84 10 -9 ~
+E
@@ -1,8 +1,8 @@
 import json
 from pathlib import Path
-from karaoke_maker.config import Config
-from karaoke_maker.pipeline import Ctx, SongInput
-from karaoke_maker.stages import acquire
+from songtrace.config import Config
+from songtrace.pipeline import Ctx, SongInput
+from songtrace.stages import acquire

 def make_ctx(tmp_path, **kw):
    inp = SongInput(title="T", artist="A", lyrics_text="x", **kw)
@@ -24,7 +24,7 @@ def test_local_video(tmp_path, monkeypatch):
    assert any("audio.wav" in str(a[-1]) for a in calls)

 def test_youtube(tmp_path, monkeypatch):
-    def fake_dl(url, dest):
+    def fake_dl(url, dest, cfg, log):
        video = dest / "source_video.webm"
        video.write_bytes(b"v")
        thumb = dest / "source_video.jpg"
@@ -46,7 +46,7 @@ def test_outputs_are_meta_and_audio(tmp_path):

 def test_requires_exactly_one_source(tmp_path):
    import pytest
-    from karaoke_maker.pipeline import PipelineError
+    from songtrace.pipeline import PipelineError
    ctx = make_ctx(tmp_path)
    with pytest.raises(PipelineError):
        acquire.acquire_stage().run(ctx)
@@ -73,7 +73,7 @@ def test_classify_no_thumbnail(tmp_path):

 def test_classify_no_video_raises(tmp_path):
    import pytest
-    from karaoke_maker.pipeline import PipelineError
+    from songtrace.pipeline import PipelineError
    (tmp_path / "source_video.jpg").write_bytes(b"t")
    with pytest.raises(PipelineError):
        acquire._classify_downloads(tmp_path)
@@ -83,3 +83,120 @@ def test_classify_picks_largest_video_when_multiple(tmp_path):
    (tmp_path / "source_video.mp4").write_bytes(b"much larger contents here")
    video, _ = acquire._classify_downloads(tmp_path)
    assert video.name == "source_video.mp4"
+
+def test_classify_ignores_partials_left_by_a_failed_attempt(tmp_path):
+    (tmp_path / "source_video.webm").write_bytes(b"v")
+    (tmp_path / "source_video.f137.mp4.part").write_bytes(b"a much bigger partial")
+    (tmp_path / "source_video.f251.webm.ytdl").write_bytes(b"resume state")
+    video, _ = acquire._classify_downloads(tmp_path)
+    assert video.name == "source_video.webm"
+
+def test_ydl_args_default_client_adds_no_player_client(tmp_path):
+    cfg = Config(whisper_url="http://x/v1")
+    args = acquire._ydl_args(cfg, tmp_path, "default", "https://yt/x")
+    assert "--extractor-args" not in args
+    assert "--retries" in args and "--fragment-retries" in args
+    assert args[-2:] == ["--", "https://yt/x"]
+
+def test_ydl_args_named_client_pot_provider_and_cookies(tmp_path):
+    cfg = Config(whisper_url="http://x/v1",
+                 ytdlp_pot_provider_url="http://pot:4416",
+                 ytdlp_cookies=tmp_path / "cookies.txt",
+                 ytdlp_impersonate="chrome")
+    args = acquire._ydl_args(cfg, tmp_path, "tv", "https://yt/x")
+    joined = " ".join(args)
+    assert "--extractor-args youtube:player_client=tv" in joined
+    assert ("--extractor-args youtubepot-bgutilhttp:base_url=http://pot:4416"
+            in joined)
+    assert f"--cookies {tmp_path / 'cookies.txt'}" in joined
+    assert "--impersonate chrome" in joined
+
+def test_ydl_args_parse_with_real_yt_dlp(tmp_path):
+    """Every option songtrace passes must exist in the bundled yt-dlp."""
+    import pytest
+    pytest.importorskip("yt_dlp")
+    from yt_dlp import parse_options
+    cfg = Config(whisper_url="http://x/v1", ytdlp_pot_provider_url="http://p:1",
+                 ytdlp_cookies=tmp_path / "c.txt")
+    parse_options(acquire._ydl_args(cfg, tmp_path, "tv", "https://yt/x"))
+
+class FakeYtDlp:
+    def __init__(self, version, behaviour, seen):
+        self.version, self._behaviour, self._seen = version, behaviour, seen
+    def run(self, args, log):
+        client = next((a for a in args if a.startswith("youtube:")), None)
+        self._seen.append((self.version, client))
+        self._behaviour(self.version, client)
+
+def _fake_candidates(monkeypatch, behaviour, versions=("2026.09.01", "2026.08.19"),
+                     refreshed=None):
+    """Stub ytdlp.candidates; `refreshed` is what a refresh turns up."""
+    seen = []
+    def candidates(cfg, log, refresh=False):
+        vs = refreshed if refresh and refreshed else versions
+        return [FakeYtDlp(v, behaviour, seen) for v in vs]
+    monkeypatch.setattr(acquire.ytdlp, "candidates", candidates)
+    return seen
+
+def test_download_falls_through_to_the_next_client(tmp_path, monkeypatch):
+    from songtrace.ytdlp import YtDlpError
+    def behaviour(version, client):
+        if client is None:
+            raise YtDlpError("ERROR: HTTP Error 403: Forbidden")
+        (tmp_path / "source_video.webm").write_bytes(b"v")
+    seen = _fake_candidates(monkeypatch, behaviour)
+    cfg = Config(whisper_url="http://x/v1", ytdlp_clients=["default", "tv"])
+    logged = []
+    video, thumb = acquire.download_youtube("https://yt/x", tmp_path, cfg,
+                                            logged.append)
+    assert video.name == "source_video.webm"
+    assert thumb is None
+    assert seen == [("2026.09.01", None),
+                    ("2026.09.01", "youtube:player_client=tv")]
+    assert any("403" in m for m in logged)
+
+def test_download_falls_back_to_the_bundled_copy(tmp_path, monkeypatch):
+    from songtrace.ytdlp import YtDlpError
+    def behaviour(version, client):
+        if version == "2026.09.01":
+            raise YtDlpError("rejected songtrace's options")
+        (tmp_path / "source_video.webm").write_bytes(b"v")
+    seen = _fake_candidates(monkeypatch, behaviour)
+    cfg = Config(whisper_url="http://x/v1", ytdlp_clients=["default"])
+    acquire.download_youtube("https://yt/x", tmp_path, cfg, lambda m: None)
+    assert [v for v, _ in seen] == ["2026.09.01", "2026.08.19"]
+
+def test_download_retries_with_a_release_found_after_failing(tmp_path, monkeypatch):
+    from songtrace.ytdlp import YtDlpError
+    def behaviour(version, client):
+        if version != "2026.10.01":
+            raise YtDlpError("ERROR: HTTP Error 403: Forbidden")
+        (tmp_path / "source_video.webm").write_bytes(b"v")
+    seen = _fake_candidates(monkeypatch, behaviour,
+                            refreshed=("2026.10.01", "2026.08.19"))
+    cfg = Config(whisper_url="http://x/v1", ytdlp_clients=["default"])
+    acquire.download_youtube("https://yt/x", tmp_path, cfg, lambda m: None)
+    assert [v for v, _ in seen] == ["2026.09.01", "2026.08.19", "2026.10.01"]
+
+def test_download_reports_versions_and_clients_when_all_fail(tmp_path, monkeypatch):
+    import pytest
+    from songtrace.pipeline import PipelineError
+    from songtrace.ytdlp import YtDlpError
+    def behaviour(version, client):
+        raise YtDlpError("ERROR: HTTP Error 403: Forbidden")
+    seen = _fake_candidates(monkeypatch, behaviour)
+    cfg = Config(whisper_url="http://x/v1", ytdlp_clients=["default", "tv"])
+    with pytest.raises(PipelineError) as e:
+        acquire.download_youtube("https://yt/x", tmp_path, cfg, lambda m: None)
+    assert "2026.09.01, 2026.08.19" in str(e.value)
+    assert "default, tv" in str(e.value)
+    assert "403" in str(e.value)
+    assert len(seen) == 4
+
+def test_download_with_no_configured_clients_still_tries_default(tmp_path, monkeypatch):
+    seen = _fake_candidates(
+        monkeypatch,
+        lambda v, c: (tmp_path / "source_video.mp4").write_bytes(b"v"))
+    cfg = Config(whisper_url="http://x/v1", ytdlp_clients=[])
+    acquire.download_youtube("https://yt/x", tmp_path, cfg, lambda m: None)
+    assert seen == [("2026.09.01", None)]
@@ -2,10 +2,10 @@ import json

 import pytest

-from karaoke_maker.config import Config
-from karaoke_maker.lyrics import normalize_lyrics
-from karaoke_maker.pipeline import Ctx, PipelineError, SongInput
-from karaoke_maker.stages import align
+from songtrace.config import Config
+from songtrace.lyrics import normalize_lyrics
+from songtrace.pipeline import Ctx, PipelineError, SongInput
+from songtrace.stages import align

 LYRICS = "it's one two three\nfour five six seven eight"

@@ -137,3 +137,26 @@ def test_words_split_across_sections(tmp_path, monkeypatch):
    check_invariants(rows, LYRICS)
    assert rows[3]["end"] <= 5.0
    assert rows[4]["start"] >= 29.0
+
+
+def test_align_drops_lines_never_sung(tmp_path, monkeypatch):
+    lyrics = "it's one two three\nfour five six seven eight\nnine ten eleven twelve"
+    transcript = [{"start": 1.0, "end": 4.0,
+                   "text": "it's one two three four five six seven eight"}]
+    rows = run_align(tmp_path, monkeypatch, lyrics, transcript)
+    assert {r["line_index"] for r in rows} == {0, 1}
+    meta = json.loads((tmp_path / "w" / "align_meta.json").read_text())
+    assert meta["dropped_lines"] == [2]
+    assert meta["missing_spans"] == []
+
+
+def test_align_flags_missing_span_for_unmatched_singing(tmp_path, monkeypatch):
+    lyrics = "it's one two three\nfour five six seven eight"
+    transcript = [
+        {"start": 1.0, "end": 4.0, "text": "it's one two three four five six seven eight"},
+        {"start": 20.0, "end": 23.0, "text": "completely unrelated ad lib bridge"},
+    ]
+    run_align(tmp_path, monkeypatch, lyrics, transcript)
+    meta = json.loads((tmp_path / "w" / "align_meta.json").read_text())
+    assert meta["dropped_lines"] == []
+    assert meta["missing_spans"] == [[19.0, 24.0]]
@@ -0,0 +1,26 @@
+import numpy as np
+import soundfile as sf
+
+from songtrace.audio_io import load_wav
+
+
+def test_load_wav_mono(tmp_path):
+    path = tmp_path / "mono.wav"
+    data = np.linspace(-0.5, 0.5, 1600, dtype=np.float32)
+    sf.write(path, data, 16000)
+    waveform, sr = load_wav(path)
+    assert sr == 16000
+    assert waveform.shape == (1, 1600)
+    assert np.allclose(waveform[0].numpy(), data, atol=1e-4)
+
+
+def test_load_wav_stereo(tmp_path):
+    path = tmp_path / "stereo.wav"
+    left = np.linspace(-0.5, 0.5, 800, dtype=np.float32)
+    right = np.linspace(0.5, -0.5, 800, dtype=np.float32)
+    sf.write(path, np.stack([left, right], axis=1), 16000)
+    waveform, sr = load_wav(path)
+    assert sr == 16000
+    assert waveform.shape == (2, 800)
+    assert np.allclose(waveform[0].numpy(), left, atol=1e-4)
+    assert np.allclose(waveform[1].numpy(), right, atol=1e-4)
@@ -1,5 +1,5 @@
-from karaoke_maker.cli import main
+from songtrace.cli import main

 def test_version(capsys):
    assert main(["--version"]) == 0
-    assert "karaoke-maker" in capsys.readouterr().out
+    assert "songtrace" in capsys.readouterr().out
@@ -2,9 +2,9 @@ import os

 import pytest

-import karaoke_maker.cli as cli
-from karaoke_maker.config import Config
-from karaoke_maker.pipeline import PipelineError, PipelineResult
+import songtrace.cli as cli
+from songtrace.config import Config
+from songtrace.pipeline import PipelineError, PipelineResult


 def setup_cfg(monkeypatch, tmp_path):
@@ -1,6 +1,6 @@
 import pytest
 from pathlib import Path
-from karaoke_maker.config import Config, load_config
+from songtrace.config import Config, load_config


 def test_load_minimal(tmp_path):
@@ -24,7 +24,7 @@ def test_paths_and_overrides(tmp_path):
 def test_env_var(tmp_path, monkeypatch):
    cfg_file = tmp_path / "c.toml"
    cfg_file.write_text('whisper_url = "http://env/v1"\n')
-    monkeypatch.setenv("KARAOKE_MAKER_CONFIG", str(cfg_file))
+    monkeypatch.setenv("SONGTRACE_CONFIG", str(cfg_file))
    assert load_config().whisper_url == "http://env/v1"


@@ -4,27 +4,27 @@ import shutil
 import pytest

 requires_full = pytest.mark.skipif(
-    not os.environ.get("KARAOKE_E2E"),
-    reason="set KARAOKE_E2E=1 (needs GPU, ffmpeg, whisper server, network)")
+    not os.environ.get("SONGTRACE_E2E"),
+    reason="set SONGTRACE_E2E=1 (needs GPU, ffmpeg, whisper server, network)")

