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.

Dockerfile

@@ -0,0 +1,19 @@
+# torch pip wheels bundle their own CUDA runtime libs; only the host NVIDIA
+# driver is needed (exposed via the nvidia container runtime), so a plain
+# Python base is sufficient — no CUDA base image required.
+FROM python:3.11-slim
+
+RUN apt-get update && apt-get install -y --no-install-recommends ffmpeg \
+    && rm -rf /var/lib/apt/lists/*
+
+WORKDIR /app
+COPY pyproject.toml ./
+COPY src ./src
+RUN pip install --no-cache-dir .
+
+ENV SONGTRACE_CONFIG=/config/config.toml \
+    TORCH_HOME=/data/models \
+    XDG_CACHE_HOME=/data/models/cache
+
+EXPOSE 8000
+CMD ["songtrace", "serve"]

LICENSE

@@ -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/>.

README.md

@@ -0,0 +1,175 @@
+# 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. 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 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
+6. **notes** — turn word timings + pitch into UltraStar notes
+7. **grid** — quantise notes to a beat grid
+8. **package** — write the final song folder (`.txt` + `.mp3` + video + cover)
+
+The shipped `.mp3` is the full original mix (vocals included) — the
+separated stems are used for analysis only and are not included in the
+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](https://www.yass-along.com/) or
+[USDX](https://usdx.eu/) and touch those spots up before playing.
+
+## Configuration
+
+Config is TOML. Resolution order:
+
+1. `--config <path>` passed explicitly on the command line
+2. the `SONGTRACE_CONFIG` environment variable
+3. `~/.config/songtrace/config.toml`
+
+Copy the example and fill in your Whisper server:
+
+```sh
+cp config.example.toml config.toml
+$EDITOR config.toml   # set whisper_url at minimum
+```
+
+See `config.example.toml` for every available field (`whisper_url`,
+`whisper_api_key`, `whisper_model`, `whisper_api`, `output_dir`, `work_dir`,
+`creator`, `demucs_model`, `crepe_model`, `device`, `bind_host`, `port`,
+`ytdlp_clients`, `ytdlp_cookies`, `ytdlp_impersonate`,
+`ytdlp_pot_provider_url`).
+
+`whisper_api` selects the wire format: `"openai"` (default) speaks the
+OpenAI-compatible `POST {whisper_url}/audio/transcriptions` API; `"asr-webservice"`
+speaks [onerahmet/openai-whisper-asr-webservice](https://github.com/ahmetoner/whisper-asr-webservice)'s
+`POST {whisper_url}/asr` API instead (point `whisper_url` at that server's base
+URL with no `/v1` suffix — `whisper_model`/`whisper_api_key` are unused for
+this backend, since the model is fixed by how that server was started).
+
+## Docker
+
+```sh
+cp config.example.toml config.toml
+$EDITOR config.toml
+docker compose up -d
+```
+
+The web UI is then at `http://localhost:8000`. Downloaded model weights
+(Demucs, wav2vec2, ...) are cached under `./data` on the host so they survive
+container recreation.
+
+Run the CLI inside the container instead of the web UI:
+
+```sh
+docker compose run --rm songtrace songtrace make \
+  --title "Song Title" --artist "Artist Name" \
+  --lyrics /data/lyrics.txt --youtube https://youtu.be/...
+```
+
+GPU passthrough requires the [NVIDIA Container
+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.
+
+## Bare-metal
+
+```sh
+pip install -e .
+songtrace make --help
+```
+
+```
+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 `songtrace serve [--config CONFIG]`.
+
+## Re-running a stage
+
+Each stage caches its output under `work_dir`. If a later stage needs
+another pass (e.g. you want to retry alignment with different lyrics, or
+pitch tracking looks off), re-run with `--force-from <stage>` to redo that
+stage and everything after it, reusing the cached output of everything
+before it:
+
+```sh
+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`.
+
+## 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 and the
+   version baked into the image does not. Hotfix a running container, then
+   rebuild so it sticks:
+
+   ```sh
+   docker compose exec songtrace pip install -U 'yt-dlp[default,curl-cffi,deno]'
+   docker compose restart songtrace
+   docker compose build --pull songtrace   # permanent
+   ```
+
+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. 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`:
+
+```sh
+pip install -e '.[dev]'
+playwright install chromium
+python scripts/screenshot_webui.py
+```
+
+## License
+
+[AGPLv3](LICENSE)

config.example.toml

@@ -0,0 +1,20 @@
+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"
+# demucs_model = "htdemucs_ft"
+# crepe_model = "full"
+# 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"

docker-compose.yml

@@ -0,0 +1,15 @@
+services:
+  songtrace:
+    build: .
+    ports:
+      - "8000:8000"
+    volumes:
+      - ./config.toml:/config/config.toml:ro
+      - ./data:/data
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: 1
+              capabilities: [gpu]

pyproject.toml

@@ -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"]
@@ -34,5 +40,8 @@ build-backend = "setuptools.build_meta"
 [tool.setuptools.packages.find]
 where = ["src"]
 
+[tool.setuptools.package-data]
+songtrace = ["static/*", "static/fonts/*"]
+
 [tool.pytest.ini_options]
 testpaths = ["tests"]

scripts/screenshot_webui.py

@@ -0,0 +1,212 @@
+"""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 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"
+
+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) -> str:
+    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}"
+
+
+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) -> None:
+    with sync_playwright() as p:
+        browser = p.chromium.launch()
+        page = browser.new_page(viewport={"width": 1280, "height": 900})
+        page.goto(base_url, wait_until="networkidle")
+        page.get_by_text(SELECTED, exact=True).click()
+        page.wait_for_timeout(300)
+        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 = start_server(Path(tmp))
+        time.sleep(1)
+        seed_jobs(base_url)
+        capture(base_url)
+    print(f"wrote {OUT_IMAGE}")
+
+
+if __name__ == "__main__":
+    main()

src/karaoke_maker/cli.py

@@ -1,16 +0,0 @@
-import argparse
-from importlib.metadata import version
-
-def build_parser() -> argparse.ArgumentParser:
-    parser = argparse.ArgumentParser(prog="karaoke-maker")
-    parser.add_argument("--version", action="store_true")
-    parser.add_subparsers(dest="command")
-    return parser
-
-def main(argv: list[str] | None = None) -> int:
-    args = build_parser().parse_args(argv)
-    if args.version:
-        print(f"karaoke-maker {version('karaoke-maker')}")
-        return 0
-    build_parser().print_help()
-    return 0

src/songtrace/audio_io.py

@@ -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

src/songtrace/cli.py

@@ -0,0 +1,83 @@
+import argparse
+import math
+import sys
+from importlib.metadata import version
+from pathlib import Path
+
+from .config import load_config
+from .pipeline import STAGE_ORDER, PipelineError, SongInput, run_pipeline
+
+
+def _positive_float(value: str) -> float:
+    bpm = float(value)
+    if not math.isfinite(bpm) or bpm <= 0:
+        raise argparse.ArgumentTypeError(f"--bpm must be a finite value > 0, got {value!r}")
+    return bpm
+
+
+def build_parser() -> argparse.ArgumentParser:
+    parser = argparse.ArgumentParser(prog="songtrace")
+    parser.add_argument("--version", action="store_true")
+    sub = parser.add_subparsers(dest="command")
+
+    make = sub.add_parser("make", help="build a karaoke song folder")
+    make.add_argument("--title", required=True)
+    make.add_argument("--artist", required=True)
+    make.add_argument("--lyrics", required=True, type=Path)
+    make.add_argument("--youtube")
+    make.add_argument("--video", type=Path)
+    make.add_argument("--genre")
+    make.add_argument("--year", type=int)
+    make.add_argument("--language")
+    make.add_argument("--cover", type=Path)
+    make.add_argument("--bpm", type=_positive_float)
+    make.add_argument("--force-from", choices=STAGE_ORDER)
+    make.add_argument("--config", type=Path)
+
+    serve = sub.add_parser("serve", help="run the web UI")
+    serve.add_argument("--config", type=Path)
+    return parser
+
+
+def _cmd_make(args) -> int:
+    if bool(args.youtube) == bool(args.video):
+        print("error: provide exactly one of --youtube / --video", file=sys.stderr)
+        return 2
+    try:
+        lyrics_text = args.lyrics.read_text(encoding="utf-8")
+    except OSError as e:
+        print(f"error: could not read --lyrics file: {e}", file=sys.stderr)
+        return 2
+    except UnicodeDecodeError as e:
+        print(f"error: --lyrics file is not valid UTF-8 text: {e}", file=sys.stderr)
+        return 2
+    cfg = load_config(args.config)
+    inp = SongInput(
+        title=args.title, artist=args.artist,
+        lyrics_text=lyrics_text,
+        youtube_url=args.youtube, video_path=args.video,
+        genre=args.genre, year=args.year, language=args.language,
+        cover_path=args.cover, bpm_override=args.bpm)
+    try:
+        result = run_pipeline(inp, cfg, force_from=args.force_from)
+    except PipelineError as e:
+        print(f"error: {e}", file=sys.stderr)
+        return 1
+    print(f"done: {result.output_dir}")
+    for w in result.warnings:
+        print(f"  review: {w}")
+    return 0
+
+
+def main(argv: list[str] | None = None) -> int:
+    args = build_parser().parse_args(argv)
+    if args.version:
+        print(f"songtrace {version('songtrace')}")
+        return 0
+    if args.command == "make":
+        return _cmd_make(args)
+    if args.command == "serve":
+        from .web import serve
+        return serve(load_config(args.config))
+    build_parser().print_help()
+    return 0

src/songtrace/config.py

@@ -0,0 +1,43 @@
+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
+
+
+_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)

src/songtrace/grid.py

@@ -0,0 +1,58 @@
+import math
+from pathlib import Path
+
+import numpy as np
+
+from .ultrastar import LineBreak, Note, UltraStarSong, format_bpm
+
+
+def detect_bpm(instrumental_path: Path) -> float:
+    import librosa
+    y, sr = librosa.load(str(instrumental_path), sr=22050, mono=True)
+    tempo, _ = librosa.beat.beat_track(y=y, sr=sr)
+    return float(np.atleast_1d(tempo)[0])
+
+
+def scale_bpm(bpm: float) -> float:
+    """Fold a tempo into UltraStar's usual [200, 400) file-BPM range."""
+    if not math.isfinite(bpm) or bpm <= 0:
+        raise ValueError(f"cannot scale bpm {bpm}")
+    while bpm < 200:
+        bpm *= 2
+    while bpm >= 400:
+        bpm /= 2
+    return bpm
+
+
+def to_ultrastar(notes: list[dict], breaks: list[float], file_bpm: float,
+                 headers: list[tuple[str, str]]) -> UltraStarSong:
+    if not notes:
+        raise ValueError("no notes to place on the beat grid")
+    gap_ms = round(notes[0]["start"] * 1000)
+    spb = 60 / (4 * file_bpm)
+
+    def beat(t: float) -> int:
+        return round((t - gap_ms / 1000) / spb)
+
+    # a line break at the same instant as a note belongs before it
+    events = [(t, 0, None) for t in breaks] + [(n["start"], 1, n) for n in notes]
+    events.sort(key=lambda e: e[:2])
+    body: list[Note | LineBreak] = []
+    # notes.py can only keep notes apart in seconds; quantising re-collides them,
+    # so the cursor keeps the emitted beats strictly ordered and non-overlapping
+    cursor = -math.inf
+    for t, is_note, n in events:
+        if not is_note:
+            b = max(beat(t), cursor)
+            body.append(LineBreak(beat=b))
+            cursor = b
+        else:
+            b = max(beat(t), cursor)
+            length = max(1, beat(t + n["duration"]) - beat(t))
+            body.append(Note(kind=":", beat=b, length=length, pitch=n["pitch"],
+                             text=n["text"]))
+            cursor = b + length
+    out_headers = list(headers)
+    out_headers.append(("BPM", format_bpm(file_bpm)))
+    out_headers.append(("GAP", str(gap_ms)))
+    return UltraStarSong(headers=out_headers, body=body)