@requires_full
 def test_full_pipeline(tmp_path):
-    from karaoke_maker.config import load_config
-    from karaoke_maker.pipeline import SongInput, run_pipeline
+    from songtrace.config import load_config
+    from songtrace.pipeline import SongInput, run_pipeline
    assert shutil.which("ffmpeg"), "ffmpeg required"
    cfg = load_config()
    cfg.output_dir = tmp_path / "out"
    cfg.work_dir = tmp_path / "work"
    inp = SongInput(
-        title=os.environ["KARAOKE_E2E_TITLE"],
-        artist=os.environ["KARAOKE_E2E_ARTIST"],
-        lyrics_text=open(os.environ["KARAOKE_E2E_LYRICS"]).read(),
-        youtube_url=os.environ.get("KARAOKE_E2E_URL"),
+        title=os.environ["SONGTRACE_E2E_TITLE"],
+        artist=os.environ["SONGTRACE_E2E_ARTIST"],
+        lyrics_text=open(os.environ["SONGTRACE_E2E_LYRICS"]).read(),
+        youtube_url=os.environ.get("SONGTRACE_E2E_URL"),
    )
    result = run_pipeline(inp, cfg)
    txts = list(result.output_dir.glob("*.txt"))
    assert len(txts) == 1
-    from karaoke_maker.ultrastar import get_header, parse_ultrastar
+    from songtrace.ultrastar import get_header, parse_ultrastar
    song = parse_ultrastar(txts[0].read_text(encoding="utf-8"))
    assert get_header(song, "BPM") and get_header(song, "GAP")
    assert len(song.body) > 20
@@ -2,11 +2,11 @@ import json

 import numpy as np
 import pytest
-from karaoke_maker.config import Config
-from karaoke_maker.grid import scale_bpm, to_ultrastar
-from karaoke_maker.pipeline import Ctx, PipelineError, SongInput
-from karaoke_maker.stages import all_stages, grid as grid_stage_mod
-from karaoke_maker.ultrastar import (
+from songtrace.config import Config
+from songtrace.grid import scale_bpm, to_ultrastar
+from songtrace.pipeline import Ctx, PipelineError, SongInput
+from songtrace.stages import all_stages, grid as grid_stage_mod
+from songtrace.ultrastar import (
    LineBreak, Note, get_header, parse_ultrastar,
 )

@@ -1,4 +1,4 @@
-from karaoke_maker.lyrics import normalize_lyrics, alignment_token
+from songtrace.lyrics import normalize_lyrics, alignment_token

 RAW = """[Verse 1]
 It's one two three
@@ -0,0 +1,129 @@
+import sys
+import types
+
+import pytest
+
+from songtrace import models
+from songtrace.stages import align, pitch
+
+
+@pytest.fixture
+def isolated_callbacks(monkeypatch):
+    """release() with only the callbacks a test registers."""
+    monkeypatch.setattr(models, "_callbacks", [])
+    return models._callbacks
+
+
+def fake_torch(available=True, initialised=True, emptied=None, reserved=None):
+    """A torch whose allocator reports `reserved` bytes, then 0 once emptied."""
+    state = {"reserved": 4 * 2 ** 20 if reserved is None else reserved}
+
+    def empty_cache():
+        emptied.append(True)
+        state["reserved"] = 0
+
+    mod = types.ModuleType("torch")
+    mod.cuda = types.SimpleNamespace(
+        is_available=lambda: available,
+        is_initialized=lambda: initialised,
+        memory_reserved=lambda: state["reserved"],
+        empty_cache=empty_cache)
+    return mod
+
+
+def test_release_without_torch_or_callbacks_does_nothing(monkeypatch,
+                                                         isolated_callbacks):
+    monkeypatch.delitem(sys.modules, "torch", raising=False)
+    assert models.release() == 0
+
+
+def test_release_runs_every_registered_callback(isolated_callbacks):
+    called = []
+    models.on_release(lambda: called.append("a"))
+    models.on_release(lambda: called.append("b"))
+    models.release()
+    assert called == ["a", "b"]
+
+
+def test_on_release_returns_the_function_so_it_works_as_a_decorator():
+    def drop():
+        pass
+    assert models.on_release(drop) is drop
+    models._callbacks.remove(drop)
+
+
+def test_release_empties_the_allocator_when_cuda_is_in_use(monkeypatch,
+                                                           isolated_callbacks):
+    emptied = []
+    monkeypatch.setitem(sys.modules, "torch", fake_torch(emptied=emptied))
+    assert models.release() == 4 * 2 ** 20
+    assert emptied == [True]
+
+
+def test_release_leaves_the_allocator_alone_without_a_gpu(monkeypatch,
+                                                          isolated_callbacks):
+    emptied = []
+    monkeypatch.setitem(sys.modules, "torch",
+                        fake_torch(available=False, emptied=emptied))
+    assert models.release() == 0
+    assert emptied == []
+
+
+def test_release_does_not_initialise_cuda_just_to_clean_up(monkeypatch,
+                                                           isolated_callbacks):
+    """A run that never touched the GPU must not create a context on the way out."""
+    emptied = []
+    monkeypatch.setitem(sys.modules, "torch",
+                        fake_torch(initialised=False, emptied=emptied))
+    assert models.release() == 0
+    assert emptied == []
+
+
+def test_align_drops_its_cached_model():
+    align._model_cache["cuda"] = object()
+    align.unload_model()
+    assert align._model_cache == {}
+
+
+def test_align_registers_its_unload():
+    assert align.unload_model in models._callbacks
+
+
+def test_pitch_drops_the_model_torchcrepe_cached(monkeypatch):
+    torchcrepe = types.ModuleType("torchcrepe")
+
+    def infer():
+        pass
+
+    infer.model = object()
+    infer.capacity = "full"
+    torchcrepe.infer = infer
+    monkeypatch.setitem(sys.modules, "torchcrepe", torchcrepe)
+    pitch.unload_model()
+    assert not hasattr(infer, "model")
+    assert not hasattr(infer, "capacity")
+
+
+def test_pitch_unload_is_fine_when_torchcrepe_was_never_imported(monkeypatch):
+    monkeypatch.delitem(sys.modules, "torchcrepe", raising=False)
+    pitch.unload_model()
+
+
+def test_pitch_unload_is_fine_when_no_model_was_cached(monkeypatch):
+    torchcrepe = types.ModuleType("torchcrepe")
+    torchcrepe.infer = lambda: None
+    monkeypatch.setitem(sys.modules, "torchcrepe", torchcrepe)
+    pitch.unload_model()
+
+
+def test_pitch_registers_its_unload():
+    assert pitch.unload_model in models._callbacks
+
+
+def test_release_reports_nothing_freed_when_the_allocator_was_already_empty(
+        monkeypatch, isolated_callbacks):
+    emptied = []
+    monkeypatch.setitem(sys.modules, "torch",
+                        fake_torch(emptied=emptied, reserved=0))
+    assert models.release() == 0
+    assert emptied == [True]
@@ -2,13 +2,13 @@ import json

 import numpy as np
 import pytest
-from karaoke_maker.config import Config
-from karaoke_maker.notes import (
+from songtrace.config import Config
+from songtrace.notes import (
    MIN_NOTE_SEC, TimedNote, build_notes, hz_to_ultrastar, segment_span,
    syllabify,
 )
-from karaoke_maker.pipeline import Ctx, SongInput
-from karaoke_maker.stages import all_stages, notes as notes_stage_mod
+from songtrace.pipeline import Ctx, SongInput
+from songtrace.stages import all_stages, notes as notes_stage_mod


 def flat_pitch(duration=10.0, hz=261.63):   # C4
@@ -78,7 +78,7 @@ def test_syllabify():
    assert syllabify("reminds", "English") == ["re", "minds"]
    assert syllabify("a", "English") == ["a"]
    assert syllabify("xyzzy", None) == ["xyzzy"]
-    assert syllabify("She's", "English") == ["She's"]
+    assert syllabify("It's", "English") == ["It's"]


 def test_syllabify_preserves_word_characters():
@@ -96,12 +96,12 @@ def test_syllabify_language_without_pyphen_dict_is_not_mishandled_as_english():
 def test_build_notes_basic_line():
    times, f0, conf = flat_pitch()
    rows = [
-        {"word": "She's", "start": 1.0, "end": 1.5, "score": 0.9, "line_index": 0},
+        {"word": "It's", "start": 1.0, "end": 1.5, "score": 0.9, "line_index": 0},
        {"word": "got", "start": 1.6, "end": 2.0, "score": 0.9, "line_index": 0},
        {"word": "Now", "start": 5.0, "end": 5.5, "score": 0.9, "line_index": 1},
    ]
    notes, breaks, warnings = build_notes(rows, times, f0, conf, "English")
-    assert notes[0].text == "She's"          # line-initial: no leading space
+    assert notes[0].text == "It's"          # line-initial: no leading space
    assert notes[1].text == " got"           # mid-line: leading space
    assert notes[2].text == "Now"            # new line: no leading space
    assert len(breaks) == 1 and 2.0 <= breaks[0] <= 5.0
@@ -278,7 +278,7 @@ def test_notes_stage_writes_json(tmp_path):
    times, f0, conf = flat_pitch()
    np.savez(work / "pitch.npz", times=times, f0=f0, conf=conf)
    (work / "alignment.json").write_text(json.dumps([
-        {"word": "She's", "start": 1.0, "end": 1.5, "score": 0.9, "line_index": 0},
+        {"word": "It's", "start": 1.0, "end": 1.5, "score": 0.9, "line_index": 0},
        {"word": "gone", "start": 3.0, "end": 3.5, "score": 0.9, "line_index": 1},
    ]))
    ctx = Ctx(inp=SongInput(title="T", artist="A", lyrics_text="x", language="English"),
@@ -287,7 +287,7 @@ def test_notes_stage_writes_json(tmp_path):
    assert stage.outputs(ctx) == [work / "notes_sec.json"]
    stage.run(ctx)
    data = json.loads((work / "notes_sec.json").read_text())
-    assert [n["text"] for n in data["notes"]] == ["She's", "gone"]
+    assert [n["text"] for n in data["notes"]] == ["It's", "gone"]
    assert data["breaks"] == [2.25]
    assert data["warnings"] == []
    assert TimedNote(**data["notes"][0]).pitch == 0
@@ -295,3 +295,42 @@ def test_notes_stage_writes_json(tmp_path):

 def test_notes_stage_registered():
    assert "notes" in [s.name for s in all_stages()]
+
+
+def test_build_notes_keeps_matched_words_within_missing_span():
+    # A missing span covers whole Whisper segments, so it can hold matched
+    # lines alongside the unmatched singing; those lines must keep their notes.
+    times, f0, conf = flat_pitch()
+    rows = [
+        {"word": "keep", "start": 1.0, "end": 1.4, "score": 0.9, "line_index": 0},
+        {"word": "held", "start": 5.0, "end": 5.4, "score": 0.9, "line_index": 1},
+    ]
+    notes, breaks, warnings = build_notes(rows, times, f0, conf, "English",
+                                          missing_spans=[(4.5, 6.0)])
+    assert [n.text for n in notes] == ["keep", "held"]
+    assert breaks == [(1.4 + 5.0) / 2]
+    assert warnings == ["missing lyrics near 0:04–0:06"]
+
+
+def test_warnings_report_dropped_lines():
+    times, f0, conf = flat_pitch()
+    rows = [{"word": "keep", "start": 1.0, "end": 1.4, "score": 0.9, "line_index": 0}]
+    ref_lines = [["keep"], ["the", "outro"]]
+    _, _, warnings = build_notes(rows, times, f0, conf, "English",
+                                 dropped_lines=[1], ref_lines=ref_lines)
+    assert warnings == ['1 line dropped (not sung) near "the outro"']
+
+
+def test_warnings_report_missing_spans():
+    times, f0, conf = flat_pitch()
+    rows = [{"word": "keep", "start": 1.0, "end": 1.4, "score": 0.9, "line_index": 0}]
+    _, _, warnings = build_notes(rows, times, f0, conf, "English",
+                                 missing_spans=[(81.0, 89.5)])
+    assert warnings == ["missing lyrics near 1:21–1:29"]
+
+
+def test_notes_carry_line_index():
+    times, f0, conf = flat_pitch()
+    rows = [{"word": "hey", "start": 1.0, "end": 1.4, "score": 0.9, "line_index": 3}]
+    notes, _, _ = build_notes(rows, times, f0, conf, "English")
+    assert notes[0].line_index == 3
@@ -4,12 +4,12 @@ from pathlib import Path
 import pytest
 from PIL import Image

-from karaoke_maker.config import Config
-from karaoke_maker.pipeline import (
+from songtrace.config import Config
+from songtrace.pipeline import (
    STAGE_ORDER, Ctx, PipelineError, SongInput, run_pipeline, song_slug,
 )
-from karaoke_maker.stages import all_stages, package
-from karaoke_maker.ultrastar import get_header, parse_ultrastar
+from songtrace.stages import all_stages, package
+from songtrace.ultrastar import get_header, parse_ultrastar

 SONG_TXT = ("#TITLE:T\n#ARTIST:A\n#MP3:A - T.mp3\n#COVER:A - T.jpg\n"
            "#VIDEO:A - T.webm\n#BPM:200\n#GAP:0\n: 0 1 0 la\nE\n")
@@ -49,19 +49,37 @@ def make_pipeline_inputs(tmp_path, cover=None):
    return inp, cfg


-def test_package_copies_compatible_video(tmp_path, monkeypatch):
-    monkeypatch.setattr(package, "ffmpeg", lambda *a: Path(a[-1]).write_bytes(b"m"))
+def test_package_remuxes_compatible_video_without_reencoding(tmp_path, monkeypatch):
+    calls = []
+    monkeypatch.setattr(package, "ffmpeg",
+                        lambda *a: (calls.append(a), Path(a[-1]).write_bytes(b"m")))
    monkeypatch.setattr(package, "ffprobe_codec", lambda p: ("webm", "vp9"))
    ctx = make_ctx(tmp_path, cover=make_cover(tmp_path))
    package.package_stage().run(ctx)
    out = tmp_path / "out" / "A - T"
    assert (out / "A - T.txt").exists()
    assert (out / "A - T.mp3").exists()
-    assert (out / "A - T.webm").read_bytes() == b"v"
+    assert (out / "A - T.webm").exists()
+    video_call = next(c for c in calls if c[-1].endswith("A - T.webm"))
+    assert "-c" in video_call and "copy" in video_call, "the video was re-encoded"
    img = Image.open(out / "A - T.jpg")
    assert img.width == img.height


+@pytest.mark.parametrize("probe,name", [(("webm", "vp9"), "A - T.webm"),
+                                        (("avi", "mpeg4"), "A - T.mp4")])
+def test_package_strips_the_videos_audio_track(tmp_path, monkeypatch, probe, name):
+    """#MP3 is the song's audio; a video that keeps its own plays as an echo."""
+    calls = []
+    monkeypatch.setattr(package, "ffmpeg",
+                        lambda *a: (calls.append(a), Path(a[-1]).write_bytes(b"m")))
+    monkeypatch.setattr(package, "ffprobe_codec", lambda p: probe)
+    ctx = make_ctx(tmp_path, cover=make_cover(tmp_path))
+    package.package_stage().run(ctx)
+    video_call = next(c for c in calls if c[-1].endswith(name))
+    assert "-an" in video_call
+
+
 def test_package_reencodes_and_rewrites_video_header(tmp_path, monkeypatch):
    monkeypatch.setattr(package, "ffmpeg", lambda *a: Path(a[-1]).write_bytes(b"m"))
    monkeypatch.setattr(package, "ffprobe_codec", lambda p: ("avi", "mpeg4"))
@@ -1,7 +1,7 @@
 import json
 from pathlib import Path
-from karaoke_maker.config import Config
-from karaoke_maker.pipeline import SongInput, Stage, run_pipeline, song_slug
+from songtrace.config import Config
+from songtrace.pipeline import SongInput, Stage, read_warnings, run_pipeline, song_slug

 def make_cfg(tmp_path):
    return Config(whisper_url="http://x/v1", output_dir=tmp_path / "out",
@@ -46,3 +46,56 @@ def test_warnings_surfaced(tmp_path):
    st = Stage(name="n", outputs=lambda ctx: [ctx.work / "n.done"], run=run)
    result = run_pipeline(inp, cfg, log=lambda m: None, stages=[st])
    assert result.warnings == ["w1"]
+
+def test_read_warnings_is_empty_without_a_notes_file(tmp_path):
+    assert read_warnings(tmp_path) == []
+
+def test_read_warnings_reads_the_notes_file(tmp_path):
+    (tmp_path / "notes_sec.json").write_text(json.dumps({"warnings": ["w1", "w2"]}))
+    assert read_warnings(tmp_path) == ["w1", "w2"]
+
+def test_gpu_is_released_after_every_stage(tmp_path, monkeypatch):
+    """Not just at the end: a later stage may wait minutes on the Whisper server."""
+    from songtrace import models
+    cfg, inp, calls = make_cfg(tmp_path), make_inp(), []
+    monkeypatch.setattr(models, "release", lambda: calls.append("release"))
+    stages = [fake_stage("s1", calls), fake_stage("s2", calls)]
+    run_pipeline(inp, cfg, log=lambda m: None, stages=stages)
+    assert calls == ["s1", "release", "s2", "release"]
+
+def test_gpu_is_released_when_a_stage_fails(tmp_path, monkeypatch):
+    import pytest
+    from songtrace import models
+    from songtrace.pipeline import PipelineError
+    cfg, inp, calls = make_cfg(tmp_path), make_inp(), []
+    monkeypatch.setattr(models, "release", lambda: calls.append("release"))
+    def boom(ctx):
+        raise PipelineError("boom")
+    stages = [Stage(name="s1", outputs=lambda ctx: [ctx.work / "s1.done"], run=boom)]
+    with pytest.raises(PipelineError):
+        run_pipeline(inp, cfg, log=lambda m: None, stages=stages)
+    assert calls == ["release"]
+
+def test_a_cached_stage_does_not_pay_for_a_release(tmp_path, monkeypatch):
+    from songtrace import models
+    cfg, inp, calls = make_cfg(tmp_path), make_inp(), []
+    stages = [fake_stage("s1", calls)]
+    run_pipeline(inp, cfg, log=lambda m: None, stages=stages)
+    calls.clear()
+    monkeypatch.setattr(models, "release", lambda: calls.append("release"))
+    run_pipeline(inp, cfg, log=lambda m: None, stages=stages)
+    assert calls == []
+
+def test_a_release_that_freed_memory_is_reported_in_the_log(tmp_path, monkeypatch):
+    from songtrace import models
+    cfg, inp, lines = make_cfg(tmp_path), make_inp(), []
+    monkeypatch.setattr(models, "release", lambda: 512 * 2 ** 20)
+    run_pipeline(inp, cfg, log=lines.append, stages=[fake_stage("s1", [])])
+    assert "[s1] released 512 MiB of GPU memory" in lines
+
+def test_a_release_that_freed_nothing_stays_out_of_the_log(tmp_path, monkeypatch):
+    from songtrace import models
+    cfg, inp, lines = make_cfg(tmp_path), make_inp(), []
+    monkeypatch.setattr(models, "release", lambda: 0)
+    run_pipeline(inp, cfg, log=lines.append, stages=[fake_stage("s1", [])])
+    assert not any("released" in line for line in lines)
@@ -1,7 +1,7 @@
 import numpy as np
-from karaoke_maker.config import Config
-from karaoke_maker.pipeline import Ctx, SongInput
-from karaoke_maker.stages import pitch
+from songtrace.config import Config
+from songtrace.pipeline import Ctx, SongInput
+from songtrace.stages import pitch


 def test_pitch_stage_writes_npz(tmp_path, monkeypatch):
@@ -3,9 +3,9 @@ from pathlib import Path

 import pytest

-from karaoke_maker.config import Config
-from karaoke_maker.pipeline import STAGE_ORDER, PipelineError, SongInput
-from karaoke_maker.rebuild import (FIELD_STAGE, changed_fields, plan,
+from songtrace.config import Config
+from songtrace.pipeline import STAGE_ORDER, PipelineError, SongInput
+from songtrace.rebuild import (FIELD_STAGE, changed_fields, earliest, plan,
                                   relocate_work, superseded_output)


@@ -124,3 +124,14 @@ def test_superseded_output_refuses_a_path_outside_the_output_root(tmp_path):
 def test_superseded_output_refuses_the_output_root_itself(tmp_path):
    root = tmp_path / "out"
    assert superseded_output(str(root), str(root / "A - T"), root) is None
+
+
+def test_earliest_picks_the_earlier_stage():
+    assert earliest("package", "separate") == "separate"
+    assert earliest("separate", "package") == "separate"
+
+
+def test_earliest_ignores_unset_stages():
+    assert earliest(None, "grid") == "grid"
+    assert earliest("grid", None) == "grid"
+    assert earliest(None, None) is None
@@ -1,9 +1,9 @@
 import numpy as np
 import soundfile as sf
 import torch
-from karaoke_maker.config import Config
-from karaoke_maker.pipeline import Ctx, SongInput
-from karaoke_maker.stages import separate
+from songtrace.config import Config
+from songtrace.pipeline import Ctx, SongInput
+from songtrace.stages import separate


 def test_separate_writes_stems(tmp_path, monkeypatch):
@@ -1,6 +1,6 @@
-from karaoke_maker.lyrics import normalize_lyrics
-from karaoke_maker.whisper_client import TranscriptSegment
-from karaoke_maker.stages.transcribe import lyrics_similarity, vocal_sections
+from songtrace.lyrics import normalize_lyrics
+from songtrace.whisper_client import TranscriptSegment
+from songtrace.stages.transcribe import classify_lyrics_match, lyrics_similarity, vocal_sections

 def seg(s, e, t=""):
    return TranscriptSegment(start=s, end=e, text=t)
@@ -26,3 +26,41 @@ def test_similarity_low_for_wrong_lyrics():
    ref = normalize_lyrics("completely different words here entirely")
    segs = [seg(0, 2, "it's one two three four five six seven eight")]
    assert lyrics_similarity(segs, ref) < 0.4
+
+def test_classify_lyrics_match_flags_extra_reference_line():
+    ref = normalize_lyrics("it's one two three\nfour five six seven eight\nnine ten eleven")
+    segs = [seg(1, 4, "it's one two three four five six seven eight")]
+    dropped, missing = classify_lyrics_match(segs, ref)
+    assert dropped == [2]
+    assert missing == []
+
+def test_classify_lyrics_match_flags_unmatched_singing():
+    ref = normalize_lyrics("it's one two three\nfour five six seven eight")
+    segs = [seg(1, 4, "it's one two three four five six seven eight"),
+            seg(20, 23, "completely unrelated ad lib bridge")]
+    dropped, missing = classify_lyrics_match(segs, ref)
+    assert dropped == []
+    assert missing == [(19.0, 24.0)]
+
+def test_classify_lyrics_match_keeps_partially_matched_lines():
+    ref = normalize_lyrics("it's one two three\nfour five six seven eight")
+    # line 1 heard with one word garbled ("five" -> "fife"): still a replace/equal
+    # match, so the line stays matched rather than dropped
+    segs = [seg(1, 4, "it's one two three four fife six seven eight")]
+    dropped, _ = classify_lyrics_match(segs, ref)
+    assert dropped == []
+
+def test_classify_lyrics_match_scopes_missing_span_to_adlib_not_whole_section():
+    # matched line and an unmatched ad-lib close enough together that
+    # vocal_sections would merge them into one section; missing_spans must
+    # stay tight around the ad-lib so the review marks the right place.
+    ref = normalize_lyrics("it's one two three\nfour five six seven eight")
+    segs = [seg(1, 4, "it's one two three four five six seven eight"),
+            seg(6, 8, "completely unrelated ad lib")]
+    sections = vocal_sections(segs)
+    assert sections == [(0.0, 9.0)]      # confirms the two segments do merge
+    dropped, missing = classify_lyrics_match(segs, ref)
+    assert dropped == []
+    assert missing == [(5.0, 9.0)]
+    for a, b in missing:
+        assert not (a <= 1.0 < b)        # matched line's start (1.0) untouched
@@ -1,5 +1,5 @@
 from pathlib import Path
-from karaoke_maker.ultrastar import (
+from songtrace.ultrastar import (
    Note, LineBreak, parse_ultrastar, serialize_ultrastar,
    get_header, format_bpm, parse_bpm,
 )
@@ -10,8 +10,8 @@ def load(name):
    return (FIXTURES / name).read_text(encoding="utf-8", errors="replace")

 def test_parse_headers_and_notes():
-    song = parse_ultrastar(load("sanctuary.txt"))
-    assert get_header(song, "artist") == "Utada Hikaru"
+    song = parse_ultrastar(load("chart_crlf.txt"))
+    assert get_header(song, "artist") == "Test Artist"
    assert parse_bpm(get_header(song, "BPM")) == 217.98
    first = song.body[0]
    assert first == Note(kind=":", beat=0, length=3, pitch=16, text="ah")
@@ -19,16 +19,16 @@ def test_parse_headers_and_notes():
    assert any(n.kind == "F" for n in song.body if isinstance(n, Note))

 def test_note_text_keeps_leading_space():
-    song = parse_ultrastar(load("sanctuary.txt"))
+    song = parse_ultrastar(load("chart_crlf.txt"))
    notes = [n for n in song.body if isinstance(n, Note)]
    assert notes[1].text == " ah"

 def test_negative_pitch():
-    song = parse_ultrastar(load("sweet_child.txt"))
+    song = parse_ultrastar(load("chart_latin1.txt"))
    assert any(n.pitch == -9 for n in song.body if isinstance(n, Note))

 def test_round_trip():
-    for name in ("sweet_child.txt", "sanctuary.txt"):
+    for name in ("chart_latin1.txt", "chart_crlf.txt"):
        song = parse_ultrastar(load(name))
        again = parse_ultrastar(serialize_ultrastar(song))
        assert again == song
@@ -1,14 +1,16 @@
 import io
+import json
 import re
+import shutil
 import threading
 import time

 import pytest
 from fastapi.testclient import TestClient

-from karaoke_maker.config import Config
-from karaoke_maker.pipeline import STAGE_ORDER, PipelineError, PipelineResult, song_slug
-from karaoke_maker.web import STAGE_LINE, create_app
+from songtrace.config import Config
+from songtrace.pipeline import STAGE_ORDER, PipelineError, PipelineResult, song_slug
+from songtrace.web import STAGE_LINE, create_app


 def make_client(tmp_path, runner):
@@ -190,7 +192,7 @@ def test_cover_upload_kept_after_success(tmp_path):

 def test_cleanup_uploads_removes_an_inherited_upload_dir(tmp_path):
    """A rebuild's video may live under its parent's id, not its own."""
-    from karaoke_maker.web import _cleanup_uploads
+    from songtrace.web import _cleanup_uploads
    cfg = Config(whisper_url="u", work_dir=tmp_path / "work",
                 output_dir=tmp_path / "out")
    parent = cfg.work_dir / "uploads" / "parent"
@@ -201,7 +203,7 @@ def test_cleanup_uploads_removes_an_inherited_upload_dir(tmp_path):


 def test_cleanup_uploads_ignores_paths_outside_the_upload_root(tmp_path):
-    from karaoke_maker.web import _cleanup_uploads
+    from songtrace.web import _cleanup_uploads
    cfg = Config(whisper_url="u", work_dir=tmp_path / "work",
                 output_dir=tmp_path / "out")
    outside = tmp_path / "somewhere" / "keep.mp4"
@@ -212,7 +214,7 @@ def test_cleanup_uploads_ignores_paths_outside_the_upload_root(tmp_path):


 def test_cleanup_uploads_tolerates_a_missing_file(tmp_path):
-    from karaoke_maker.web import _cleanup_uploads
+    from songtrace.web import _cleanup_uploads
    cfg = Config(whisper_url="u", work_dir=tmp_path / "work",
                 output_dir=tmp_path / "out")
    _cleanup_uploads(cfg, "child", cfg.work_dir / "uploads" / "child" / "gone.mp4")
@@ -340,13 +342,13 @@ def test_fonts_are_served(tmp_path):


 def test_font_licence_is_bundled():
-    from karaoke_maker.web import STATIC
+    from songtrace.web import STATIC
    licence = (STATIC / "fonts" / "OFL.txt").read_text(encoding="utf-8", errors="replace")
    assert "SIL OPEN FONT LICENSE" in licence.upper()


 def _index_html():
-    from karaoke_maker.web import STATIC
+    from songtrace.web import STATIC
    return (STATIC / "index.html").read_text(encoding="utf-8")