src/songtrace/languages.py

@@ -0,0 +1,19 @@
+from dataclasses import dataclass
+
+
+@dataclass(frozen=True)
+class Language:
+    iso: str             # whisper / num2words language code
+    pyphen: str | None   # pyphen dictionary code, None if no dictionary exists
+
+
+LANGUAGES: dict[str, Language] = {
+    "english": Language("en", "en_US"),
+    "german": Language("de", "de_DE"),
+    "spanish": Language("es", "es"),
+    "french": Language("fr", "fr"),
+    "italian": Language("it", "it"),
+    "portuguese": Language("pt", "pt"),
+    "dutch": Language("nl", "nl_NL"),
+    "japanese": Language("ja", None),
+}

src/songtrace/lyrics.py

@@ -0,0 +1,44 @@
+import re
+from dataclasses import dataclass
+from num2words import num2words
+
+from .languages import LANGUAGES
+
+SECTION_RE = re.compile(r"^\s*[\[(].*[\])]\s*$")
+
+
+@dataclass
+class RefLyrics:
+    lines: list[list[str]]
+    words: list[str]
+    word_line_index: list[int]
+
+
+def normalize_lyrics(text: str) -> RefLyrics:
+    lines = []
+    for raw in text.splitlines():
+        if not raw.strip() or SECTION_RE.match(raw):
+            continue
+        lines.append(raw.split())
+    words, idx = [], []
+    for i, line in enumerate(lines):
+        words.extend(line)
+        idx.extend([i] * len(line))
+    return RefLyrics(lines, words, idx)
+
+
+def alignment_token(word: str, language: str | None = None) -> str:
+    entry = LANGUAGES.get((language or "english").lower())
+    lang = entry.iso if entry else "en"
+
+    def expand(m):
+        try:
+            return " " + num2words(int(m[0]), lang=lang) + " "
+        except NotImplementedError:
+            return " " + num2words(int(m[0])) + " "
+
+    word = word.replace("’", "'").replace("ʼ", "'")
+    word = re.sub(r"\d+", expand, word)
+    word = word.lower()
+    word = re.sub(r"[^a-z'À-ɏЀ-ӿ ]", "", word)
+    return re.sub(r"\s+", "", word)

src/songtrace/media.py

@@ -0,0 +1,23 @@
+import json
+import subprocess
+from pathlib import Path
+
+from .pipeline import PipelineError
+
+
+def ffmpeg(*args: str) -> None:
+    result = subprocess.run(["ffmpeg", "-y", *args], capture_output=True, text=True)
+    if result.returncode != 0:
+        raise PipelineError(f"ffmpeg failed: {result.stderr[-2000:]}")
+
+
+def ffprobe_codec(path: Path) -> tuple[str, str]:
+    result = subprocess.run(
+        ["ffprobe", "-v", "quiet", "-print_format", "json",
+         "-show_format", "-show_streams", str(path)],
+        capture_output=True, text=True)
+    if result.returncode != 0:
+        raise PipelineError(f"ffprobe failed on {path}")
+    data = json.loads(result.stdout)
+    video = next((s for s in data["streams"] if s["codec_type"] == "video"), {})
+    return data["format"]["format_name"], video.get("codec_name", "")

src/songtrace/notes.py

@@ -0,0 +1,198 @@
+import math
+from dataclasses import dataclass
+
+import numpy as np
+import pyphen
+
+from .languages import LANGUAGES
+
+CONF_MASK = 0.5          # pitch frames below this are unvoiced
+LOW_SCORE = 0.5          # word alignment score below this -> low_confidence
+SPLIT_SEMITONES = 0.8
+MIN_SEG_SEC = 0.05
+MIN_NOTE_SEC = 0.05      # alignment may hand us zero-length words
+HOP_SEC = 0.01           # pitch stage hop
+
+_SYLL_SEP = "\0"         # sentinel so literal hyphens survive syllabification
+
+
+@dataclass
+class TimedNote:
+    start: float
+    duration: float
+    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:
+    return 69 + 12 * math.log2(hz / 440.0)
+
+
+def hz_to_ultrastar(hz: float) -> int:
+    return round(hz_to_midi_float(hz)) - 60
+
+
+def segment_span(times, f0, conf, start: float, end: float
+                 ) -> list[tuple[float, float, float]]:
+    """Voiced pitch segments within [start, end) as (start, end, median_midi)."""
+    mask = (times >= start) & (times < end) & (conf >= CONF_MASK) & (f0 > 0)
+    t = times[mask]
+    if len(t) == 0:
+        return []
+    midi = 69 + 12 * np.log2(f0[mask] / 440.0)
+    segs: list[tuple[float, float, float]] = []
+
+    def emit(a: int, b: int) -> None:
+        if float(t[b - 1]) - float(t[a]) + HOP_SEC >= MIN_SEG_SEC:
+            segs.append((float(t[a]), float(t[b - 1]) + HOP_SEC,
+                         float(np.median(midi[a:b]))))
+
+    seg_start, dev_start = 0, None
+    for i in range(1, len(t)):
+        stable_end = i if dev_start is None else dev_start
+        if abs(midi[i] - np.median(midi[seg_start:stable_end])) >= SPLIT_SEMITONES:
+            # only split once the deviation has held for MIN_SEG_SEC, so that
+            # short tracker glitches don't fragment a sustained note
+            if dev_start is None:
+                dev_start = i
+            elif t[i] - t[dev_start] + HOP_SEC >= MIN_SEG_SEC:
+                emit(seg_start, dev_start)
+                seg_start, dev_start = dev_start, None
+        else:
+            dev_start = None
+    emit(seg_start, len(t))
+    return segs
+
+
+def syllabify(word: str, language: str | None) -> list[str]:
+    entry = LANGUAGES.get((language or "english").lower())
+    code = entry.pyphen if entry else "en_US"
+    if code is None:
+        return [word]
+    try:
+        dic = pyphen.Pyphen(lang=code)
+    except KeyError:
+        return [word]
+    parts = [p for p in dic.inserted(word, hyphen=_SYLL_SEP).split(_SYLL_SEP) if p]
+    return parts or [word]
+
+
+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] = []
+    prev_line, prev_end, last_seen_line = None, 0.0, None
+    for row in rows:
+        # last_seen_line tracks every row (skipped or not) so a run of
+        # skipped rows can't trigger a duplicate break on the next row of
+        # the same still-unseen line; prev_line/prev_end only advance on
+        # rows that actually emit notes, so a skipped row can't corrupt
+        # line_initial or break timing for its own line's later words.
+        if last_seen_line is not None and row["line_index"] > last_seen_line:
+            breaks.append((prev_end + row["start"]) / 2)
+        last_seen_line = row["line_index"]
+        line_initial = row["line_index"] != prev_line
+        if any(a <= row["start"] < b for a, b in missing_spans):
+            continue
+        prev_line, prev_end = row["line_index"], row["end"]
+        segs = segment_span(times, f0, conf, row["start"], row["end"])
+        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,
+                               row["line_index"]))
+            continue
+        low = row["score"] < LOW_SCORE
+        sylls = syllabify(row["word"], language)
+        if len(segs) < len(sylls):
+            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,
+                               row["line_index"]))
+    _deoverlap(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, line_index: int) -> TimedNote:
+    return TimedNote(float(start), float(duration), int(pitch), text, low,
+                     int(line_index))
+
+
+def _deoverlap(notes: list[TimedNote]) -> None:
+    """Give every note a positive length without letting notes collide.
+
+    Notes are grown to MIN_NOTE_SEC (alignment hands us zero-length words, and
+    subdividing a short word by syllable makes very short slices) but only into
+    space the next note isn't using; where there is no space at all the rest of
+    the song is nudged later.
+    """
+    limits = [math.inf] * len(notes)
+    for i in range(len(notes) - 2, -1, -1):
+        limits[i] = (notes[i + 1].start if notes[i + 1].start > notes[i].start
+                     else limits[i + 1])
+    cursor = -math.inf
+    for note, limit in zip(notes, limits):
+        note.start = max(note.start, cursor)
+        room = limit - note.start
+        note.duration = (min(max(note.duration, MIN_NOTE_SEC), room)
+                         if room > 0 else MIN_NOTE_SEC)
+        cursor = note.start + note.duration
+
+
+def _text(syllable: str, index: int, line_initial: bool) -> str:
+    return syllable if index or line_initial else " " + syllable
+
+
+def _subdivide(segs, target: int):
+    """Cut segments into equal slices until there is one per syllable."""
+    parts = [1] * len(segs)
+    while sum(parts) < target:
+        i = max(range(len(segs)),
+                key=lambda j: (segs[j][1] - segs[j][0]) / parts[j])
+        parts[i] += 1
+    out = []
+    for (s, e, midi), n in zip(segs, parts):
+        step = (e - s) / n
+        out.extend((s + k * step, s + (k + 1) * step, midi) for k in range(n))
+    return out
+
+
+def _fmt(t: float) -> str:
+    return f"{int(t // 60)}:{int(t % 60):02d}"
+
+
+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:
+            run.append(n)
+        elif run:
+            first, last = run[0], run[-1]
+            warnings.append(
+                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

src/songtrace/pipeline.py

@@ -0,0 +1,79 @@
+import json
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Callable
+
+from .config import Config
+
+class PipelineError(Exception):
+    pass
+
+@dataclass
+class SongInput:
+    title: str
+    artist: str
+    lyrics_text: str
+    youtube_url: str | None = None
+    video_path: Path | None = None
+    genre: str | None = None
+    year: int | None = None
+    language: str | None = None
+    cover_path: Path | None = None      # local file or URL already fetched by front-end
+    bpm_override: float | None = None
+
+@dataclass
+class Stage:
+    name: str
+    outputs: Callable[["Ctx"], list[Path]]
+    run: Callable[["Ctx"], None]
+
+@dataclass
+class Ctx:
+    inp: SongInput
+    cfg: Config
+    work: Path                          # per-song work dir (created)
+    log: Callable[[str], None]
+
+@dataclass
+class PipelineResult:
+    output_dir: Path
+    warnings: list[str] = field(default_factory=list)
+
+STAGE_ORDER = ["acquire", "separate", "transcribe", "align", "pitch", "notes", "grid", "package"]
+
+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()
+
+def run_pipeline(inp: SongInput, cfg: Config, log=print,
+                 force_from: str | None = None,
+                 stages: list[Stage] | None = None) -> PipelineResult:
+    if stages is None:
+        stages = _real_stages()
+    names = [s.name for s in stages]
+    if force_from is not None and force_from not in names:
+        raise PipelineError(f"unknown stage {force_from!r}, expected one of {names}")
+    work = cfg.work_dir / song_slug(inp)
+    work.mkdir(parents=True, exist_ok=True)
+    ctx = Ctx(inp=inp, cfg=cfg, work=work, log=log)
+    forcing = False
+    for stage in stages:
+        if stage.name == force_from:
+            forcing = True
+        if not forcing and all(p.exists() for p in stage.outputs(ctx)):
+            log(f"[{stage.name}] cached, skipping")
+            continue
+        log(f"[{stage.name}] running")
+        stage.run(ctx)
+    return PipelineResult(output_dir=cfg.output_dir / song_slug(inp),
+                          warnings=read_warnings(work))

src/songtrace/rebuild.py

@@ -0,0 +1,73 @@
+import shutil
+from pathlib import Path
+
+from .config import Config
+from .pipeline import STAGE_ORDER, PipelineError, SongInput, song_slug
+
+# The earliest stage that reads each SongInput field. Editing a field means every
+# stage from here onward is stale. Adding a field to SongInput without adding it
+# here would let an edit invalidate nothing at all, so tests pin this to the
+# dataclass.
+FIELD_STAGE = {
+    "youtube_url": "acquire",
+    "video_path": "acquire",
+    "language": "transcribe",
+    "lyrics_text": "align",
+    "title": "grid",
+    "artist": "grid",
+    "genre": "grid",
+    "year": "grid",
+    "bpm_override": "grid",
+    "cover_path": "package",
+}
+
+
+def changed_fields(old: SongInput, new: SongInput) -> list[str]:
+    return [f for f in FIELD_STAGE if getattr(old, f) != getattr(new, f)]
+
+
+def plan(old: SongInput, new: SongInput) -> str | None:
+    """The stage to restart from, or None when nothing changed.
+
+    None is not "do nothing": run_pipeline with no force_from resumes at the
+    first stage whose outputs are missing, which is exactly what a rerun wants.
+    """
+    changed = changed_fields(old, new)
+    if not changed:
+        return None
+    return min((FIELD_STAGE[f] for f in changed), key=STAGE_ORDER.index)
+
+
+def relocate_work(old: SongInput, new: SongInput, cfg: Config) -> bool:
+    """Move the cached work dir when title/artist changed its slug.
+
+    Nothing inside a work dir depends on the slug except song.txt, which `grid`
+    rewrites — and FIELD_STAGE sends title/artist edits to `grid`. So a retitle
+    costs a rename plus grid and package, not the whole GPU pipeline.
+    """
+    old_dir = cfg.work_dir / song_slug(old)
+    new_dir = cfg.work_dir / song_slug(new)
+    if old_dir == new_dir or not old_dir.exists():
+        return False
+    if new_dir.exists():
+        raise PipelineError(
+            f"a work folder already exists for {song_slug(new)!r}; "
+            "rename or remove it before rebuilding")
+    new_dir.parent.mkdir(parents=True, exist_ok=True)
+    shutil.move(str(old_dir), str(new_dir))
+    return True
+
+
+def superseded_output(old_dir: str | None, new_dir: str,
+                      output_root: Path) -> Path | None:
+    """The previous build's folder, when the rebuild replaced it under a new slug.
+
+    Returns None unless the path is strictly inside output_root, so a bad job
+    record cannot turn the caller's removal into an arbitrary delete.
+    """
+    if not old_dir:
+        return None
+    old, new = Path(old_dir).resolve(), Path(new_dir).resolve()
+    if old == new or output_root.resolve() not in old.parents:
+        return None
+    return old

src/songtrace/stages/__init__.py

@@ -0,0 +1,14 @@
+from ..pipeline import Stage
+from .acquire import acquire_stage
+from .align import align_stage
+from .grid import grid_stage
+from .notes import notes_stage
+from .package import package_stage
+from .pitch import pitch_stage
+from .separate import separate_stage
+from .transcribe import transcribe_stage
+
+
+def all_stages() -> list[Stage]:
+    return [acquire_stage(), separate_stage(), transcribe_stage(), align_stage(),
+            pitch_stage(), notes_stage(), grid_stage(), package_stage()]

src/songtrace/stages/acquire.py

@@ -0,0 +1,109 @@
+import json
+import shutil
+from pathlib import Path
+
+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
+
+
+class _YdlLog:
+    """Routes yt-dlp's own messages into the job log, minus progress noise."""
+
+    def __init__(self, log):
+        self._log = log
+
+    def debug(self, msg: str) -> None:
+        if not msg.startswith("[debug] "):
+            self._log(msg)
+
+    def warning(self, msg: str) -> None:
+        self._log(msg)
+
+    error = warning
+
+
+def _ydl_opts(cfg: Config, dest: Path, client: str, log) -> dict:
+    extractor_args: dict[str, dict[str, list[str]]] = {}
+    if client != "default":
+        extractor_args["youtube"] = {"player_client": [client]}
+    if cfg.ytdlp_pot_provider_url:
+        extractor_args["youtubepot-bgutilhttp"] = {
+            "base_url": [cfg.ytdlp_pot_provider_url]}
+    opts = {
+        "outtmpl": str(dest / "source_video.%(ext)s"),
+        "format": "bestvideo*+bestaudio/best",
+        "writethumbnail": True,
+        "noplaylist": True,
+        "noprogress": True,
+        "logger": _YdlLog(log),
+        "retries": 10,
+        "fragment_retries": 10,
+        "extractor_retries": 5,
+        "extractor_args": extractor_args,
+    }
+    if cfg.ytdlp_cookies:
+        opts["cookiefile"] = str(cfg.ytdlp_cookies)
+    if cfg.ytdlp_impersonate:
+        from yt_dlp.networking.impersonate import ImpersonateTarget
+        opts["impersonate"] = ImpersonateTarget.from_str(cfg.ytdlp_impersonate)
+    return opts
+
+
+def download_youtube(url: str, dest: Path, cfg: Config,
+                     log=print) -> tuple[Path, Path | None]:
+    import yt_dlp
+    clients = list(cfg.ytdlp_clients) or ["default"]
+    last: Exception | None = None
+    for client in clients:
+        try:
+            with yt_dlp.YoutubeDL(_ydl_opts(cfg, dest, client, log)) as ydl:
+                ydl.extract_info(url, download=True)
+            return _classify_downloads(dest)
+        except Exception as e:
+            last = e
+            log(f"yt-dlp player_client={client} failed: {e}")
+    raise PipelineError(
+        f"yt-dlp {yt_dlp.version.__version__} could not download {url} after "
+        f"trying player_client={', '.join(clients)}. Last error: {last}. "
+        "YouTube blocks change often: upgrade yt-dlp, then 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,
+    )

src/songtrace/stages/align.py

@@ -0,0 +1,129 @@
+import json
+from pathlib import Path
+
+from ..lyrics import alignment_token, normalize_lyrics
+from ..pipeline import Ctx, PipelineError, Stage
+from .transcribe import classify_lyrics_match, load_transcript, lyrics_similarity, vocal_sections
+
+SIMILARITY_THRESHOLD = 0.4
+
+_model_cache = {}
+
+
+def _get_model(device: str):
+    if device not in _model_cache:
+        from torchaudio.pipelines import MMS_FA as bundle
+        _model_cache[device] = bundle.get_model().to(device)
+    return _model_cache[device]
+
+
+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
+
+    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(
+        waveform[:, lo:hi], sr, bundle.sample_rate)
+    # MMS_FA only models a-z and apostrophe; anything else becomes the star token.
+    vocab = bundle.get_dict()
+    tokens = ["".join(c if c in vocab else "*" for c in t) for t in tokens]
+    with torch.inference_mode():
+        emission, _ = _get_model(device)(waveform.to(device))
+        spans = bundle.get_aligner()(emission[0], bundle.get_tokenizer()(tokens))
+    frame = (hi - lo) / sr / emission.shape[1]
+    return [(start + s[0].start * frame,
+             start + s[-1].end * frame,
+             sum(t.score for t in s) / len(s)) for s in spans]
+
+
+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 = len(line_indices)
+    groups, elapsed, start = [], 0.0, 0
+    for i, (a, b) in enumerate(sections):
+        elapsed += b - a
+        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
+
+
+def _interpolate(rows: list[dict], lo_bound: float, hi_bound: float) -> None:
+    """Give rows left without times evenly spaced spans between their aligned neighbours."""
+    i = 0
+    while i < len(rows):
+        if rows[i]["start"] is not None:
+            i += 1
+            continue
+        j = i
+        while j < len(rows) and rows[j]["start"] is None:
+            j += 1
+        lo = rows[i - 1]["end"] if i else lo_bound
+        hi = max(rows[j]["start"] if j < len(rows) else hi_bound, lo)
+        step = (hi - lo) / (j - i)
+        for k in range(i, j):
+            rows[k]["start"] = lo + (k - i) * step
+            rows[k]["end"] = rows[k]["start"] + step
+        i = j
+
+
+def _enforce_monotonic(rows: list[dict]) -> None:
+    prev = 0.0
+    for r in rows:
+        r["start"] = max(r["start"], prev)
+        r["end"] = prev = max(r["end"], r["start"])
+
+
+def _align_section(vocals: Path, sec: tuple[float, float], words: list[tuple[str, int]],
+                   language: str | None, device: str) -> list[dict]:
+    tokens = [alignment_token(w, language) for w, _ in words]
+    alignable = [t for t in tokens if t]
+    aligned = iter(force_align_section(vocals, sec[0], sec[1], alignable, device)
+                   if alignable else [])
+    rows, blank = [], (None, None, 0.0)
+    for (word, line_index), token in zip(words, tokens):
+        start, end, score = next(aligned, blank) if token else blank
+        rows.append({"word": word, "start": start, "end": end,
+                     "score": score, "line_index": line_index})
+    _interpolate(rows, sec[0], sec[1])
+    return rows
+
+
+def _run(ctx: Ctx) -> None:
+    ref = normalize_lyrics(ctx.inp.lyrics_text)
+    segments = sorted(load_transcript(ctx), key=lambda s: s.start)
+    sim = lyrics_similarity(segments, ref)
+    if sim < SIMILARITY_THRESHOLD:
+        raise PipelineError(
+            f"lyrics don't match audio (similarity {sim:.2f}) — "
+            "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(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:
+    return Stage(name="align",
+                 outputs=lambda ctx: [ctx.work / "alignment.json"],
+                 run=_run)

src/songtrace/stages/grid.py

@@ -0,0 +1,40 @@
+import json
+from pathlib import Path
+
+from .. import grid
+from ..pipeline import Ctx, PipelineError, Stage, song_slug
+from ..ultrastar import serialize_ultrastar
+
+
+def _run(ctx: Ctx) -> None:
+    data = json.loads((ctx.work / "notes_sec.json").read_text())
+    meta = json.loads((ctx.work / "meta.json").read_text())
+    if ctx.inp.bpm_override is not None:
+        bpm = ctx.inp.bpm_override
+    else:
+        try:
+            bpm = grid.scale_bpm(grid.detect_bpm(ctx.work / "instrumental.wav"))
+        except ValueError:
+            raise PipelineError("could not detect a usable tempo - is the audio silent?")
+    slug = song_slug(ctx.inp)
+    video_ext = Path(meta["video_file"]).suffix
+    headers = [("TITLE", ctx.inp.title), ("ARTIST", ctx.inp.artist)]
+    if ctx.inp.language:
+        headers.append(("LANGUAGE", ctx.inp.language))
+    if ctx.inp.genre:
+        headers.append(("GENRE", ctx.inp.genre))
+    if ctx.inp.year:
+        headers.append(("YEAR", str(ctx.inp.year)))
+    headers += [("CREATOR", ctx.cfg.creator), ("MP3", f"{slug}.mp3"),
+                ("COVER", f"{slug}.jpg"), ("VIDEO", f"{slug}{video_ext}")]
+    try:
+        song = grid.to_ultrastar(data["notes"], data["breaks"], bpm, headers)
+    except ValueError:
+        raise PipelineError("no notes were generated - the lyrics may not match the audio")
+    (ctx.work / "song.txt").write_text(serialize_ultrastar(song), encoding="utf-8")
+
+
+def grid_stage() -> Stage:
+    return Stage(name="grid",
+                 outputs=lambda ctx: [ctx.work / "song.txt"],
+                 run=_run)