@@ -359,6 +361,32 @@ def test_index_has_required_hooks():
        assert f'name="{name}"' in html, name


+def test_video_source_fields_are_outside_advanced_details():
+    html = _index_html()
+    form_start = html.index('id="song-form"')
+    details_start = html.index("<details>", form_start)
+    before = html[form_start:details_start]
+    assert 'name="youtube_url"' in before
+    assert 'name="video"' in before
+    after = html[details_start:]
+    for name in ("genre", "year", "language", "bpm", "cover"):
+        assert f'name="{name}"' in after
+
+
+def test_song_form_lyrics_uses_a_two_column_grid():
+    css = _page_css(_index_html())
+    assert re.search(r"\.song-form-grid\s*\{[^}]*grid-template-columns:\s*1fr 1fr",
+                     css, re.S)
+
+
+def test_song_form_grid_stacks_on_narrow_screens():
+    css = _page_css(_index_html())
+    assert re.search(
+        r"@media \(max-width: 720px\)\s*\{[^}]*\.song-form-grid\s*\{[^}]*"
+        r"grid-template-columns:\s*1fr\s*;",
+        css, re.S)
+
+
 def test_index_makes_no_external_requests():
    import re as _re
    html = _index_html()
@@ -383,10 +411,53 @@ def test_index_respects_reduced_motion():
    assert "prefers-reduced-motion" in _index_html()


-def test_index_parses_warnings_with_endash():
+def test_index_has_lyrics_review_hook():
    html = _index_html()
-    assert "–" in html, "warning parser must handle the en-dash separator"
-    assert "data-start" in html and "data-end" in html
+    assert 'id="lyrics-review"' in html
+    assert 'id="warnings"' not in html
+
+
+def test_index_review_uses_svg_not_canvas():
+    html = _index_html()
+    assert "createElementNS" in html
+    assert "<canvas" not in html
+
+
+def test_index_review_fetches_review_endpoint():
+    html = _index_html()
+    assert "/review" in html
+
+
+def test_index_lyrics_review_uses_endash_for_time_ranges():
+    html = _index_html()
+    body = _fn_body(html, "missingSpanNode")
+    assert "–" in body
+
+
+def test_index_review_lays_out_words_in_time():
+    """Words must be positioned by their sung timing (percentage-based), not
+    packed as plain flow text, so notes can line up under their own word."""
+    html = _index_html()
+    body = _fn_body(html, "lineTimeline")
+    assert "review-word" in body
+    assert "style.left" in body
+
+
+def test_lyrics_review_cache_clears_on_job_status_change():
+    """Regression: switching between done/failed/running jobs must clear the
+    cache so a previously-viewed done job re-fetches when re-selected."""
+    html = _index_html()
+    body = _fn_body(html, "renderLyricsReview")
+    # The "failed" branch must clear the cache before rendering the placeholder
+    assert re.search(
+        r'if\s*\(\s*job\.status\s*===\s*"failed"\s*\)\s*\{[^}]*'
+        r'delete\s+box\.dataset\.jobId',
+        body, re.S), "failed branch must delete cache key"
+    # The "not done" (queued/running) branch must clear the cache before rendering
+    assert re.search(
+        r'if\s*\(\s*job\.status\s*!==\s*"done"\s*\)\s*\{[^}]*'
+        r'delete\s+box\.dataset\.jobId',
+        body, re.S), "not done branch must delete cache key"


 def test_index_has_copy_path_control():
@@ -513,11 +584,19 @@ def _submit_body(html):
    return html[start:html.index("\n});\n", start)]


+def _listener_body(html, prefix):
+    start = html.index(prefix)
+    return html[start:html.index("\n});\n", start)]
+
+
 def _rebuild_flaws(html):
    """Ways rebuild mode could quietly stop working."""
    flaws, submit = [], _submit_body(html)
    if "rebuildOf ?" not in submit or "/rebuild" not in submit:
        flaws.append("the form always posts to /api/jobs, so edits create a new song")
+    if "form.reset()" in submit or "exitRebuild()" in submit:
+        flaws.append("a successful submission empties the form, so correcting "
+                     "or varying the song just sent means retyping it")

    actions = _fn_body(html, "renderActions")
    if not re.search(r"if \(.*dataset\.jobId === job\.id.*\) return;", actions):
@@ -525,12 +604,128 @@ def _rebuild_flaws(html):
    if '"failed"' not in actions or '"done"' not in actions:
        flaws.append("the action row does not key off the job status")

-    start = _fn_body(html, "startRebuild")
-    if "/input" not in start or "form.elements[" not in start:
+    load = _fn_body(html, "loadInputs")
+    if "/input" not in load or "form.elements[" not in load:
        flaws.append("rebuild mode does not prefill the form from the job's input")
+
+    select = _fn_body(html, "selectJob")
+    if "loadInputs(" not in select:
+        flaws.append("selecting a queue entry does not bring its inputs back up")
+    if "formDirty" not in select or "confirm(" not in select:
+        flaws.append("selecting a queue entry discards unsaved edits without asking")
+
+    if "if (rebuildOf === id) exitRebuild();" not in _fn_body(html, "deleteJob"):
+        flaws.append("deleting the loaded job leaves the form pointed at it")
+
+    # selecting a queue entry loads it, so the sidebar button is now the only
+    # way back to a blank form
+    create = _listener_body(html, 'createSong.addEventListener("click"')
+    if "exitRebuild()" not in create:
+        flaws.append("the queue's CREATE SONG button does not clear the form")
+    if "formDirty" not in create or "confirm(" not in create:
+        flaws.append("the queue's CREATE SONG button discards edits without asking")
+    if "clearDetail(" not in create:
+        flaws.append("CREATE SONG empties the form but leaves the previous "
+                     "song's review and player sitting under it")
+    if "composing" not in create or "composing" not in _fn_body(html, "refreshJobs"):
+        flaws.append("the blank page CREATE SONG made is selected back over by "
+                     "the next poll of the queue")
+
+    mode = _fn_body(html, "setFormMode")
+    if 'querySelector("details").open' not in mode:
+        flaws.append("prefilled genre/year/language/bpm stay hidden behind the "
+                     "closed Advanced summary, so they rebuild unseen")
+    if "submitButton.disabled" not in _fn_body(html, "setSubmitLock"):
+        flaws.append("the form lets you submit a rebuild the API will answer 409")
    return flaws


+def _rebuild_full_and_delete_flaws(html):
+    flaws = []
+    actions = _fn_body(html, "renderActions")
+    for action in ('"rebuild-full"', '"delete"'):
+        if action not in actions:
+            flaws.append(f"{action} is not offered from renderActions")
+
+    rf = _fn_body(html, "rebuildFullJob")
+    if "/rebuild-full" not in rf or 'method: "POST"' not in rf:
+        flaws.append("rebuildFullJob does not POST to /rebuild-full")
+
+    d = _fn_body(html, "deleteJob")
+    confirm_idx = d.find("confirm(")
+    fetch_idx = d.find("fetch(")
+    if confirm_idx == -1:
+        flaws.append("deleteJob does not confirm before deleting")
+    elif fetch_idx != -1 and confirm_idx > fetch_idx:
+        flaws.append("deleteJob fetches before confirming")
+    if 'method: "DELETE"' not in d:
+        flaws.append("deleteJob does not send a DELETE request")
+    return flaws
+
+
+def test_rebuild_full_and_delete_controls_are_wired_up():
+    assert _rebuild_full_and_delete_flaws(_index_html()) == []
+
+
+def _empty_state_flaws(html):
+    refresh = _fn_body(html, "refreshJobs")
+    if "clearDetail(" not in refresh:
+        return ["refreshJobs never resets the panel when no job is selected"]
+    return []
+
+
+def test_review_panel_resets_when_no_job_is_selected():
+    assert _empty_state_flaws(_index_html()) == []
+
+
+def _playback_flaws(html):
+    """Ways the review could stop tracking the media element."""
+    flaws = []
+    source = _fn_body(html, "setPlayerSource")
+    if "media/audio" not in source:
+        flaws.append("the player is not pointed at the job's packaged audio, "
+                     "which is the file the game plays and the notes track")
+    if "media/video" not in source:
+        flaws.append("the packaged video is never shown alongside the audio")
+
+    # the mp3 and the video carry the same recording; playing both is an echo
+    if not re.search(r"<video[^>]*\bmuted\b", html):
+        flaws.append("the background video is not muted, so its own audio "
+                     "plays on top of the mp3")
+    pump = _fn_body(html, "pumpPlayback")
+    if "slaveVideo(" not in pump:
+        flaws.append("the picture is never pulled back onto the audio clock, "
+                     "so it drifts out of sync over a long song")
+
+    if "requestAnimationFrame" not in _fn_body(html, "pumpPlayback"):
+        flaws.append("sync is not driven by the frame loop, so the playhead "
+                     "steps at timeupdate's ~4Hz instead of following the audio")
+
+    if "media.paused" not in _fn_body(html, "focusActive"):
+        flaws.append("the view is scrolled even while paused, fighting a reader "
+                     "who scrolled away to look at something else")
+
+    sync = _fn_body(html, "syncPlayback")
+    if "clearEntry(" not in sync:
+        flaws.append("the line the playhead left keeps its highlight")
+    clear = _fn_body(html, "clearEntry")
+    for name in ("review-word-sung", "review-word-active", "review-line-active"):
+        if name not in clear:
+            flaws.append(f"{name} is never taken off a line the playhead left")
+
+    # every branch that gives up on rendering a review has to unload the media
+    # too, or the previous song stays loaded and playable under the new one
+    render = _fn_body(html, "renderLyricsReview")
+    if render.count("resetPlayer()") < render.count("emptyNote("):
+        flaws.append("a review the page cannot render leaves the previous "
+                     "song's media loaded in the player")
+    return flaws
+
+
+def test_review_playback_tracks_the_media_element():
+    assert _playback_flaws(_index_html()) == []
+
+
 # each mutation is a plausible regression; the checker above must reject all of them
 MUTATIONS = [
    (_focus_flaws, "queue: card moved unconditionally",
@@ -559,6 +754,57 @@ MUTATIONS = [
    (_rebuild_flaws, "rebuild: action row rebuilt on every poll",
     "if (reviewActions.dataset.jobId === job.id && shown.join() === wanted.join()) return;",
     ""),
+    (_rebuild_full_and_delete_flaws, "delete: confirmation dialog removed",
+     'if (!confirm("Delete this song? This removes all its files and cannot be undone.")) return;',
+     ""),
+    (_rebuild_full_and_delete_flaws, "rebuild-full: no longer offered on a done job",
+     '["rebuild-full", "delete"]', '["delete"]'),
+    (_rebuild_flaws, "prefill: selecting a queue entry no longer loads its inputs",
+     "  if (reload) await loadInputs(id, JOBS.find(j => j.id === id)?.status);\n", ""),
+    (_rebuild_flaws, "prefill: unsaved edits discarded without asking",
+     '  if (reload && formDirty && !confirm("Discard your unsaved edits to this form?")) return;\n',
+     ""),
+    (_rebuild_flaws, "prefill: deleting the loaded job leaves the form pointed at it",
+     "  if (rebuildOf === id) exitRebuild();\n", ""),
+    (_rebuild_flaws, "prefill: advanced fields stay hidden behind a closed summary",
+     'form.querySelector("details").open =', "const unusedAdvancedOpen ="),
+    (_rebuild_flaws, "prefill: an in-progress job can be submitted for rebuild anyway",
+     "submitButton.disabled = locked;", ""),
+    (_rebuild_flaws, "submit: a successful submission empties the form again",
+     "  rebuildOf = null;\n  formDirty = false;\n  setFormMode(null, null);\n  refreshJobs();",
+     "  exitRebuild();\n  refreshJobs();"),
+    (_rebuild_flaws, "create: the queue button no longer clears the form",
+     "  exitRebuild();\n  clearDetail(\"nothing selected\");\n", ""),
+    (_rebuild_flaws, "create: the review is left behind under a blank form",
+     '  clearDetail("nothing selected");\n', ""),
+    (_rebuild_flaws, "create: the next poll selects a song back into the blank page",
+     "  composing = true;\n", ""),
+    (_rebuild_flaws, "create: the queue button discards edits without asking",
+     '  if (formDirty && !confirm("Discard your unsaved edits to this form?")) return;\n  composing = true;',
+     "  composing = true;"),
+    (_playback_flaws, "playback: sync falls back to the timeupdate event's ~4Hz",
+     "syncFrame = media.paused ? null : requestAnimationFrame(pumpPlayback);",
+     "syncFrame = null;"),
+    (_playback_flaws, "playback: the view is scrolled even while paused",
+     "if (!activeEntry || media.paused) return;", "if (!activeEntry) return;"),
+    (_playback_flaws, "playback: the line the playhead left keeps its highlight",
+     "    if (activeEntry) clearEntry(activeEntry);\n", ""),
+    (_playback_flaws, "playback: a word stays lit after the playhead moves on",
+     'w.el.classList.remove("review-word-sung", "review-word-active");',
+     'w.el.classList.remove("review-word-sung");'),
+    (_playback_flaws, "playback: an unrenderable review leaves the old song loaded",
+     "    resetPlayer();\n", ""),
+    (_playback_flaws, "playback: the player is not pointed at the packaged audio",
+     'const src = available?.audio ? `/api/jobs/${job.id}/media/audio` : "";',
+     'const src = "";'),
+    (_playback_flaws, "playback: the background video plays its own audio too",
+     '<video id="player-video" muted playsinline',
+     '<video id="player-video" playsinline'),
+    (_playback_flaws, "playback: the picture is never re-synced to the audio",
+     "  slaveVideo(false);\n", ""),
+    (_empty_state_flaws, "empty state: clearDetail no longer wired into refreshJobs",
+     "if (selectedId) refreshDetail();\n  else clearDetail(composing ? \"nothing selected\" : undefined);",
+     "if (selectedId) refreshDetail();"),
 ]