src/songtrace/stages/notes.py

@@ -0,0 +1,33 @@
+import json
+from dataclasses import asdict
+
+import numpy as np
+
+from ..lyrics import normalize_lyrics
+from ..notes import build_notes
+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,
+        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,
+        "warnings": warnings,
+    }))
+
+
+def notes_stage() -> Stage:
+    return Stage(name="notes",
+                 outputs=lambda ctx: [ctx.work / "notes_sec.json"],
+                 run=_run)

src/songtrace/stages/package.py

@@ -0,0 +1,80 @@
+import json
+import shutil
+from pathlib import Path
+
+from PIL import Image, UnidentifiedImageError
+
+from ..media import ffmpeg, ffprobe_codec
+from ..pipeline import Ctx, Stage, song_slug
+from ..ultrastar import parse_ultrastar, serialize_ultrastar
+
+GOOD_VIDEO = {("webm", "vp9"), ("webm", "vp8"),
+              ("mov,mp4,m4a,3gp,3g2,mj2", "h264")}
+
+
+def prepare_cover(src: Path, dest_jpg: Path) -> None:
+    img = Image.open(src).convert("RGB")
+    side = min(img.size)
+    left = (img.width - side) // 2
+    top = (img.height - side) // 2
+    img = img.crop((left, top, left + side, top + side))
+    if side > 800:
+        img = img.resize((800, 800))
+    img.save(dest_jpg, "JPEG", quality=90)
+
+
+def _set_header(txt_path: Path, key: str, value: str | None) -> None:
+    song = parse_ultrastar(txt_path.read_text(encoding="utf-8"))
+    replaced = False
+    headers = []
+    for k, v in song.headers:
+        if k.upper() == key.upper():
+            replaced = True
+            if value is not None:
+                headers.append((key, value))
+        else:
+            headers.append((k, v))
+    if value is not None and not replaced:
+        headers.append((key, value))
+    song.headers = headers
+    txt_path.write_text(serialize_ultrastar(song), encoding="utf-8")
+
+
+def _run(ctx: Ctx) -> None:
+    slug = song_slug(ctx.inp)
+    out = ctx.cfg.output_dir / slug
+    out.mkdir(parents=True, exist_ok=True)
+    txt = out / f"{slug}.txt"
+    shutil.copyfile(ctx.work / "song.txt", txt)
+    ffmpeg("-i", str(ctx.work / "audio.wav"), "-b:a", "320k",
+           str(out / f"{slug}.mp3"))
+    meta = json.loads((ctx.work / "meta.json").read_text())
+    src_video = ctx.work / meta["video_file"]
+    fmt, codec = ffprobe_codec(src_video)
+    if (fmt, codec) in GOOD_VIDEO:
+        shutil.copyfile(src_video, 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"))
+        _set_header(txt, "VIDEO", f"{slug}.mp4")
+    cover_src = ctx.inp.cover_path
+    if cover_src is None and meta.get("thumbnail"):
+        cover_src = ctx.work / meta["thumbnail"]
+    cover_ok = False
+    if cover_src and Path(cover_src).exists():
+        try:
+            prepare_cover(Path(cover_src), out / f"{slug}.jpg")
+            cover_ok = True
+        except UnidentifiedImageError:
+            ctx.log("cover file is not a valid image; dropping #COVER header")
+    else:
+        ctx.log("no cover available; dropping #COVER header")
+    if not cover_ok:
+        _set_header(txt, "COVER", None)
+    (ctx.work / "package.done").write_text("")
+
+
+def package_stage() -> Stage:
+    def outputs(ctx: Ctx):
+        return [ctx.work / "package.done"]
+    return Stage(name="package", outputs=outputs, run=_run)

src/songtrace/stages/pitch.py

@@ -0,0 +1,36 @@
+from pathlib import Path
+
+import numpy as np
+
+from ..pipeline import Ctx, Stage
+
+
+def track_pitch(vocals_path: Path, model: str, device: str):
+    import torchaudio
+    import torchcrepe
+
+    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(
+        waveform, 16000, hop_length=160, model=model,
+        fmin=50.0, fmax=1100.0, device=device,
+        return_periodicity=True, batch_size=512)
+    f0 = f0[0].cpu().numpy()
+    conf = periodicity[0].cpu().numpy()
+    times = np.arange(len(f0)) * 0.01
+    return times, f0, conf
+
+
+def _run(ctx: Ctx) -> None:
+    times, f0, conf = track_pitch(ctx.work / "vocals.wav",
+                                  ctx.cfg.crepe_model, ctx.cfg.device)
+    np.savez(ctx.work / "pitch.npz", times=times, f0=f0, conf=conf)
+
+
+def pitch_stage() -> Stage:
+    return Stage(name="pitch",
+                 outputs=lambda ctx: [ctx.work / "pitch.npz"],
+                 run=_run)

src/songtrace/stages/separate.py

@@ -0,0 +1,30 @@
+from pathlib import Path
+
+import soundfile as sf
+
+from ..pipeline import Ctx, Stage
+
+
+def separate_audio(audio_path: Path, model: str, device: str):
+    from demucs.api import Separator
+    separator = Separator(model=model, device=device)
+    _, stems = separator.separate_audio_file(audio_path)
+    return stems
+
+
+def _run(ctx: Ctx) -> None:
+    stems = separate_audio(ctx.work / "audio.wav",
+                           ctx.cfg.demucs_model, ctx.cfg.device)
+    vocals = stems["vocals"]
+    others = [t for name, t in stems.items() if name != "vocals"]
+    instrumental = sum(others[1:], start=others[0])
+    sf.write(ctx.work / "vocals.wav", vocals.cpu().numpy().T, 44100)
+    sf.write(ctx.work / "instrumental.wav", instrumental.cpu().numpy().T, 44100)
+
+
+def separate_stage() -> Stage:
+    return Stage(
+        name="separate",
+        outputs=lambda ctx: [ctx.work / "vocals.wav", ctx.work / "instrumental.wav"],
+        run=_run,
+    )

src/songtrace/stages/transcribe.py

@@ -0,0 +1,100 @@
+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) — scoped tightly to the unmatched audio itself, not
+    the whole vocal section it falls in, so matched lines sharing a section
+    with an ad-lib keep their notes.
+    """
+    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]

src/songtrace/static/fonts/OFL.txt

@@ -0,0 +1,93 @@
+Copyright 2018 The Chakra Petch Project Authors (https://github.com/m4rc1e/Chakra-Petch.git) ChakraPetch-LightItalic.ttf: Copyright 2018 The Chakra Petch Project Authors (https://github.com/m4rc1e/Chakra-Petch.git) ChakraPetch-Regular.ttf: Copyright 2018 The Chakra Petch Project Authors (https://github.com/m4rc1e/Chakra-Petch.git) ChakraPetch-Italic.ttf: Copyright 2018 The Chakra Petch Project Authors (https://github.com/m4rc1e/Chakra-Petch.git) ChakraPetch-Medium.ttf: Copyright 2018 The Chakra Petch Project Authors (https://github.com/m4rc1e/Chakra-Petch.git) ChakraPetch-MediumItalic.ttf: Copyright 2018 The Chakra Petch Project Authors (https://github.com/m4rc1e/Chakra-Petch.git) ChakraPetch-SemiBold.ttf: Copyright 2018 The Chakra Petch Project Authors (https://github.com/m4rc1e/Chakra-Petch.git) ChakraPetch-SemiBoldItalic.ttf: Copyright 2018 The Chakra Petch Project Authors (https://github.com/m4rc1e/Chakra-Petch.git) ChakraPetch-Bold.ttf: Copyright 2018 The Chakra Petch Project Authors (https://github.com/m4rc1e/Chakra-Petch.git) ChakraPetch-BoldItalic.ttf: Copyright 2018 The Chakra Petch Project Authors (https://github.com/m4rc1e/Chakra-Petch.git)
+
+This Font Software is licensed under the SIL Open Font License, Version 1.1.
+This license is copied below, and is also available with a FAQ at:
+http://scripts.sil.org/OFL
+
+
+-----------------------------------------------------------
+SIL OPEN FONT LICENSE Version 1.1 - 26 February 2007
+-----------------------------------------------------------
+
+PREAMBLE
+The goals of the Open Font License (OFL) are to stimulate worldwide
+development of collaborative font projects, to support the font creation
+efforts of academic and linguistic communities, and to provide a free and
+open framework in which fonts may be shared and improved in partnership
+with others.
+
+The OFL allows the licensed fonts to be used, studied, modified and
+redistributed freely as long as they are not sold by themselves. The
+fonts, including any derivative works, can be bundled, embedded,
+redistributed and/or sold with any software provided that any reserved
+names are not used by derivative works. The fonts and derivatives,
+however, cannot be released under any other type of license. The
+requirement for fonts to remain under this license does not apply
+to any document created using the fonts or their derivatives.
+
+DEFINITIONS
+"Font Software" refers to the set of files released by the Copyright
+Holder(s) under this license and clearly marked as such. This may
+include source files, build scripts and documentation.
+
+"Reserved Font Name" refers to any names specified as such after the
+copyright statement(s).
+
+"Original Version" refers to the collection of Font Software components as
+distributed by the Copyright Holder(s).
+
+"Modified Version" refers to any derivative made by adding to, deleting,
+or substituting -- in part or in whole -- any of the components of the
+Original Version, by changing formats or by porting the Font Software to a
+new environment.
+
+"Author" refers to any designer, engineer, programmer, technical
+writer or other person who contributed to the Font Software.
+
+PERMISSION & CONDITIONS
+Permission is hereby granted, free of charge, to any person obtaining
+a copy of the Font Software, to use, study, copy, merge, embed, modify,
+redistribute, and sell modified and unmodified copies of the Font
+Software, subject to the following conditions:
+
+1) Neither the Font Software nor any of its individual components,
+in Original or Modified Versions, may be sold by itself.
+
+2) Original or Modified Versions of the Font Software may be bundled,
+redistributed and/or sold with any software, provided that each copy
+contains the above copyright notice and this license. These can be
+included either as stand-alone text files, human-readable headers or
+in the appropriate machine-readable metadata fields within text or
+binary files as long as those fields can be easily viewed by the user.
+
+3) No Modified Version of the Font Software may use the Reserved Font
+Name(s) unless explicit written permission is granted by the corresponding
+Copyright Holder. This restriction only applies to the primary font name as
+presented to the users.
+
+4) The name(s) of the Copyright Holder(s) or the Author(s) of the Font
+Software shall not be used to promote, endorse or advertise any
+Modified Version, except to acknowledge the contribution(s) of the
+Copyright Holder(s) and the Author(s) or with their explicit written
+permission.
+
+5) The Font Software, modified or unmodified, in part or in whole,
+must be distributed entirely under this license, and must not be
+distributed under any other license. The requirement for fonts to
+remain under this license does not apply to any document created
+using the Font Software.
+
+TERMINATION
+This license becomes null and void if any of the above conditions are
+not met.
+
+DISCLAIMER
+THE FONT SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO ANY WARRANTIES OF
+MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT
+OF COPYRIGHT, PATENT, TRADEMARK, OR OTHER RIGHT. IN NO EVENT SHALL THE
+COPYRIGHT HOLDER BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
+INCLUDING ANY GENERAL, SPECIAL, INDIRECT, INCIDENTAL, OR CONSEQUENTIAL
+DAMAGES, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+FROM, OUT OF THE USE OR INABILITY TO USE THE FONT SOFTWARE OR FROM
+OTHER DEALINGS IN THE FONT SOFTWARE.