@@ -594,7 +840,7 @@ def test_rebuild_mode_reuses_the_one_form():

 def test_index_has_rebuild_hooks():
    html = _index_html()
-    for hook in ("review-actions", "review-lineage", "cancel-rebuild",
+    for hook in ("review-actions", "review-lineage", "create-song",
                 'data-keep="video"', 'data-keep="cover"'):
        assert hook in html, hook

@@ -776,7 +1022,7 @@ def test_rebuild_renames_the_work_dir_on_a_retitle(tmp_path):
    seen = []
    client, cfg = make_client(tmp_path, capture_runner_factory(seen))
    parent = make_done_job(client)
-    (cfg.work_dir / "A - T").mkdir(parents=True)
+    (cfg.work_dir / "A - T").mkdir(parents=True, exist_ok=True)
    (cfg.work_dir / "A - T" / "audio.wav").write_bytes(b"x")
    child = client.post(f"/api/jobs/{parent}/rebuild",
                        data=rebuild_form(title="T2")).json()["id"]
@@ -789,7 +1035,7 @@ def test_rebuild_renames_the_work_dir_on_a_retitle(tmp_path):
 def test_rebuild_fails_the_job_on_a_work_dir_collision(tmp_path):
    client, cfg = make_client(tmp_path, ok_runner)
    parent = make_done_job(client)
-    (cfg.work_dir / "A - T").mkdir(parents=True)
+    (cfg.work_dir / "A - T").mkdir(parents=True, exist_ok=True)
    (cfg.work_dir / "A - T2").mkdir(parents=True)
    child = client.post(f"/api/jobs/{parent}/rebuild",
                        data=rebuild_form(title="T2")).json()["id"]
@@ -817,7 +1063,7 @@ def test_rebuild_inherits_the_parents_source(tmp_path):
                        files={"video": ("v.mp4", io.BytesIO(b"data"), "video/mp4")})
    parent = resp.json()["id"]
    wait_status(client, parent, "done")
-    (cfg.work_dir / "A - T").mkdir(parents=True)
+    (cfg.work_dir / "A - T").mkdir(parents=True, exist_ok=True)
    for name in ("meta.json", "audio.wav"):
        (cfg.work_dir / "A - T" / name).write_bytes(b"x")
    child = client.post(f"/api/jobs/{parent}/rebuild",
@@ -903,6 +1149,59 @@ def test_rebuild_of_an_unknown_job_is_404(tmp_path):
    assert client.post("/api/jobs/nope/rebuild", data=rebuild_form()).status_code == 404


+def test_rebuild_full_forces_every_stage_from_the_first(tmp_path):
+    seen = []
+    client, _ = make_client(tmp_path, capture_runner_factory(seen))
+    parent = make_done_job(client)
+    child = client.post(f"/api/jobs/{parent}/rebuild-full").json()["id"]
+    wait_status(client, child, "done")
+    assert seen[1]["force_from"] == STAGE_ORDER[0]
+    assert seen[1]["inp"] == seen[0]["inp"]
+
+
+def test_rebuild_full_creates_a_linked_job(tmp_path):
+    client, _ = make_client(tmp_path, ok_runner)
+    parent = make_done_job(client)
+    resp = client.post(f"/api/jobs/{parent}/rebuild-full")
+    assert resp.status_code == 201
+    child = resp.json()["id"]
+    assert child != parent
+    job = wait_status(client, child, "done")
+    assert job["rebuilt_from"] == parent
+
+
+def test_rebuild_full_logs_that_it_is_a_forced_rebuild(tmp_path):
+    client, _ = make_client(tmp_path, ok_runner)
+    parent = make_done_job(client)
+    child = client.post(f"/api/jobs/{parent}/rebuild-full").json()["id"]
+    job = wait_status(client, child, "done")
+    rebuild_lines = [l for l in job["log"] if l.startswith("[rebuild]")]
+    assert len(rebuild_lines) == 1
+    assert STAGE_ORDER[0] in rebuild_lines[0]
+
+
+def test_rebuild_full_of_an_unknown_job_is_404(tmp_path):
+    client, _ = make_client(tmp_path, ok_runner)
+    assert client.post("/api/jobs/nope/rebuild-full").status_code == 404
+
+
+def test_rebuild_full_of_a_running_job_is_409(tmp_path):
+    release = threading.Event()
+
+    def blocking(inp, cfg, log=print, force_from=None, stages=None):
+        release.wait(5)
+        return PipelineResult(output_dir=cfg.output_dir / "X")
+
+    client, _ = make_client(tmp_path, blocking)
+    parent = client.post("/api/jobs", data=rebuild_form()).json()["id"]
+    try:
+        wait_status(client, parent, "running")
+        assert client.post(f"/api/jobs/{parent}/rebuild-full").status_code == 409
+    finally:
+        release.set()
+    wait_status(client, parent, "done")
+
+
 def slug_runner(inp, cfg, log=print, force_from=None, stages=None):
    out = cfg.output_dir / song_slug(inp)
    out.mkdir(parents=True, exist_ok=True)
@@ -951,6 +1250,76 @@ def test_rebuild_of_a_failed_parent_removes_nothing(tmp_path):
    assert (cfg.output_dir / "A - T" / "keep.txt").exists()