src/songtrace/ultrastar.py

@@ -0,0 +1,73 @@
+import re
+from dataclasses import dataclass
+
+NOTE_RE = re.compile(r"^([:*FRG]) (-?\d+) (\d+) (-?\d+) (.*)$")
+
+
+@dataclass
+class Note:
+    kind: str
+    beat: int
+    length: int
+    pitch: int
+    text: str
+
+
+@dataclass
+class LineBreak:
+    beat: int
+
+
+@dataclass
+class UltraStarSong:
+    headers: list[tuple[str, str]]
+    body: list[Note | LineBreak]
+
+
+def parse_ultrastar(text: str) -> UltraStarSong:
+    headers, body = [], []
+    for line in text.splitlines():
+        line = line.rstrip("\r")
+        if not line.strip():
+            continue
+        if line.startswith("#"):
+            key, _, value = line[1:].partition(":")
+            headers.append((key, value))
+        elif line.startswith("E"):
+            break
+        elif line.startswith("-"):
+            nums = line[1:].split()
+            body.append(LineBreak(beat=int(nums[0])))
+        else:
+            m = NOTE_RE.match(line)
+            if m:
+                body.append(Note(m[1], int(m[2]), int(m[3]), int(m[4]), m[5]))
+    return UltraStarSong(headers, body)
+
+
+def serialize_ultrastar(song: UltraStarSong) -> str:
+    out = [f"#{k}:{v}" for k, v in song.headers]
+    for item in song.body:
+        if isinstance(item, LineBreak):
+            out.append(f"- {item.beat}")
+        else:
+            out.append(f"{item.kind} {item.beat} {item.length} {item.pitch} {item.text}")
+    out.append("E")
+    return "\n".join(out) + "\n"
+
+
+def get_header(song: UltraStarSong, key: str) -> str | None:
+    for k, v in song.headers:
+        if k.upper() == key.upper():
+            return v
+    return None
+
+
+def format_bpm(bpm: float) -> str:
+    if bpm == int(bpm):
+        return str(int(bpm))
+    return f"{bpm:.2f}".replace(".", ",")
+
+
+def parse_bpm(value: str) -> float:
+    return float(value.replace(",", "."))

src/songtrace/web.py

@@ -0,0 +1,417 @@
+import json
+import math
+import queue
+import re
+import shutil
+import threading
+import time
+import uuid
+from dataclasses import asdict, dataclass, field
+from pathlib import Path
+
+from fastapi import FastAPI, File, Form, HTTPException, UploadFile
+from fastapi.concurrency import run_in_threadpool
+from fastapi.responses import FileResponse
+from fastapi.staticfiles import StaticFiles
+
+from . import rebuild as rebuild_policy
+from .config import Config
+from .lyrics import normalize_lyrics
+from .pipeline import (STAGE_ORDER, PipelineError, SongInput, read_warnings,
+                       run_pipeline, song_slug)
+
+STATIC = Path(__file__).parent / "static"
+STAGE_LINE = re.compile(r"^\[(\w+)\] (?:running|cached, skipping)")
+
+
+@dataclass
+class Job:
+    id: str
+    title: str
+    artist: str
+    status: str = "queued"
+    stage: str | None = None
+    log: list[str] = field(default_factory=list)
+    output_dir: str | None = None
+    warnings: list[str] = field(default_factory=list)
+    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
+
+
+def _safe_filename(name: str | None, default: str) -> str:
+    safe = Path(name).name if name else ""
+    return safe if safe not in ("", ".", "..") else default
+
+
+async def _save_upload(upload: UploadFile, dest: Path) -> None:
+    with dest.open("wb") as f:
+        await run_in_threadpool(shutil.copyfileobj, upload.file, f)
+
+
+async def _store_uploads(cfg: Config, job_id: str, video: UploadFile | None,
+                         cover: UploadFile | None) -> tuple[Path | None, Path | None]:
+    updir = cfg.work_dir / "uploads" / job_id
+    paths = []
+    for upload, default in ((video, "video"), (cover, "cover")):
+        if not (upload and upload.filename):
+            paths.append(None)
+            continue
+        updir.mkdir(parents=True, exist_ok=True)
+        dest = updir / _safe_filename(upload.filename, default)
+        await _save_upload(upload, dest)
+        paths.append(dest)
+    return paths[0], paths[1]
+
+
+def _cleanup_uploads(cfg: Config, job_id: str, video_path: Path | None) -> None:
+    """Drop the video upload a finished job no longer needs.
+
+    Covers are left behind: they are small, and a rebuild inherits cover_path, so
+    removing one would silently strip #COVER from the rebuilt song. A rebuild can
+    also inherit its parent's upload, so the directory is not always uploads/<job_id>.
+    Never raises — the caller has already marked the job done.
+    """
+    uploads = (cfg.work_dir / "uploads").resolve()
+    dirs = {uploads / job_id}
+    if video_path is not None:
+        target = Path(video_path)
+        parent = target.resolve().parent
+        if parent.parent == uploads:
+            try:
+                target.unlink(missing_ok=True)
+            except OSError:
+                pass
+            dirs.add(parent)
+    for d in dirs:
+        try:
+            d.rmdir()
+        except OSError:
+            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")
+    jobs: dict[str, Job] = {}
+    order: list[str] = []
+    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 job.force_stage:
+            log(f"[rebuild] restarting from {job.force_stage} (forced rebuild from scratch)")
+            return job.force_stage
+        if not job.rebuilt_from:
+            return None
+        with lock:
+            parent = jobs.get(job.rebuilt_from)
+        if parent is None:
+            return None
+        force_from = 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)
+        log(f"[rebuild] restarting from {force_from or 'the first incomplete stage'}"
+            f" ({', '.join(changed) if changed else 'nothing'} changed)")
+        return force_from
+
+    def _drop_superseded_output(job: Job, log) -> None:
+        if not job.rebuilt_from:
+            return
+        with lock:
+            parent = jobs.get(job.rebuilt_from)
+        if parent is None or parent.status != "done":
+            return
+        old = rebuild_policy.superseded_output(parent.output_dir, job.output_dir,
+                                                cfg.output_dir)
+        if old is None:
+            return
+        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()
+            with lock:
+                job = jobs[job_id]
+                job.status = "running"
+                job.started_at = time.time()
+
+            def log(msg: str, job=job):
+                with lock:
+                    job.log.append(msg)
+                    if m := STAGE_LINE.match(msg):
+                        job.stage = m[1]
+
+            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)
+                    job.warnings = list(result.warnings)
+                _drop_superseded_output(job, log)
+                with lock:
+                    job.status = "done"
+                # uploads survive a failure so the job stays rerunnable; acquire
+                # has already copied the video into the work dir by now.
+                _cleanup_uploads(cfg, job_id, job.inp.video_path)
+            except BaseException as e:
+                with lock:
+                    job.status = "failed"
+                    job.error = str(e)
+            finally:
+                with lock:
+                    job.finished_at = time.time()
+
+    threading.Thread(target=worker, daemon=True).start()
+
+    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, force_stage=force_stage)
+        with lock:
+            jobs[job_id] = job
+            order.append(job_id)
+        todo.put(job_id)
+        return {"id": job_id}
+
+    def _finished_job(job_id: str) -> Job:
+        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")
+            return job
+
+    @app.get("/")
+    def index():
+        return FileResponse(STATIC / "index.html")
+
+    @app.post("/api/jobs", status_code=201)
+    async def submit(title: str = Form(...), artist: str = Form(...),
+                      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),
+                      video: UploadFile | None = File(None),
+                      cover: UploadFile | None = File(None)):
+        youtube_url, genre, language = (youtube_url or None, genre or None,
+                                        language or None)
+        has_video = bool(video and video.filename)
+        if bool(youtube_url) == has_video:
+            raise HTTPException(400, "provide exactly one of youtube_url / video")
+        if bpm is not None and (not math.isfinite(bpm) or bpm <= 0):
+            raise HTTPException(400, f"bpm must be a finite value > 0, got {bpm!r}")
+
+        job_id = uuid.uuid4().hex
+        video_path, cover_path = await _store_uploads(cfg, job_id, video, cover)
+
+        inp = SongInput(title=title, artist=artist, lyrics_text=lyrics,
+                         youtube_url=youtube_url, video_path=video_path,
+                         genre=genre, year=year, language=language,
+                         cover_path=cover_path, bpm_override=bpm)
+        return _enqueue(job_id, inp)
+
+    @app.post("/api/jobs/{job_id}/rerun", status_code=201)
+    def rerun(job_id: str):
+        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."""
+        if inp.video_path is None or inp.video_path.exists():
+            return
+        if force_from != "acquire":
+            # acquire is cached under the old slug until the worker renames it
+            for slug in {song_slug(parent_inp), song_slug(inp)}:
+                work = cfg.work_dir / slug
+                if (work / "meta.json").exists() and (work / "audio.wav").exists():
+                    return
+        raise HTTPException(400, "the original video upload is no longer available; "
+                                 "attach the video file again")
+
+    @app.post("/api/jobs/{job_id}/rebuild", status_code=201)
+    async def rebuild(job_id: str, title: str = Form(...), artist: str = Form(...),
+                      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),
+                      video: UploadFile | None = File(None),
+                      cover: UploadFile | None = File(None)):
+        parent = _finished_job(job_id)
+        # 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)
+        has_video = bool(video and video.filename)
+        if youtube_url and has_video:
+            raise HTTPException(400, "provide exactly one of youtube_url / video")
+        if bpm is not None and (not math.isfinite(bpm) or bpm <= 0):
+            raise HTTPException(400, f"bpm must be a finite value > 0, got {bpm!r}")
+
+        new_id = uuid.uuid4().hex
+        video_path, cover_path = await _store_uploads(cfg, new_id, video, cover)
+        if not youtube_url and not has_video:
+            youtube_url, video_path = parent.inp.youtube_url, parent.inp.video_path
+        if cover_path is None:
+            cover_path = parent.inp.cover_path
+
+        inp = SongInput(title=title, artist=artist, lyrics_text=lyrics,
+                        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)
+
+    @app.get("/api/stages")
+    def list_stages():
+        return {"stages": list(STAGE_ORDER)}
+
+    @app.get("/api/jobs")
+    def list_jobs():
+        with lock:
+            return [{"id": j.id, "title": j.title, "artist": j.artist,
+                      "status": j.status, "stage": j.stage,
+                      "created_at": j.created_at, "started_at": j.started_at,
+                      "finished_at": j.finished_at, "rebuilt_from": j.rebuilt_from}
+                     for j in (jobs[i] for i in reversed(order))]
+
+    @app.get("/api/jobs/{job_id}")
+    def get_job(job_id: str):
+        with lock:
+            job = jobs.get(job_id)
+            if job is None:
+                raise HTTPException(404)
+            return {"id": job.id, "title": job.title, "artist": job.artist,
+                    "status": job.status, "stage": job.stage,
+                    "log": job.log[-100:], "output_dir": job.output_dir,
+                    "warnings": list(job.warnings), "error": job.error,
+                    "created_at": job.created_at, "started_at": job.started_at,
+                    "finished_at": job.finished_at, "rebuilt_from": job.rebuilt_from}
+
+    @app.get("/api/jobs/{job_id}/input")
+    def get_input(job_id: str):
+        with lock:
+            job = jobs.get(job_id)
+            if job is None:
+                raise HTTPException(404)
+            inp = job.inp
+        # keys are the form's field names so the page can prefill by name;
+        # paths are reduced to filenames, which is all the browser can use.
+        return {"title": inp.title, "artist": inp.artist, "lyrics": inp.lyrics_text,
+                "youtube_url": inp.youtube_url, "genre": inp.genre, "year": inp.year,
+                "language": inp.language, "bpm": inp.bpm_override,
+                "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)
+
+    @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)]
+        return {"lines": lines, "missing_spans": missing_spans}
+
+    return app
+
+
+def serve(cfg: Config) -> int:
+    import uvicorn
+    uvicorn.run(create_app(cfg), host=cfg.bind_host, port=cfg.port)
+    return 0

src/songtrace/whisper_client.py

@@ -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()

tests/fixtures/sanctuary.txt

@@ -0,0 +1,472 @@
+#TITLE:Sanctuary
+#ARTIST:Utada Hikaru
+#LANGUAGE:English
+#YEAR:2009
+#CREATOR:Subkei
+#GENRE:J-Pop
+#COVER:Utada Hikaru - Sanctuary [CO].jpg
+#MP3:Utada Hikaru - Sanctuary.mp3
+#VIDEO:Utada Hikaru - Sanctuary.mp4
+#BPM:217,98
+#GAP:4880
+#MEDLEYSTARTBEAT:2471
+#MEDLEYENDBEAT:2990
+: 0 3 16 ah
+: 5 3 20  ah
+: 11 3 23  ah
+: 17 3 16  ah
+: 22 28 15  ah
+: 51 2 16  ah
+: 54 31 16 ~
+- 95
+F 113 3 18 wonk
+F 117 3 21  uoy
+F 121 3 21  naht
+F 125 3 21  snoit
+F 129 3 21 cef
+F 133 3 18 fa
+F 137 3 18  erom
+F 144 2 18  deen
+F 147 4 18  I
+- 180
+: 425 3 16 In
+: 429 3 21  you
+: 433 2 20 ~
+: 436 4 16 ~
+: 441 2 18  and
+: 445 4 20  I
+: 451 4 18 ~
+- 459
+: 461 3 16 there's
+: 465 3 16  a
+: 469 5 20  new
+: 475 4 23  land
+: 481 4 16 ~
+: 487 27 15 ~
+: 515 4 16 ~
+: 521 17 16 ~
+: 539 1 18 ~
+: 541 3 20 ~
+- 547
+: 549 1 16 An
+: 551 2 18 ~
+: 556 3 18 gels
+: 560 4 18  in
+: 565 7 20  flight
+: 573 2 18 ~
+- 576
+F 578 2 18 wonk
+F 581 3 21  uoy
+F 585 3 21  naht
+F 589 3 21  snoit
+F 593 3 21 cef
+F 597 3 18 fa
+F 601 3 18  erom
+F 608 2 18  deen
+F 611 3 18  I
+- 615
+: 617 1 16 My
+: 620 1 21  sanc
+: 624 2 20 tu
+: 628 4 16 a
+: 634 8 18 ry
+- 646
+: 648 3 16 my
+: 653 1 21  sanc
+: 657 2 20 tu
+: 661 3 16 a
+: 665 4 16 ry
+: 670 2 18 ~
+F 674 4 13  yeah
+- 680
+: 682 2 16 Where
+: 686 1 21  fears
+: 689 1 20 ~
+: 691 5 16 ~
+: 697 1 18  and
+* 701 4 20  lies
+* 707 4 18 ~
+- 715
+: 717 3 16 melt
+: 721 3 16  a
+: 725 4 20 way
+: 732 4 23 ~
+: 737 3 16 ~
+: 742 28 15 ~
+: 771 3 16 ~
+: 775 19 16 ~
+: 795 1 18 ~
+: 797 3 20 ~
+- 804
+: 806 1 16 Mu
+: 808 2 18 ~
+: 812 3 18 sic
+: 817 3 18  will
+: 822 7 20  tie
+: 830 2 18 ~
+- 833
+F 835 1 18 wonk
+F 837 3 21  uoy
+F 841 3 21  naht
+F 845 3 21  snoit
+F 849 3 21 cef
+F 853 3 18 fa
+F 857 3 18  erom
+F 864 2 18  deen
+F 867 3 18  I
+- 871
+: 873 2 16 What's
+: 877 1 21  left
+: 881 1 20  of
+: 884 3 16  me
+: 888 1 16 ~
+: 890 7 18 ~
+- 903
+: 905 2 16 What's
+: 909 1 21  left
+: 913 1 20  of
+: 916 3 16  me
+: 921 2 16 ~
+: 925 3 18 ~
+: 933 24 13 now?
+- 986
+: 1069 3 13 I
+: 1074 6 16  watch
+: 1082 3 16  you
+: 1087 9 18 ~
+- 1102
+: 1105 3 18 fast
+: 1109 3 20 ~
+* 1117 2 13  a
+* 1122 5 16 sleep
+: 1129 3 18 ~
+: 1133 15 16 ~
+- 1162
+: 1189 3 20 All
+: 1193 6 21 ~
+: 1201 3 18  I
+: 1205 6 20 ~
+: 1213 5 16  fear
+: 1220 20 13 ~
+- 1246
+: 1249 2 11 means
+: 1255 7 13  no
+: 1266 5 18 thing
+: 1273 41 20 ~
+- 1319
+: 1321 2 16 In
+: 1325 2 21  you
+: 1328 2 20 ~
+: 1332 3 16 ~
+: 1337 2 18  and
+: 1341 4 20  I
+: 1347 4 18 ~
+- 1355
+: 1357 3 16 there's
+: 1361 3 16  a
+* 1365 5 20  new
+: 1371 4 23  land
+: 1377 4 16 ~
+: 1383 27 15 ~
+: 1411 4 16 ~
+: 1417 17 16 ~
+: 1435 1 18 ~
+: 1437 3 20 ~
+- 1442
+: 1444 1 16 An
+: 1446 2 18 ~
+: 1452 3 18 gels
+: 1457 3 18  in
+: 1462 6 20  flight
+: 1469 2 18 ~
+- 1472
+F 1474 1 18 wonk
+F 1476 3 21  uoy
+F 1480 3 21  naht
+F 1484 3 21  snoit
+F 1488 3 21 cef
+F 1492 3 18 fa
+F 1496 3 18  erom
+F 1503 2 18  deen
+F 1506 3 18  I
+- 1511
+: 1513 1 16 My
+: 1516 1 21  sanc
+: 1520 2 20 tu
+: 1524 4 16 a
+: 1530 8 18 ry
+- 1541
+: 1543 3 16 my
+: 1548 1 21  sanc
+: 1552 2 20 tu
+: 1556 3 16 a
+: 1560 4 16 ry
+: 1565 2 18 ~
+F 1570 3 13  yeah
+- 1575
+: 1577 2 16 Where
+* 1581 1 21  fears
+* 1584 1 20 ~
+: 1586 5 16 ~
+: 1592 1 18  and
+: 1596 4 20  lies
+: 1602 4 18 ~
+- 1611
+: 1613 3 16 melt
+: 1617 3 16  a
+: 1621 4 20 way
+: 1628 4 23 ~
+: 1633 3 16 ~
+: 1638 27 15 ~
+: 1666 4 16 ~
+: 1672 18 16 ~
+: 1691 1 18 ~
+: 1693 3 20 ~
+- 1699
+: 1701 1 16 Mu
+: 1703 2 18 ~
+: 1707 3 18 sic
+: 1712 3 18  will
+: 1717 7 20  tie
+: 1725 2 18 ~
+- 1728
+F 1730 2 18 wonk
+F 1733 3 21  uoy
+F 1737 3 21  naht
+F 1741 3 21  snoit
+F 1745 3 21 cef
+F 1749 3 18 fa
+F 1753 3 18  erom
+F 1760 2 18  deen
+F 1763 3 18  I
+- 1767
+: 1769 2 16 What's
+: 1773 1 21  left
+: 1777 1 20  of
+: 1780 3 16  me
+: 1785 1 16 ~
+: 1787 7 18 ~
+- 1798
+: 1800 2 16 What's
+: 1804 1 21  left
+* 1808 1 20  of
+* 1811 3 16  me
+: 1815 1 16 ~
+: 1817 7 18 ~
+- 1831
+F 1834 2 21 snwod
+F 1837 2 20  dna
+F 1841 3 18  spu
+F 1845 3 20  y
+F 1849 3 20 nam
+F 1853 3 18  os
+- 1870
+: 1897 3 20 My
+* 1901 2 20  heart's
+: 1904 1 18  a
+: 1906 3 20  bat
+: 1910 3 21 tle
+: 1914 5 21 ground
+- 1926
+F 1929 2 16 snoit
+F 1932 6 16 om
+F 1941 2 18 e
+F 1945 2 18 ~
+F 1948 2 18  eurt
+F 1952 2 18  deen
+F 1955 2 16  I
+- 1971
+F 1987 3 18 wonk
+F 1991 3 21  uoy
+F 1995 3 21  naht
+F 1999 3 21  snoit
+F 2003 3 21 cef
+F 2007 3 18 fa
+F 2011 3 18  erom
+F 2018 2 18  deen
+F 2021 3 18  I
+- 2038
+F 2059 2 16 snoit
+F 2062 6 16 om
+F 2070 3 18 e
+F 2075 2 18 ~
+F 2078 2 18  eurt
+F 2082 2 18  deen
+F 2085 2 16 I
+- 2116
+: 2219 3 13 You
+: 2224 6 16  show
+: 2232 3 16  me
+: 2237 7 18 ~
+- 2252
+: 2255 3 18 how
+: 2259 3 20 ~
+: 2267 2 13  to
+: 2272 5 16  see
+- 2282
+* 2284 3 13 That
+* 2289 6 16  no
+* 2297 3 16 thing
+* 2302 5 18 ~
+- 2313
+: 2316 3 13 is
+: 2321 3 16  whole
+: 2325 2 15 ~
+: 2328 2 13  and
+: 2331 1 13 ~
+: 2333 2 15 ~
+- 2338
+: 2340 3 20 no
+: 2344 6 21 ~
+: 2352 3 18 ~
+: 2356 6 20 ~
+: 2364 5 16 thing
+: 2371 12 13 ~
+- 2395
+: 2399 2 11 is
+: 2403 9 13  bro
+: 2416 5 18 ken
+: 2423 41 20 ~
+- 2470
+: 2472 2 16 In
+* 2476 2 21  you
+* 2479 2 20 ~
+* 2483 3 16 ~
+: 2488 2 18  and
+: 2492 4 20  I
+: 2499 3 18 ~
+- 2506
+: 2508 3 16 There's
+: 2512 3 16  a
+: 2516 5 20  new
+: 2522 4 23  land
+: 2528 4 16 ~
+: 2534 27 15 ~
+: 2562 4 16 ~
+: 2568 17 16 ~
+: 2586 1 18 ~
+: 2588 3 20 ~
+- 2593
+: 2595 1 16 An
+: 2597 2 18 ~
+: 2603 3 18 gels
+: 2608 3 18  in
+: 2613 6 20  flight
+: 2620 2 18 ~
+- 2623
+F 2625 2 18 wonk
+F 2628 3 21  uoy
+F 2632 3 21  naht
+F 2636 3 21  snoit
+F 2640 3 21 cef
+F 2644 3 18 fa
+F 2648 3 18  erom
+F 2655 2 18  deen
+F 2658 2 18  I
+- 2661
+: 2663 2 16 My
+: 2668 1 21  sanc
+: 2672 1 20 tu
+: 2675 4 16 a
+: 2681 7 18 ry
+- 2693
+: 2695 3 16 My
+: 2700 1 21  sanc
+: 2704 2 20 tu
+: 2708 3 16 a
+: 2712 4 16 ry
+: 2717 2 18 ~
+F 2721 4 13  yeah
+- 2726
+: 2728 2 16 Where
+: 2732 1 21  fears
+: 2735 1 20 ~
+: 2737 5 16 ~
+: 2743 1 18  and
+: 2747 4 20  lies
+: 2754 3 18 ~
+- 2761
+: 2763 3 16 melt
+: 2767 3 16  a
+: 2771 4 20 way
+: 2778 4 23 ~
+: 2783 4 16 ~
+: 2789 27 15 ~
+: 2817 4 16 ~
+: 2823 17 16 ~
+: 2841 1 18 ~
+: 2843 3 20 ~
+- 2849
+: 2851 1 16 Mu
+: 2853 2 18 ~
+: 2857 3 18 sic
+* 2862 3 18  will
+* 2868 7 20  tie
+: 2877 2 18 ~
+- 2880
+F 2882 1 18 wonk
+F 2884 3 21  uoy
+F 2888 3 21  naht
+F 2892 3 21  snoit
+F 2896 3 21 cef
+F 2900 3 18 fa
+F 2904 3 18  erom
+F 2911 2 18  deen
+F 2914 2 18  I
+- 2917
+: 2919 2 16 What's
+: 2923 1 21  left
+: 2927 1 20  of
+: 2930 3 16  me
+: 2935 1 16 ~
+: 2937 7 18 ~
+- 2949
+: 2951 2 16 What's
+: 2955 1 21  left
+* 2959 1 20  of
+* 2962 3 16  me
+* 2967 2 16 ~
+: 2971 3 18 ~
+: 2979 12 13  now?
+- 3005
+: 3026 2 11 uh
+: 3029 2 13 ~
+: 3032 3 16 ~
+: 3036 8 15 ~
+: 3045 4 18 ~
+: 3051 4 16 ~
+: 3056 3 15 ~
+: 3060 3 13 ~
+- 3066
+: 3068 1 16 My
+: 3070 2 18 ~
+: 3074 12 18  fears
+- 3096
+: 3100 1 16 my
+: 3102 2 18 ~
+: 3107 3 18  lies
+: 3111 2 16 ~
+: 3114 1 13 ~
+- 3129
+* 3164 3 16 melt
+* 3169 3 16  a
+* 3173 4 20 way
+* 3179 4 23 ~
+* 3185 3 16 ~
+* 3190 27 15 ~
+* 3218 3 16 ~
+* 3223 19 16 ~
+- 3248
+: 3250 4 18 ah
+: 3256 126 20 ~
+- 3392
+F 3410 1 18 wonk
+F 3412 3 21  uoy
+F 3416 3 21  naht
+F 3420 3 21  snoit
+F 3424 3 21 cef
+F 3428 3 18 fa
+F 3432 3 18  erom
+F 3439 2 18  deen
+F 3442 3 18  I
+E