+def test_delete_removes_work_and_output_and_the_job(tmp_path):
+    client, cfg = make_client(tmp_path, slug_runner)
+    job_id = make_done_job(client)
+    (cfg.work_dir / "A - T").mkdir(parents=True, exist_ok=True)
+    (cfg.work_dir / "A - T" / "audio.wav").write_bytes(b"x")
+    assert (cfg.output_dir / "A - T").exists()
+
+    resp = client.delete(f"/api/jobs/{job_id}")
+
+    assert resp.status_code == 204
+    assert not (cfg.work_dir / "A - T").exists()
+    assert not (cfg.output_dir / "A - T").exists()
+    assert client.get(f"/api/jobs/{job_id}").status_code == 404
+    assert job_id not in [j["id"] for j in client.get("/api/jobs").json()]
+
+
+def test_delete_tolerates_a_work_dir_that_is_already_gone(tmp_path):
+    client, cfg = make_client(tmp_path, slug_runner)
+    job_id = make_done_job(client)
+    shutil.rmtree(cfg.work_dir / "A - T")
+    resp = client.delete(f"/api/jobs/{job_id}")
+    assert resp.status_code == 204
+
+
+def test_delete_of_an_unknown_job_is_404(tmp_path):
+    client, _ = make_client(tmp_path, ok_runner)
+    assert client.delete("/api/jobs/nope").status_code == 404
+
+
+def test_delete_of_a_running_job_is_409(tmp_path):
+    release = threading.Event()
+
+    def blocking(inp, cfg, log=print, force_from=None, stages=None):
+        release.wait(5)
+        return PipelineResult(output_dir=cfg.output_dir / "X")
+
+    client, _ = make_client(tmp_path, blocking)
+    parent = client.post("/api/jobs", data=rebuild_form()).json()["id"]
+    try:
+        wait_status(client, parent, "running")
+        assert client.delete(f"/api/jobs/{parent}").status_code == 409
+    finally:
+        release.set()
+    wait_status(client, parent, "done")
+    assert client.get(f"/api/jobs/{parent}").status_code == 200
+
+
+def test_delete_of_a_job_with_an_active_rebuild_is_409(tmp_path):
+    """A done job's slug must not be deleted out from under a rebuild in flight."""
+    release = threading.Event()
+    calls = []
+
+    def runner(inp, cfg, log=print, force_from=None, stages=None):
+        calls.append(1)
+        if len(calls) > 1:
+            release.wait(5)
+        return PipelineResult(output_dir=cfg.output_dir / "X")
+
+    client, _ = make_client(tmp_path, runner)
+    parent = make_done_job(client)
+    child = client.post(f"/api/jobs/{parent}/rebuild-full").json()["id"]
+    try:
+        wait_status(client, child, "running")
+        assert client.delete(f"/api/jobs/{parent}").status_code == 409
+    finally:
+        release.set()
+    wait_status(client, child, "done")
+    assert client.get(f"/api/jobs/{parent}").status_code == 200
+
+
 def test_a_plain_job_never_removes_anything(tmp_path):
    client, cfg = make_client(tmp_path, slug_runner)
    make_done_job(client)
@@ -958,3 +1327,422 @@ def test_a_plain_job_never_removes_anything(tmp_path):
    stray.mkdir(parents=True)
    make_done_job(client, title="T2")
    assert stray.exists()
+
+
+def test_discover_jobs_ignores_a_work_dir_without_a_sidecar(tmp_path):
+    from songtrace.web import _discover_jobs
+    cfg = Config(whisper_url="u", work_dir=tmp_path / "work", output_dir=tmp_path / "out")
+    (cfg.work_dir / "A - T").mkdir(parents=True)
+    assert _discover_jobs(cfg) == []
+
+
+def test_discover_jobs_reconstructs_a_completed_song(tmp_path):
+    from songtrace.web import _discover_jobs
+    cfg = Config(whisper_url="u", work_dir=tmp_path / "work", output_dir=tmp_path / "out")
+    client = TestClient(create_app(cfg, runner=slug_runner))
+    make_done_job(client)
+
+    discovered = _discover_jobs(cfg)
+
+    assert len(discovered) == 1
+    job = discovered[0]
+    assert job.title == "T" and job.artist == "A" and job.status == "done"
+    assert job.output_dir == str(cfg.output_dir / "A - T")
+    assert job.error is None
+
+
+def test_discover_jobs_reconstructs_a_song_whose_output_is_missing_as_failed(tmp_path):
+    from songtrace.web import _discover_jobs
+    cfg = Config(whisper_url="u", work_dir=tmp_path / "work", output_dir=tmp_path / "out")
+    client = TestClient(create_app(cfg, runner=slug_runner))
+    make_done_job(client)
+    shutil.rmtree(cfg.output_dir / "A - T")
+
+    discovered = _discover_jobs(cfg)
+
+    assert len(discovered) == 1
+    assert discovered[0].status == "failed"
+    assert discovered[0].error == "interrupted before completion"
+    assert discovered[0].output_dir is None
+
+
+def test_discovered_jobs_are_visible_from_a_fresh_app(tmp_path):
+    cfg = Config(whisper_url="u", work_dir=tmp_path / "work", output_dir=tmp_path / "out")
+    client1 = TestClient(create_app(cfg, runner=slug_runner))
+    make_done_job(client1)
+
+    client2 = TestClient(create_app(cfg, runner=slug_runner))
+    jobs = client2.get("/api/jobs").json()
+    assert len(jobs) == 1
+    assert jobs[0]["title"] == "T" and jobs[0]["status"] == "done"
+
+
+def review_runner_factory(notes, dropped_lines=None, missing_spans=None, warnings=None,
+                          words=None):
+    def runner(inp, cfg, log=print, force_from=None, stages=None):
+        work = cfg.work_dir / song_slug(inp)
+        work.mkdir(parents=True, exist_ok=True)
+        (work / "notes_sec.json").write_text(json.dumps({
+            "notes": notes, "breaks": [], "warnings": warnings or []}))
+        if dropped_lines is not None or missing_spans is not None:
+            (work / "align_meta.json").write_text(json.dumps({
+                "dropped_lines": dropped_lines or [],
+                "missing_spans": missing_spans or []}))
+        if words is not None:
+            (work / "alignment.json").write_text(json.dumps(words))
+        out = cfg.output_dir / song_slug(inp)
+        out.mkdir(parents=True, exist_ok=True)
+        return PipelineResult(output_dir=out, warnings=warnings or [])
+    return runner
+
+
+def test_review_endpoint_groups_notes_by_line(tmp_path):
+    notes = [
+        {"start": 1.0, "duration": 0.4, "pitch": 0, "text": "hi",
+         "low_confidence": False, "line_index": 0},
+        {"start": 5.0, "duration": 0.3, "pitch": 2, "text": "there",
+         "low_confidence": True, "line_index": 1},
+    ]
+    client, _ = make_client(tmp_path, review_runner_factory(
+        notes, dropped_lines=[2], missing_spans=[[8.0, 9.5]]))
+    resp = client.post("/api/jobs", data={
+        "title": "T", "artist": "A", "lyrics": "hi\nthere\ngone",
+        "youtube_url": "http://y"})
+    job_id = resp.json()["id"]
+    wait_status(client, job_id, "done")
+    body = client.get(f"/api/jobs/{job_id}/review").json()
+    assert body["lines"] == [
+        {"index": 0, "text": "hi", "dropped": False,
+         "notes": [{"start": 1.0, "end": 1.4, "pitch": 0, "low_confidence": False}],
+         "words": []},
+        {"index": 1, "text": "there", "dropped": False,
+         "notes": [{"start": 5.0, "end": 5.3, "pitch": 2, "low_confidence": True}],
+         "words": []},
+        {"index": 2, "text": "gone", "dropped": True, "notes": [], "words": []},
+    ]
+    assert body["missing_spans"] == [{"start": 8.0, "end": 9.5}]
+
+
+def test_review_endpoint_includes_per_word_timing(tmp_path):
+    notes = [
+        {"start": 1.0, "duration": 0.4, "pitch": 0, "text": "hi",
+         "low_confidence": False, "line_index": 0},
+        {"start": 5.0, "duration": 0.3, "pitch": 2, "text": "there",
+         "low_confidence": True, "line_index": 1},
+    ]
+    words = [
+        {"word": "hi", "start": 1.0, "end": 1.4, "score": 0.9, "line_index": 0},
+        {"word": "there", "start": 5.0, "end": 5.3, "score": 0.9, "line_index": 1},
+    ]
+    client, _ = make_client(tmp_path, review_runner_factory(
+        notes, dropped_lines=[2], missing_spans=[[8.0, 9.5]], words=words))
+    resp = client.post("/api/jobs", data={
+        "title": "T", "artist": "A", "lyrics": "hi\nthere\ngone",
+        "youtube_url": "http://y"})
+    job_id = resp.json()["id"]
+    wait_status(client, job_id, "done")
+    body = client.get(f"/api/jobs/{job_id}/review").json()
+    assert body["lines"][0]["words"] == [{"text": "hi", "start": 1.0, "end": 1.4}]
+    assert body["lines"][1]["words"] == [{"text": "there", "start": 5.0, "end": 5.3}]
+    assert body["lines"][2]["dropped"] is True
+    assert body["lines"][2]["words"] == []
+
+
+def test_review_of_a_failed_job_is_404(tmp_path):
+    def bad_runner(*a, **kw):
+        raise PipelineError("boom")
+    client, _ = make_client(tmp_path, bad_runner)
+    resp = client.post("/api/jobs", data={
+        "title": "T", "artist": "A", "lyrics": "x", "youtube_url": "http://y"})
+    job_id = resp.json()["id"]
+    wait_status(client, job_id, "failed")
+    assert client.get(f"/api/jobs/{job_id}/review").status_code == 404
+
+
+def test_review_of_a_running_job_is_409(tmp_path):
+    release = threading.Event()
+
+    def blocking(inp, cfg, log=print, force_from=None, stages=None):
+        release.wait(5)
+        return PipelineResult(output_dir=cfg.output_dir / "X")
+
+    client, _ = make_client(tmp_path, blocking)
+    resp = client.post("/api/jobs", data={
+        "title": "T", "artist": "A", "lyrics": "x", "youtube_url": "http://y"})
+    job_id = resp.json()["id"]
+    try:
+        wait_status(client, job_id, "running")
+        assert client.get(f"/api/jobs/{job_id}/review").status_code == 409
+    finally:
+        release.set()
+    wait_status(client, job_id, "done")
+
+
+def test_review_of_an_unknown_job_is_404(tmp_path):
+    client, _ = make_client(tmp_path, ok_runner)
+    assert client.get("/api/jobs/nope/review").status_code == 404
+
+
+def test_review_missing_when_pipeline_never_wrote_notes(tmp_path):
+    client, _ = make_client(tmp_path, ok_runner)
+    job_id = make_done_job(client)
+    assert client.get(f"/api/jobs/{job_id}/review").status_code == 404
+
+
+# --- review playback -------------------------------------------------------
+
+MEDIA_FILES = {"A - T.mp3": b"ID3audio", "A - T.mp4": b"\0\0\0\x18ftypvideo",
+               "A - T.jpg": b"\xff\xd8cover"}
+MEDIA_HEADERS = [("TITLE", "T"), ("ARTIST", "A"), ("MP3", "A - T.mp3"),
+                 ("COVER", "A - T.jpg"), ("VIDEO", "A - T.mp4")]
+
+
+def media_runner_factory(files=None, headers=None, write_txt=True):
+    """A run that leaves behind a song folder, the way package does."""
+    def runner(inp, cfg, log=print, force_from=None, stages=None):
+        slug = song_slug(inp)
+        out = cfg.output_dir / slug
+        out.mkdir(parents=True, exist_ok=True)
+        for name, body in (MEDIA_FILES if files is None else files).items():
+            (out / name).write_bytes(body)
+        if write_txt:
+            rows = [f"#{k}:{v}" for k, v in
+                    (MEDIA_HEADERS if headers is None else headers)]
+            (out / f"{slug}.txt").write_text("\n".join(rows) + "\n: 0 4 0 la\nE\n",
+                                             encoding="utf-8")
+        return PipelineResult(output_dir=out)
+    return runner
+
+
+def test_media_serves_the_files_the_song_txt_names(tmp_path):
+    client, _ = make_client(tmp_path, media_runner_factory())
+    job_id = make_done_job(client)
+    for kind, name in (("audio", "A - T.mp3"), ("video", "A - T.mp4"),
+                       ("cover", "A - T.jpg")):
+        resp = client.get(f"/api/jobs/{job_id}/media/{kind}")
+        assert resp.status_code == 200, kind
+        assert resp.content == MEDIA_FILES[name], kind
+
+
+def test_media_serves_whatever_the_header_was_rewritten_to(tmp_path):
+    """package transcodes to .mp4 and rewrites #VIDEO; the extension is not fixed."""
+    files = {**MEDIA_FILES, "A - T.webm": b"transcoded"}
+    headers = [(k, "A - T.webm" if k == "VIDEO" else v) for k, v in MEDIA_HEADERS]
+    client, _ = make_client(tmp_path, media_runner_factory(files, headers))
+    job_id = make_done_job(client)
+    assert client.get(f"/api/jobs/{job_id}/media/video").content == b"transcoded"
+
+
+def test_media_supports_range_requests(tmp_path):
+    """Without 206 the browser cannot seek, which is the point of the player."""
+    client, _ = make_client(tmp_path, media_runner_factory())
+    job_id = make_done_job(client)
+    resp = client.get(f"/api/jobs/{job_id}/media/video", headers={"Range": "bytes=0-3"})
+    assert resp.status_code == 206
+    assert resp.headers["content-range"] == f"bytes 0-3/{len(MEDIA_FILES['A - T.mp4'])}"
+    assert resp.content == MEDIA_FILES["A - T.mp4"][:4]
+
+
+def test_media_of_an_unknown_kind_is_404(tmp_path):
+    client, _ = make_client(tmp_path, media_runner_factory())
+    job_id = make_done_job(client)
+    assert client.get(f"/api/jobs/{job_id}/media/instrumental").status_code == 404
+
+
+def test_media_refuses_a_header_pointing_outside_the_song_folder(tmp_path):
+    secret = tmp_path / "secret.txt"
+    secret.write_text("not yours")
+    headers = [(k, "../../secret.txt" if k == "MP3" else v) for k, v in MEDIA_HEADERS]
+    client, _ = make_client(tmp_path, media_runner_factory(headers=headers))
+    job_id = make_done_job(client)
+    assert client.get(f"/api/jobs/{job_id}/media/audio").status_code == 404
+
+
+def test_media_refuses_an_absolute_header_path(tmp_path):
+    secret = tmp_path / "secret.txt"
+    secret.write_text("not yours")
+    headers = [(k, str(secret) if k == "MP3" else v) for k, v in MEDIA_HEADERS]
+    client, _ = make_client(tmp_path, media_runner_factory(headers=headers))
+    job_id = make_done_job(client)
+    assert client.get(f"/api/jobs/{job_id}/media/audio").status_code == 404
+
+
+def test_media_404s_when_the_named_file_is_missing(tmp_path):
+    client, _ = make_client(tmp_path, media_runner_factory(files={}))
+    job_id = make_done_job(client)
+    assert client.get(f"/api/jobs/{job_id}/media/audio").status_code == 404
+
+
+def test_media_404s_when_the_header_was_dropped(tmp_path):
+    """package strips #COVER when there is no usable cover image."""
+    headers = [(k, v) for k, v in MEDIA_HEADERS if k != "COVER"]
+    client, _ = make_client(tmp_path, media_runner_factory(headers=headers))
+    job_id = make_done_job(client)
+    assert client.get(f"/api/jobs/{job_id}/media/cover").status_code == 404
+    assert client.get(f"/api/jobs/{job_id}/media/audio").status_code == 200
+
+
+def test_media_404s_when_there_is_no_song_txt(tmp_path):
+    client, _ = make_client(tmp_path, media_runner_factory(write_txt=False))
+    job_id = make_done_job(client)
+    assert client.get(f"/api/jobs/{job_id}/media/audio").status_code == 404
+
+
+def test_media_of_a_failed_job_is_404(tmp_path):
+    def bad_runner(*a, **kw):
+        raise PipelineError("boom")
+    client, _ = make_client(tmp_path, bad_runner)
+    job_id = client.post("/api/jobs", data={
+        "title": "T", "artist": "A", "lyrics": "x", "youtube_url": "http://y"}).json()["id"]
+    wait_status(client, job_id, "failed")
+    assert client.get(f"/api/jobs/{job_id}/media/audio").status_code == 404
+
+
+def test_media_of_a_running_job_is_409(tmp_path):
+    release = threading.Event()
+
+    def blocking(inp, cfg, log=print, force_from=None, stages=None):
+        release.wait(5)
+        return PipelineResult(output_dir=cfg.output_dir / "X")
+
+    client, _ = make_client(tmp_path, blocking)
+    job_id = client.post("/api/jobs", data={
+        "title": "T", "artist": "A", "lyrics": "x", "youtube_url": "http://y"}).json()["id"]
+    try:
+        wait_status(client, job_id, "running")
+        assert client.get(f"/api/jobs/{job_id}/media/audio").status_code == 409
+    finally:
+        release.set()
+    wait_status(client, job_id, "done")
+
+
+def test_media_of_an_unknown_job_is_404(tmp_path):
+    client, _ = make_client(tmp_path, media_runner_factory())
+    assert client.get("/api/jobs/nope/media/audio").status_code == 404
+
+
+def test_review_reports_which_media_the_player_can_load(tmp_path):
+    notes = [{"start": 1.0, "duration": 0.4, "pitch": 0, "text": "hi",
+              "low_confidence": False, "line_index": 0}]
+    review = review_runner_factory(notes)
+    media = media_runner_factory(headers=[(k, v) for k, v in MEDIA_HEADERS
+                                          if k != "COVER"])
+
+    def runner(inp, cfg, **kw):
+        media(inp, cfg, **kw)
+        return review(inp, cfg, **kw)
+
+    client, _ = make_client(tmp_path, runner)
+    job_id = make_done_job(client, lyrics="hi")
+    body = client.get(f"/api/jobs/{job_id}/review").json()
+    assert body["media"] == {"audio": True, "video": True, "cover": False}
+
+
+# --- restarting a rebuild from a chosen stage ------------------------------
+
+def force_from_runner_factory(seen):
+    def runner(inp, cfg, log=print, force_from=None, stages=None):
+        seen.append(force_from)
+        out = cfg.output_dir / song_slug(inp)
+        out.mkdir(parents=True, exist_ok=True)
+        return PipelineResult(output_dir=out)
+    return runner
+
+
+def rebuild_with(client, job_id, **extra):
+    data = {"title": "T", "artist": "A", "lyrics": "x", "youtube_url": "http://y"}
+    data.update(extra)
+    return client.post(f"/api/jobs/{job_id}/rebuild", data=data)
+
+
+def test_rebuild_without_edits_restarts_from_the_stage_asked_for(tmp_path):
+    """The point of the control: a change in songtrace invalidates no field."""
+    seen = []
+    client, _ = make_client(tmp_path, force_from_runner_factory(seen))
+    parent = make_done_job(client)
+    resp = rebuild_with(client, parent, force_from="package")
+    assert resp.status_code == 201
+    wait_status(client, resp.json()["id"], "done")
+    assert seen[-1] == "package"
+
+
+def test_rebuild_without_a_stage_still_plans_from_the_edit(tmp_path):
+    seen = []
+    client, _ = make_client(tmp_path, force_from_runner_factory(seen))
+    parent = make_done_job(client)
+    resp = rebuild_with(client, parent, lyrics="different words")
+    wait_status(client, resp.json()["id"], "done")
+    assert seen[-1] == "align"
+
+
+def test_an_edit_wins_when_it_invalidates_more_than_the_stage_asked_for(tmp_path):
+    """A choice may widen what is redone, never narrow it below correctness."""
+    seen = []
+    client, _ = make_client(tmp_path, force_from_runner_factory(seen))
+    parent = make_done_job(client)
+    resp = rebuild_with(client, parent, lyrics="different words",
+                        force_from="package")
+    wait_status(client, resp.json()["id"], "done")
+    assert seen[-1] == "align"
+
+
+def test_the_stage_asked_for_wins_when_it_is_earlier_than_the_edit(tmp_path):
+    seen = []
+    client, _ = make_client(tmp_path, force_from_runner_factory(seen))
+    parent = make_done_job(client)
+    resp = rebuild_with(client, parent, title="Renamed", force_from="separate")
+    wait_status(client, resp.json()["id"], "done")
+    assert seen[-1] == "separate"
+
+
+def test_rebuild_rejects_a_stage_that_is_not_in_the_pipeline(tmp_path):
+    client, _ = make_client(tmp_path, ok_runner)
+    parent = make_done_job(client)
+    resp = rebuild_with(client, parent, force_from="deechoing")
+    assert resp.status_code == 400
+    assert "deechoing" in resp.json()["detail"]
+
+
+def test_a_blank_stage_is_the_same_as_not_asking(tmp_path):
+    seen = []
+    client, _ = make_client(tmp_path, force_from_runner_factory(seen))
+    parent = make_done_job(client)
+    resp = rebuild_with(client, parent, force_from="")
+    wait_status(client, resp.json()["id"], "done")
+    assert seen[-1] is None
+
+
+def test_rebuild_from_scratch_still_restarts_at_the_first_stage(tmp_path):
+    seen = []
+    client, _ = make_client(tmp_path, force_from_runner_factory(seen))
+    parent = make_done_job(client)
+    resp = client.post(f"/api/jobs/{parent}/rebuild-full")
+    wait_status(client, resp.json()["id"], "done")
+    assert seen[-1] == STAGE_ORDER[0]
+
+
+def test_the_restart_control_offers_every_stage():
+    html = _index_html()
+    assert 'name="force_from"' in html
+    stages = _fn_body(html, "loadStages")
+    assert "force_from" in stages and "STAGES.map" in stages
+    mode = _fn_body(html, "setFormMode")
+    assert "restart-field" in mode, "the control is shown outside rebuild mode"
+
+
+def test_a_textareas_crlf_does_not_read_as_a_lyrics_edit(tmp_path):
+    """Browsers submit textareas with CRLF; the stored lyrics use LF."""
+    seen = []
+    client, _ = make_client(tmp_path, force_from_runner_factory(seen))
+    parent = make_done_job(client, lyrics="one\ntwo")
+    resp = rebuild_with(client, parent, lyrics="one\r\ntwo")
+    wait_status(client, resp.json()["id"], "done")
+    assert seen[-1] is None, "the rebuild redid the alignment for nothing"
+    body = client.get(f"/api/jobs/{resp.json()['id']}/input").json()
+    assert body["lyrics"] == "one\ntwo"
+
+
+def test_submitting_crlf_lyrics_stores_them_with_lf(tmp_path):
+    client, _ = make_client(tmp_path, ok_runner)
+    job_id = make_done_job(client, lyrics="one\r\ntwo")
+    assert client.get(f"/api/jobs/{job_id}/input").json()["lyrics"] == "one\ntwo"
@@ -1,7 +1,7 @@
 import httpx
 from pathlib import Path
-from karaoke_maker.config import Config
-from karaoke_maker.whisper_client import TranscriptSegment, transcribe
+from songtrace.config import Config
+from songtrace.whisper_client import TranscriptSegment, transcribe

 def test_transcribe_parses_segments(tmp_path):
    audio = tmp_path / "v.wav"
@@ -36,3 +36,25 @@ def test_transcribe_maps_language_to_iso_code(tmp_path):
    transcribe(cfg, audio, language="Japanese", client=client)
    assert 'name="language"' in seen["body"]
    assert "\r\n\r\nja\r\n" in seen["body"]
+
+
+def test_transcribe_asr_webservice_hits_fixed_asr_endpoint(tmp_path):
+    audio = tmp_path / "v.wav"
+    audio.write_bytes(b"RIFF")
+    seen = {}
+
+    def handler(request: httpx.Request) -> httpx.Response:
+        seen["url"] = str(request.url)
+        seen["body"] = request.read().decode("utf-8", errors="replace")
+        return httpx.Response(200, json={
+            "text": "hello world",
+            "segments": [{"start": 0.5, "end": 2.0, "text": " hello world"}],
+        })
+
+    client = httpx.Client(transport=httpx.MockTransport(handler))
+    cfg = Config(whisper_url="http://w:9000", whisper_api="asr-webservice")
+    segs = transcribe(cfg, audio, language="Japanese", client=client)
+    assert segs == [TranscriptSegment(start=0.5, end=2.0, text="hello world")]
+    assert seen["url"] == "http://w:9000/asr?output=json&language=ja"
+    assert 'name="audio_file"' in seen["body"]
+    assert 'name="model"' not in seen["body"]
@@ -0,0 +1,139 @@
+import os
+import time
+
+import pytest
+
+from songtrace import ytdlp
+from songtrace.config import Config
+from songtrace.ytdlp import YtDlp, YtDlpError
+
+
+def fake_package(path, version="2026.09.01", body=""):
+    """A stand-in yt_dlp that answers --version and otherwise runs `body`."""
+    pkg = path / "yt_dlp"
+    pkg.mkdir(parents=True)
+    (pkg / "__init__.py").touch()
+    (pkg / "__main__.py").write_text(
+        "import sys\n"
+        f"if '--version' in sys.argv: print({version!r}); sys.exit()\n"
+        + body)
+    return path
+
+
+@pytest.fixture
+def cache(tmp_path, monkeypatch):
+    monkeypatch.setenv("XDG_CACHE_HOME", str(tmp_path / "cache"))
+    monkeypatch.setattr(ytdlp, "bundled", lambda: YtDlp("2026.8.19"))
+    return ytdlp.cache_dir()
+
+
+@pytest.fixture
+def pypi(monkeypatch):
+    """Stubs PyPI and pip: `latest` is the newest release, `installs` logs pip."""
+    state = {"latest": "2026.9.1", "installs": []}
+    monkeypatch.setattr(ytdlp, "_latest_release", lambda: state["latest"])
+    def pip_install(target, version):
+        state["installs"].append(version)
+        fake_package(target, version)
+    monkeypatch.setattr(ytdlp, "_pip_install", pip_install)
+    return state
+
+
+def cfg(version="auto"):
+    return Config(whisper_url="http://x/v1", ytdlp_version=version)
+
+
+def test_run_uses_the_install_on_its_path_and_logs_its_output(tmp_path):
+    path = fake_package(tmp_path, body="print('[youtube] hello', sys.argv[1:])\n")
+    logged = []
+    YtDlp("2026.09.01", path).run(["--", "https://yt/x"], logged.append)
+    assert logged == ["[youtube] hello ['--ignore-config', '--', 'https://yt/x']"]
+
+
+def test_run_raises_the_error_line(tmp_path):
+    path = fake_package(tmp_path, body=(
+        "print('ERROR: HTTP Error 403: Forbidden', file=sys.stderr)\n"
+        "print('trailing noise')\nsys.exit(1)\n"))
+    with pytest.raises(YtDlpError, match="^ERROR: HTTP Error 403"):
+        YtDlp("2026.09.01", path).run([], lambda m: None)
+
+
+def test_run_flags_a_usage_error_as_incompatible(tmp_path):
+    path = fake_package(tmp_path, body=(
+        "print('error: no such option: --retries', file=sys.stderr)\n"
+        "sys.exit(2)\n"))
+    with pytest.raises(YtDlpError, match="2026.09.01 rejected.*--retries"):
+        YtDlp("2026.09.01", path).run([], lambda m: None)
+
+
+def test_auto_installs_a_newer_release_and_tries_it_first(cache, pypi):
+    found = ytdlp.candidates(cfg(), lambda m: None)
+    assert [y.version for y in found] == ["2026.9.1", "2026.8.19"]
+    assert found[0].path == cache / "2026.9.1"
+    assert not [p for p in cache.iterdir() if p.name.startswith(".tmp-")]
+
+
+def test_auto_checks_pypi_once_a_day(cache, pypi):
+    ytdlp.candidates(cfg(), lambda m: None)
+    pypi["latest"] = "2026.10.1"
+    found = ytdlp.candidates(cfg(), lambda m: None)
+    assert found[0].version == "2026.9.1"
+    found = ytdlp.candidates(cfg(), lambda m: None, refresh=True)
+    assert found[0].version == "2026.10.1"
+    assert not (cache / "2026.9.1").exists()
+    old = time.time() - ytdlp.CHECK_INTERVAL - 1
+    os.utime(cache / ".checked", (old, old))
+    pypi["latest"] = "2026.11.1"
+    assert ytdlp.candidates(cfg(), lambda m: None)[0].version == "2026.11.1"
+
+
+def test_auto_uses_only_bundled_when_it_is_newest(cache, pypi):
+    pypi["latest"] = "2026.8.19"
+    assert [y.version for y in ytdlp.candidates(cfg(), lambda m: None)] == [
+        "2026.8.19"]
+    assert pypi["installs"] == []
+
+
+def test_auto_survives_pypi_being_unreachable(cache, pypi, monkeypatch):
+    ytdlp.candidates(cfg(), lambda m: None)
+    def down():
+        raise OSError("network unreachable")
+    monkeypatch.setattr(ytdlp, "_latest_release", down)
+    logged = []
+    found = ytdlp.candidates(cfg(), logged.append, refresh=True)
+    assert [y.version for y in found] == ["2026.9.1", "2026.8.19"]
+    assert any("network unreachable" in m for m in logged)
+
+
+def test_an_install_that_fails_its_check_is_discarded(cache, pypi, monkeypatch):
+    monkeypatch.setattr(ytdlp, "_pip_install",
+                        lambda target, v: fake_package(target, "2026.1.1"))
+    logged = []
+    assert [y.version for y in ytdlp.candidates(cfg(), logged.append)] == [
+        "2026.8.19"]
+    assert any("failed its check" in m for m in logged)
+    assert list(cache.iterdir()) == [cache / ".checked"]
+
+
+def test_bundled_never_touches_pypi(cache, pypi):
+    pypi["latest"] = None
+    assert [y.version for y in ytdlp.candidates(cfg("bundled"), print)] == [
+        "2026.8.19"]
+    assert not cache.exists()
+
+
+def test_pin_installs_that_release_even_if_older(cache, pypi):
+    found = ytdlp.candidates(cfg("2026.07.01"), lambda m: None)
+    assert [y.version for y in found] == ["2026.07.01", "2026.8.19"]
+    assert pypi["installs"] == ["2026.07.01"]
+
+
+def test_pin_matching_bundled_installs_nothing(cache, pypi):
+    found = ytdlp.candidates(cfg("2026.08.19"), lambda m: None)
+    assert [y.version for y in found] == ["2026.8.19"]
+    assert pypi["installs"] == []
+
+
+def test_pin_must_be_a_release_number(cache, pypi):
+    with pytest.raises(YtDlpError, match="ytdlp_version"):
+        ytdlp.candidates(cfg("latest; rm -rf /"), lambda m: None)
Author	SHA1	Message	Date
branderson	4272189bca	tests: replace copyrighted lyrics with placeholder text The align, transcribe, lyrics and notes tests quoted lines from published songs, and the UltraStar parser fixtures were complete community charts of commercial songs. Replace them with counting words and two hand-written charts that keep what the parser tests rely on: CRLF line endings, Latin-1 bytes, a decimal-comma BPM, freestyle/golden/rap notes, negative pitch and a two-number line break.	2026-10-09 17:17:02 -07:00
branderson	936641e13c	Merge pull request 'Keep matched words that fall inside a missing-lyrics span' (#13 ) from missing-notes into main Reviewed-on: #13 Reviewed-by: branderson <brad@branderson.io>	2026-10-09 16:54:17 -07:00
branderson	793817ce42	notes: keep matched words that fall inside a missing-lyrics span A missing span covers whole Whisper segments, so it often also holds lines that were heard and matched. Skipping every aligned word starting inside one dropped those lines' notes and left empty line breaks in the chart. Words from lines never heard are already excluded via dropped_lines, so the span now only drives the warning and the review marker.	2026-10-09 16:52:56 -07:00
branderson	f3125370d3	Merge pull request 'Keep yt-dlp current at runtime, with the bundled copy as fallback' (#12 ) from ytdlp-runtime into main Reviewed-on: #12 Reviewed-by: branderson <brad@branderson.io>	2026-10-09 16:43:24 -07:00
branderson	6d297a8bfc	acquire: keep yt-dlp current at runtime, with the bundled copy as fallback YouTube breaks old yt-dlp releases faster than songtrace is released. songtrace now checks PyPI daily, and again after a download fails outright, installs a newer stable release into its cache dir and tries it before the bundled copy, which is never modified. yt-dlp runs in a subprocess through its CLI, all from one module, so a cached release never mixes with the imported one. ytdlp_version pins a release or turns updating off.	2026-10-09 14:36:29 -07:00
branderson	4af78fdd1d	Merge pull request 'Release the GPU after every stage' (#11 ) from gpu-unload into main Reviewed-on: #11 Reviewed-by: branderson <brad@branderson.io>	2026-10-06 13:44:44 -07:00
branderson	cb485d3c6b	Merge pull request 'webui: play a finished song against its review' (#10 ) from review-playback into main Reviewed-on: #10 Reviewed-by: branderson <brad@branderson.io>	2026-10-06 13:44:23 -07:00
branderson	f187bfd6c0	Release the GPU after every stage songtrace serve is long-lived but only wants the GPU in bursts, and until now it kept everything it ever loaded. align cached wav2vec2 in a module dict that was never cleared, torchcrepe caches its model on its own infer() function, and even once the last reference to a model goes, PyTorch keeps the freed blocks in its allocator rather than returning them to the driver. On a shared box that is the whole GPU held against one idle process. Stages now hand it back as soon as they are done with it. A stage that caches a model registers how to drop it with @on_release, so models.py needs to know nothing about the pipeline, and run_pipeline calls release() in a finally after each stage — not once per job, because transcribe sits between separate and align waiting on the Whisper server, which can be minutes of a GPU nothing is using. release() answers the bytes it handed back and the stage logs a line when that is not zero, which is the only part of this visible from outside. It will not initialise CUDA just to clean up after a run that never touched it. The cost is reloading each model once per job rather than once per process, seconds against stages that run for minutes. The CUDA context itself stays until the process exits; releasing that would mean a subprocess per stage.	2026-09-29 00:01:20 -07:00
branderson	22c7f6d03f	webui: clear the page for a new song, and let a rebuild redo a stage Two problems found in review. CREATE SONG emptied the form but left the previous song's review, player, output path and queue selection in place, which reads as though the blank form and that song belong together. It now clears the page down to a blank sheet, and a `composing` flag stops the next poll selecting the newest song back into it. The reason it cleared so little is that clearDetail reached for an element id that no longer exists, so it threw partway through and left everything after the first two lines untouched. It now targets the review box and takes the reason to show, since "no songs queued yet" is wrong when you have simply deselected. Separately, rebuild plans from the fields that changed, which cannot see a change in songtrace itself: rebuilding a song packaged before the video's audio was stripped found every stage cached and kept the echo. The rebuild form gains "restart from", the CLI's --force-from, which combines with the planner rather than replacing it — an explicit stage may widen what is rerun but never narrow it below what the edit requires. Also: a browser submits a textarea with CRLF, so every rebuild from the page looked like a lyrics edit and redid the alignment. Lyrics are normalised to LF on the way in. And [hidden] is only display:none, which .field's display:flex was winning, so the new control showed in new-song mode despite the property being set.	2026-09-25 23:52:28 -07:00