tests/fixtures/sweet_child.txt

@@ -0,0 +1,494 @@
+#TITLE:Sweet Child O´Mine
+#ARTIST:Guns N' Roses
+#LANGUAGE:English
+#GENRE:Rock
+#YEAR:1987
+#MP3:Guns N' Roses - Sweet Child O´Mine.mp3
+#COVER:Guns N' Roses - Sweet Child O´Mine.jpg
+#VIDEO:Guns N' Roses - Sweet Child O´Mine.webm
+#CREATOR:sanicsip
+#FIXER:Pacayaka
+#BPM:256
+#GAP:49673
+#VIDEOGAP:0,3
+#START:13
+#MEDLEYSTARTBEAT:1564
+#MEDLEYENDBEAT:2057
+: 0 3 8 She's
+: 5 2 8  got
+: 8 1 8  a
+: 13 2 6  smile
+: 16 8 8 ~
+: 25 1 6  that
+: 27 2 5  it
+: 32 7 6  seems
+: 41 1 8  to
+: 45 1 5  me
+: 47 5 4 ~
+: 53 1 3 ~
+- 55
+: 57 1 1 re
+: 60 7 6 minds
+: 69 2 5  me
+: 73 1 1  of
+: 76 1 1  child
+: 79 9 6 ~
+: 91 5 6 hood
+- 99
+: 101 3 6 mem
+: 105 4 5 o
+: 111 6 3 ries
+: 118 1 1 ~
+- 120
+: 122 3 1 where
+: 126 2 4  eve
+: 129 5 6 ~
+: 135 2 5 ~
+: 138 1 1 ry
+: 142 2 4 thing
+: 145 4 6 ~
+: 150 2 5  was
+: 154 1 1  as
+: 158 2 5  fresh
+: 161 4 6 ~
+: 166 3 5  as
+: 171 1 1  the
+: 174 5 6  bright
+: 181 5 6  blue
+: 190 8 6  sky
+: 200 5 5 ~
+- 222
+: 258 4 8 Now
+: 264 4 8  and
+: 271 10 8  then
+: 283 2 6  when
+: 287 2 5  I
+: 291 3 6  see
+: 295 1 5  her
+: 297 1 6 ~
+: 299 2 8 ~
+: 304 8 5  face
+- 313
+: 315 3 6 she
+: 319 3 8 ~
+: 324 3 6  takes
+: 330 2 5  me
+: 334 1 1  a
+: 338 2 3 way
+: 341 3 6 ~
+: 345 2 5  to
+: 349 3 1  that
+- 354
+: 356 3 6 spe
+: 360 3 6 cial
+: 364 2 8 ~
+: 369 1 6  place
+: 371 4 5 ~
+: 377 2 5  and
+: 381 1 6  if
+: 384 2 5  I
+: 389 3 6  stare
+: 393 2 5 ~
+: 398 3 6  too
+: 402 2 5 ~
+: 405 4 6  long
+: 410 1 5 ~
+- 415
+: 417 1 1 I'd
+: 420 1 1  prob
+: 423 1 6 a
+: 425 2 6 bly
+: 429 5 6  break
+: 436 5 6  down
+: 442 2 6  and
+* 445 1 8 ~
+: 450 7 6  cry
+: 458 8 5 ~
+- 483
+: 512 3 6 Whoah
+: 517 24 8 ~
+: 543 3 10  oh
+: 548 2 8  ooh
+: 551 3 8  whoah
+: 555 1 6 ~
+: 559 3 6  sweet
+: 563 1 5 ~
+: 567 2 5  child
+: 570 3 6 ~
+: 575 1 6  o'
+: 578 8 6  mine
+: 587 12 5 ~
+- 616
+: 643 1 6 Whoah
+: 645 11 8 ~
+: 659 7 6  oh
+: 668 7 5  oh
+: 677 2 5  oh
+: 680 3 6 ~
+: 684 3 5 ~
+: 689 5 1  sweet
+: 697 7 3  love
+: 706 2 5  of
+: 711 26 1  mine
+- 771
+: 1041 5 8 She's
+: 1048 6 8  got
+: 1056 2 6  eyes
+: 1059 7 8 ~
+: 1068 3 6  of
+: 1073 1 5  the
+: 1076 3 6  blu
+: 1080 3 6 est
+: 1084 2 8 ~
+: 1089 11 3  skies
+: 1102 5 1  as
+: 1110 5 6  if
+: 1118 7 6  they
+- 1128
+: 1130 4 6 thought
+: 1135 2 5  of
+: 1140 2 5  rain
+: 1143 4 6 ~
+: 1148 4 5 ~
+- 1162
+* 1171 4 8 I
+: 1176 2 5  hate
+: 1179 3 6 ~
+: 1184 2 6  to
+: 1187 3 5 ~
+: 1192 2 6  look
+: 1196 10 6  in
+: 1208 3 6 to
+: 1213 6 5  those
+: 1221 2 5  eyes
+: 1224 7 6 ~
+: 1233 4 6  and
+- 1239
+: 1241 4 6 see
+: 1246 1 5 ~
+: 1249 3 1  an
+: 1253 1 3  ounce
+: 1256 5 6 ~
+: 1262 1 5 ~
+: 1265 2 1  of
+: 1269 15 1  pain
+- 1294
+: 1301 3 6 Her
+: 1305 2 6  hair
+: 1308 5 8 ~
+: 1314 2 8  re
+: 1318 6 8 minds
+: 1326 4 8  me
+: 1331 3 6  of
+: 1335 2 5  a
+: 1339 6 6  warm
+: 1347 2 6  safe
+: 1350 3 8 ~
+: 1355 6 5  place
+- 1362
+: 1364 6 1 where
+: 1372 3 6  as
+: 1376 2 5 ~
+: 1380 2 1  a
+: 1384 2 5  child
+: 1387 7 6 ~
+* 1395 3 8  I'd
+: 1399 5 6  hide
+: 1405 1 5 ~
+: 1408 11 1 ~
+- 1428
+: 1431 1 6 and
+: 1433 2 8 ~
+: 1437 2 5  pray
+: 1440 3 6 ~
+: 1445 3 6  for
+: 1449 2 5  the
+: 1453 7 6  thun
+: 1461 6 6 der
+- 1471
+: 1473 4 6 and
+: 1478 2 5  the
+: 1482 1 5  rain
+: 1484 6 6 ~
+- 1491
+: 1493 1 5 to
+: 1496 8 6  qui
+: 1505 2 5 et
+: 1509 2 1 ly
+: 1514 1 5  pass
+: 1516 6 6 ~
+* 1526 2 8  me
+: 1529 5 6  by
+: 1535 1 5 ~
+: 1538 10 1 ~
+- 1558
+: 1566 3 6 Whoah
+: 1570 19 8 ~
+: 1591 3 10  oh
+: 1595 3 8  ooh
+: 1600 2 8  whoah
+: 1603 1 6 ~
+: 1607 4 6  sweet
+: 1612 1 5 ~
+: 1615 3 5  child
+: 1620 2 6 ~
+: 1623 1 6  o'
+: 1629 7 6  mine
+: 1637 11 5 ~
+- 1658
+: 1675 2 12 Whoah
+: 1678 20 13 ~
+: 1699 1 12 ~
+: 1701 5 13 ~
+: 1711 1 8  oh
+: 1713 4 10 ~
+: 1719 7 8  oh
+: 1727 3 8  oh
+: 1731 3 10 ~
+: 1736 1 8 ~
+- 1737
+: 1739 3 6 Sweet
+: 1743 1 5 ~
+: 1748 7 6  love
+* 1757 2 8  of
+: 1761 4 6  mine
+: 1766 1 5 ~
+: 1770 16 1 ~
+- 1803
+* 1829 8 8 Oh
+* 1842 4 10  oh
+* 1850 4 13  oh
+* 1858 2 13  oh
+* 1861 5 15 ~
+* 1867 1 13 ~
+- 1869
+: 1871 3 13 Sweet
+: 1875 1 10 ~
+: 1880 1 10  child
+: 1882 3 13 ~
+: 1886 1 10 ~
+: 1888 2 8  o'
+: 1892 2 8  mine
+: 1895 5 10 ~
+: 1901 2 13 ~
+: 1905 2 10 ~
+: 1909 9 8 ~
+- 1924
+: 1926 2 8 Ooh
+: 1929 4 10 ~
+: 1934 2 8 ~
+: 1937 3 5 ~
+: 1942 3 8  yeah
+: 1951 1 8  yeah
+: 1953 5 10 ~
+: 1959 2 8 ~
+- 1962
+: 1964 1 8 Ooh
+: 1966 7 10 ~
+: 1974 7 8 ~
+: 1982 7 6 ~
+: 1991 2 5  ooh
+: 1994 4 6 ~
+: 1999 1 5 ~
+- 2002
+: 2004 3 1 sweet
+: 2010 9 3  love
+: 2021 1 5  of
+: 2025 1 3  mine
+: 2027 1 5 ~
+: 2029 2 3 ~
+: 2032 25 1 ~
+- 2091
+: 3176 4 3 Where
+: 3181 3 1  do
+: 3185 2 -2  we
+: 3190 8 -4  go?
+- 3204
+: 3207 1 1 Where
+: 3209 4 3 ~
+: 3214 2 1  do
+: 3217 2 -2  we
+: 3221 6 -4  go
+: 3229 8 -6  now?
+- 3239
+: 3241 3 3 Where
+: 3245 3 1  do
+: 3249 2 -2  we
+: 3253 11 -4  go?
+- 3274
+: 3282 1 -4 Ooh
+: 3284 4 -2 ~
+: 3289 3 -4  oh
+: 3293 3 -2 ~
+: 3297 3 -4  oh
+: 3301 1 -6 ~
+- 3304
+: 3306 2 -2 Where
+: 3309 2 -4  do
+: 3312 2 -6  we
+: 3316 4 -2  go
+: 3321 4 -4 ~
+: 3326 6 -2 ~?
+- 3342
+: 3355 5 -4 Oh
+: 3362 7 -2 ~
+: 3370 2 -2  where
+: 3373 2 -2  do
+: 3376 2 -2  we
+: 3380 3 -4  go
+: 3384 1 -6 ~
+: 3387 13 -9  now?
+- 3410
+: 3430 1 -6 Where
+: 3433 1 -2 ~
+: 3435 2 -2  do
+: 3439 2 1  we
+: 3443 3 -4  go
+: 3447 8 -2 ~?
+- 3465
+: 3474 7 3 Ooh
+: 3483 12 1  oh
+: 3497 2 3  where
+: 3500 2 1  do
+: 3503 2 -2  we
+: 3507 6 -4  go
+: 3514 13 -6  now?
+- 3537
+: 3558 1 -4 ay
+: 3560 6 -2 ~
+: 3567 1 -4  ay
+: 3569 5 -2 ~
+: 3575 1 -4  ay
+: 3577 5 -2 ~
+: 3583 1 -4  ay
+: 3585 5 -2 ~
+: 3591 1 -4  ay
+: 3593 6 -2 ~
+: 3600 1 -4  ay
+: 3602 5 -2 ~
+: 3608 2 -4  ay
+: 3611 2 -2 ~
+- 3614
+: 3616 4 10 Oh
+: 3623 1 13  where
+: 3625 1 15 ~
+: 3627 3 13  do
+: 3631 3 10  we
+: 3635 6 8  go
+: 3643 5 6  now?
+- 3649
+* 3651 2 6 ah
+* 3655 2 10  ah
+* 3659 2 13  ah
+* 3663 2 10  ah
+* 3666 1 8  ah
+* 3668 2 10  ah
+* 3673 2 8  ah
+* 3676 2 6  oh
+* 3680 2 8  ah
+* 3683 1 6  oh
+- 3684
+: 3686 1 8 where
+: 3688 2 10 ~
+: 3691 2 8  do
+: 3695 2 6  we
+: 3699 1 8  go
+: 3701 11 10 ~?
+- 3719
+* 3722 2 13 ah
+* 3725 12 15 ~
+* 3738 8 13 ~
+* 3748 1 10  oh
+* 3752 2 15  where
+* 3756 2 13  do
+* 3760 2 10  we
+* 3764 7 8  go
+* 3772 10 6  now?
+- 3792
+: 3816 3 10 Where
+: 3821 2 10  do
+: 3825 2 13  we
+: 3829 1 8  go
+: 3831 4 10 ~
+: 3837 1 13 ~
+: 3841 2 10 ~
+: 3845 1 13 ~
+: 3849 1 10 ~
+: 3851 1 8 ~
+: 3853 1 10 ~
+: 3855 1 8 ~
+: 3857 1 10 ~
+: 3859 1 8 ~
+: 3861 2 10 ~
+: 3865 2 8 ~
+: 3868 3 6 ~
+: 3872 3 8 ~
+: 3876 2 6 ~?
+- 3879
+: 3881 2 10 where
+: 3884 1 8  do
+: 3886 1 10 ~
+: 3888 2 10  we
+: 3892 3 8  go
+: 3897 1 6 ~
+: 3901 11 3  now?
+- 3922
+: 3943 1 8 Where
+: 3945 2 10 ~
+: 3949 2 8  do
+: 3953 2 6  we
+: 3957 2 10  go
+: 3960 2 8 ~
+: 3963 10 10 ~?
+- 3983
+: 3990 8 10 Whoah
+: 4000 3 13 ~
+: 4004 3 10 ~
+: 4008 3 10  where
+: 4012 3 13  do
+: 4017 2 10  we
+: 4021 5 8  go
+: 4028 10 6  now?
+- 4039
+* 4041 3 15 no
+* 4046 3 13  no
+* 4051 4 10  no
+* 4057 3 9  no
+* 4062 3 8  no
+* 4067 3 6  no
+* 4072 9 3  no
+- 4087
+: 4090 3 8 sweet
+: 4094 2 10 ~
+: 4101 3 8  child
+: 4105 2 6 ~
+: 4109 7 3 ~
+- 4122
+: 4124 3 10 sweet
+: 4128 1 8 ~
+: 4130 3 10 ~
+: 4137 3 10  child
+: 4142 2 13 ~
+: 4145 4 10 ~
+: 4150 2 13 ~
+: 4153 4 10 ~
+: 4158 1 8 ~
+: 4160 2 10 ~
+: 4163 2 8 ~
+: 4166 3 6 ~
+: 4171 2 10 ~
+: 4175 2 13 ~
+: 4179 3 10 ~
+: 4183 4 13 ~
+: 4188 4 15 ~
+: 4194 2 13 ~
+: 4199 1 10 ~
+- 4200
+: 4202 1 8 o'
+: 4204 2 10 ~
+: 4208 2 8  mine
+: 4211 2 6 ~
+: 4214 46 3 ~
+: 4262 33 1 ~
+: 4297 49 3 ~
+E