branderson	7c4359eb39	webui: play a finished song against its review Checking whether the notes line up meant packaging the song, finding it in the game and starting it there. The review panel now plays the song folder itself: a transport above the lyrics, the line being sung highlighted, words lighting up as they are reached, and a playhead sweeping each line against its notes. Clicking a line or a word plays from there. Playback is modelled on the game rather than on whatever is convenient: #MP3 is the audio and the clock, #VIDEO is muted background slaved to it, and the times are the same seconds the notes were generated from. A word that lights up late here is late in the game. package now strips the audio track from the video it writes. It is the same recording as the #MP3, so a player that does not mute the video plays both a few frames apart, which is heard as an echo. The compatible-video branch remuxes with -c copy -an rather than copying the file, so it still does not re-encode. The browser player mutes the video as well, which covers song folders built before this. A new /api/jobs/{id}/media/{audio,video,cover} serves what the song's own #MP3/#VIDEO/#COVER headers name, refusing anything resolving outside the song folder. Range requests come from FileResponse, so seeking works. Also, per review: CREATE SONG moves out of the form and into the top of the queue, where it is always available, and a successful submission keeps the fields it sent instead of emptying them.	2026-09-25 23:33:23 -07:00
branderson	ba8bf9f4af	Merge pull request 'webui: bring a selected song's inputs back into the form' (#9 ) from webui-prefill into main Reviewed-on: #9 Reviewed-by: branderson <brad@branderson.io>	2026-09-25 23:09:30 -07:00
branderson	be67e01943	webui: bring a selected song's inputs back into the form Selecting a queue entry filled the review panel but left the form empty, so the only way to adjust a song was the EDIT & REBUILD button, and the only way to see what a song was built from was to click it. Selection now loads the job's inputs, which is what the form is for: pick a song, change a line of lyrics, hit REBUILD SONG. That makes the EDIT & REBUILD action redundant, so it goes. CANCEL becomes NEW SONG INSTEAD, which is what it does now that the form is rarely empty. The landing page deliberately keeps its blank form: the job auto-selected on load fills the review panel only, so opening the page is still the way to start a new song. Three things the old button-driven flow never had to handle: - Advanced is opened when the job has a genre, year, language, bpm or a kept cover. Those carry into a rebuild either way, so leaving them behind a closed summary meant rebuilding with values that were never on screen. - Submit is disabled, with the reason shown, while the selected job is queued or running, since /rebuild answers 409 until it settles. - Selecting a different entry asks before discarding edits, and deleting the job the form is holding drops it rather than leaving the form pointed at an id that would 404.	2026-09-25 22:48:43 -07:00
branderson	b5850042be	Merge pull request 'Fix YouTube 403s: install yt-dlp's YouTube extras and retry across player clients' (#8 ) from ytdlp-resilience into main Reviewed-on: #8 Reviewed-by: branderson <brad@branderson.io>	2026-09-25 18:32:19 -07:00
branderson	55e8297672	acquire: install yt-dlp's YouTube extras and retry across player clients yt-dlp declares `dependencies = []` and puts everything in extras, so the bare `yt-dlp` requirement installed it with no yt-dlp-ejs and no JavaScript runtime. Without those it cannot solve YouTube's player challenges and falls back to media URLs that YouTube rejects with HTTP 403, which is the failure seen when adding a song. Require `yt-dlp[default,curl-cffi,deno]` instead; the deno extra ships the JS runtime as a wheel, so the image needs no extra apt packages. Around that, make the stage survive the next time YouTube changes something: the download is retried against each player client in `ytdlp_clients` (default `["default", "tv"]`, the latter not needing a PO token) before giving up, yt-dlp's retry counts are raised, and its own log lines are routed into the job log so a PO-token demand can be told apart from a rate limit. The final error names the running yt-dlp version, since a stale one is the usual cause. Cookies, TLS impersonation and a bgutil PO token provider URL are now config fields, so an operator can work around a block without a code change. `_classify_downloads` skips .part/.ytdl scratch files, which retries make far more likely to be lying around.	2026-09-25 18:24:40 -07:00
branderson	a7622c2d01	Merge pull request 'Add rebuild-from-scratch, delete, and cross-restart persistence for songs' (#5 ) from song-rebuild-persistence into main Reviewed-on: #5	2026-07-28 03:25:48 -07:00
branderson	c489dd5742	docs: regenerate webui.png	2026-07-28 03:26:08 -07:00
branderson	f1b967b487	Merge remote-tracking branch 'origin/main' into song-rebuild-persistence # Conflicts: # src/songtrace/web.py # tests/test_web.py	2026-07-28 03:24:04 -07:00
branderson	e141cc5773	Fix DELETE slug-collision race and widen _discover_jobs error handling - DELETE now checks under the shared lock whether another job with the same slug is queued/running before removing the job entry, so deleting a done parent can no longer rmtree a rebuild's active work/output directories. - _discover_jobs now wraps the iterdir/stat/read_warnings calls in the same try/except as the JSON parse, so one corrupt or torn sidecar file (e.g. from a killed mid-write) no longer crashes the whole app at boot. - Fixed test_delete_tolerates_a_work_dir_that_is_already_gone, which had become vacuous since _write_sidecar now eagerly creates the work dir; it now removes the dir before asserting DELETE tolerates its absence. - Added a regression test covering the DELETE slug-collision race.	2026-07-28 03:06:09 -07:00
branderson	92948b2ea7	webui: add rebuild-from-scratch and delete actions	2026-07-28 02:52:43 -07:00
branderson	c8c3d941ca	web: persist song input as a sidecar and reload jobs on startup	2026-07-28 02:38:55 -07:00
branderson	7b8fce5b6e	web: add delete endpoint	2026-07-28 02:34:09 -07:00
branderson	cb10400a21	web: add force-rebuild-from-scratch endpoint	2026-07-28 02:31:23 -07:00
branderson	98e4f9532f	pipeline: extract read_warnings for reuse by startup discovery	2026-07-28 02:28:27 -07:00
branderson	77bebcf9e1	Merge pull request 'Webui: open video source, restructure lyrics form, and add a visual per-line review' (#4 ) from worktree-webui-lyrics-review into main Reviewed-on: #4	2026-07-28 01:42:56 -07:00
branderson	986f26b8ed	Fix note bars hidden under alternating word-lane tint .review-lane spans the full timeline box height (top:0; bottom:0) to group each word with its note visually, but as a positioned element it painted after the unpositioned .review-roll SVG, so the lane's background covered the note bars for every other word. Give .review-roll position: relative so DOM order (lanes, then notes) governs paint order instead.	2026-07-28 01:43:04 -07:00
branderson	d557ac6386	Use real Greensleeves melody in webui screenshot fixture Replace arbitrary demo pitches with the actual melody transcribed from Gilbert DeBenedetti's public-domain "Level 4, Flowing" arrangement (E minor, 6/8), converted to UltraStar semitones, with one note per word across both demo lines.	2026-07-28 01:35:44 -07:00
branderson	5023ea69d3	review: lay out lyrics words along a shared time axis with notes aligned beneath The review panel used to show a full line of packed text with one undifferentiated strip of notes below it, so it was impossible to tell which notes belonged to which word. The forced-alignment stage already produces per-word timestamps in alignment.json, but they were never surfaced past the pipeline. /api/jobs/{id}/review now includes each line's per-word timing, and the frontend spaces words out by their real sung timing (karaoke-style) instead of normal text flow, with the note strip sharing that same time axis so each note lands directly under its word. Alternating lane backgrounds make the word/note pairing unambiguous. Dropped lines and jobs without alignment data keep the old plain-text rendering.	2026-07-28 01:22:48 -07:00
branderson	0599f6b665	docs: regenerate webui.png with the new lyrics review panel	2026-07-28 00:48:42 -07:00
branderson	2debfb171c	transcribe: scope missing_spans to the unmatched audio itself, not the whole vocal section Also fix a secondary ripple in build_notes(): a skipped row (inside a missing span) no longer advances prev_line/prev_end, so it can't corrupt line-initial spacing or break timing for its own line's later words.	2026-07-28 00:48:19 -07:00
branderson	ac904cb09f	screenshot: select the new Greensleeves fixture so it actually appears in the regenerated screenshot	2026-07-28 00:48:19 -07:00
branderson	8e89b08571	screenshot fixture: add songs with dropped lines, missing span, low-confidence notes	2026-07-28 00:48:08 -07:00
branderson	6a98ea0f21	fix: clear lyrics-review cache on job status change renderLyricsReview was caching its job ID only in the done path, allowing stale data to persist when switching between done/running/failed jobs. Now clears the cache in the failed and not-done branches so a previously- viewed done job always re-fetches when re-selected. Regression test: test_lyrics_review_cache_clears_on_job_status_change verifies both branches clear the cache key before rendering placeholders.	2026-07-27 23:36:56 -07:00
branderson	903e83dbcc	webui: replace flat warnings list with per-line SVG piano-roll review	2026-07-27 23:24:39 -07:00
branderson	a4e0b45953	web: add GET /api/jobs/{id}/review for per-line note data	2026-07-27 23:18:26 -07:00
branderson	ffd662fb38	notes: skip notes in unmatched spans, warn on dropped lines and missing lyrics	2026-07-27 23:13:39 -07:00
branderson	43f0554796	align: remove unused RefLyrics import	2026-07-27 23:11:30 -07:00
branderson	ebb7fc4433	align: exclude never-sung lines from proportional distribution, record align_meta.json - Added classify_lyrics_match import to wire lyric classification into align stage - Changed _assign_lines_to_sections signature to accept line_indices list instead of ref - Modified _run to classify lyrics match, filter dropped lines, and write align_meta.json - Sort segments by start time before processing for correct matching of non-chronological transcripts - Added tests for dropped line detection and missing span flagging	2026-07-27 23:06:02 -07:00
branderson	cf1c7410ab	transcribe: classify reference lines as matched/dropped, flag unmatched singing	2026-07-27 22:58:58 -07:00
branderson	5bc4855264	webui: move video source out of Advanced, give lyrics a tall column	2026-07-27 22:51:36 -07:00
branderson	a650c2f7ec	Move webui image to top	2026-07-27 19:59:46 -07:00
branderson	283da7e1a9	Fix yass link	2026-07-27 19:54:36 -07:00
branderson	292fa25d5e	Fix Melody Mania link	2026-07-27 19:51:53 -07:00
branderson	0eaab6aeed	Remove karaoke maker reference	2026-07-27 19:49:30 -07:00
branderson	7a48942f1c	Merge pull request 'Rename project from karaoke-maker to songtrace' (#3 ) from songtrace-rename into main Reviewed-on: branderson/karaoke-maker#3	2026-07-27 18:42:57 -07:00
branderson	6c15c3d8f1	docs: fix migration guidance for old console script and orphaned container	2026-07-27 18:36:42 -07:00
branderson	d2b7986cb9	docs: rename to songtrace and add tagline	2026-07-27 18:26:50 -07:00
branderson	3f90f4857f	refactor: restyle web UI wordmark as SONG / TRACE	2026-07-27 18:24:21 -07:00
branderson	5553614fdf	refactor: rename config env vars and default config path to songtrace	2026-07-27 18:21:30 -07:00
branderson	8a87534c5d	refactor: rename CLI command and distribution to songtrace	2026-07-27 18:18:11 -07:00
branderson	322f8eb98b	refactor: move package from karaoke_maker to songtrace	2026-07-27 18:14:23 -07:00
branderson	a3078bb2fc	Convert to AGPLv3	2026-07-27 17:37:56 -07:00
branderson	bd1b4aeb6e	Add GPLv3 license and webui screenshot	2026-07-27 17:28:07 -07:00
branderson	f2a97f4f10	Fix torchcodec issue	2026-07-27 10:05:59 -07:00
branderson	fca9e874f3	Support whisper asr-webservice	2026-07-27 09:53:03 -07:00
branderson	bd5593c3ed	Merge pull request 'Rerun failed tracks and edit-and-rebuild existing ones' (#2 ) from rerun-rebuild into main Reviewed-on: branderson/karaoke-maker#2	2026-07-27 09:19:56 -07:00