tests/test_acquire.py

@@ -0,0 +1,177 @@
+import json
+from pathlib import Path
+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)
+    cfg = Config(whisper_url="http://x/v1", work_dir=tmp_path)
+    work = tmp_path / "A - T"
+    work.mkdir()
+    return Ctx(inp=inp, cfg=cfg, work=work, log=lambda m: None)
+
+def test_local_video(tmp_path, monkeypatch):
+    src = tmp_path / "in.mp4"
+    src.write_bytes(b"fake")
+    calls = []
+    monkeypatch.setattr(acquire, "ffmpeg", lambda *a: calls.append(a))
+    ctx = make_ctx(tmp_path, video_path=src)
+    acquire.acquire_stage().run(ctx)
+    meta = json.loads((ctx.work / "meta.json").read_text())
+    assert meta["video_file"] == "source_video.mp4"
+    assert (ctx.work / "source_video.mp4").read_bytes() == b"fake"
+    assert any("audio.wav" in str(a[-1]) for a in calls)
+
+def test_youtube(tmp_path, monkeypatch):
+    def fake_dl(url, dest, cfg, log):
+        video = dest / "source_video.webm"
+        video.write_bytes(b"v")
+        thumb = dest / "source_video.jpg"
+        thumb.write_bytes(b"t")
+        return video, thumb
+    monkeypatch.setattr(acquire, "download_youtube", fake_dl)
+    monkeypatch.setattr(acquire, "ffmpeg", lambda *a: None)
+    ctx = make_ctx(tmp_path, youtube_url="https://yt/x")
+    acquire.acquire_stage().run(ctx)
+    meta = json.loads((ctx.work / "meta.json").read_text())
+    assert meta["video_file"] == "source_video.webm"
+    assert meta["thumbnail"] == "source_video.jpg"
+
+def test_outputs_are_meta_and_audio(tmp_path):
+    """_require_source in web.py hardcodes these filenames; pin them here."""
+    ctx = make_ctx(tmp_path, youtube_url="https://yt/x")
+    outputs = acquire.acquire_stage().outputs(ctx)
+    assert [p.name for p in outputs] == ["meta.json", "audio.wav"]
+
+def test_requires_exactly_one_source(tmp_path):
+    import pytest
+    from songtrace.pipeline import PipelineError
+    ctx = make_ctx(tmp_path)
+    with pytest.raises(PipelineError):
+        acquire.acquire_stage().run(ctx)
+
+def test_classify_ordinary_webm_and_jpg(tmp_path):
+    (tmp_path / "source_video.webm").write_bytes(b"v")
+    (tmp_path / "source_video.jpg").write_bytes(b"t")
+    video, thumb = acquire._classify_downloads(tmp_path)
+    assert video.name == "source_video.webm"
+    assert thumb.name == "source_video.jpg"
+
+def test_classify_unexpected_thumbnail_extension(tmp_path):
+    (tmp_path / "source_video.webm").write_bytes(b"v")
+    (tmp_path / "source_video.avif").write_bytes(b"t")
+    video, thumb = acquire._classify_downloads(tmp_path)
+    assert video.name == "source_video.webm"
+    assert thumb.name == "source_video.avif"
+
+def test_classify_no_thumbnail(tmp_path):
+    (tmp_path / "source_video.mp4").write_bytes(b"v")
+    video, thumb = acquire._classify_downloads(tmp_path)
+    assert video.name == "source_video.mp4"
+    assert thumb is None
+
+def test_classify_no_video_raises(tmp_path):
+    import pytest
+    from songtrace.pipeline import PipelineError
+    (tmp_path / "source_video.jpg").write_bytes(b"t")
+    with pytest.raises(PipelineError):
+        acquire._classify_downloads(tmp_path)
+
+def test_classify_picks_largest_video_when_multiple(tmp_path):
+    (tmp_path / "source_video.f137.mp4").write_bytes(b"small")
+    (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_opts_default_client_adds_no_player_client(tmp_path):
+    cfg = Config(whisper_url="http://x/v1")
+    opts = acquire._ydl_opts(cfg, tmp_path, "default", print)
+    assert "youtube" not in opts["extractor_args"]
+    assert opts["retries"] and opts["fragment_retries"] and opts["extractor_retries"]
+
+def test_ydl_opts_named_client_and_pot_provider(tmp_path):
+    cfg = Config(whisper_url="http://x/v1",
+                 ytdlp_pot_provider_url="http://pot:4416",
+                 ytdlp_cookies=tmp_path / "cookies.txt")
+    opts = acquire._ydl_opts(cfg, tmp_path, "tv", print)
+    assert opts["extractor_args"]["youtube"] == {"player_client": ["tv"]}
+    assert opts["extractor_args"]["youtubepot-bgutilhttp"] == {
+        "base_url": ["http://pot:4416"]}
+    assert opts["cookiefile"] == str(tmp_path / "cookies.txt")
+
+def _fake_yt_dlp(monkeypatch, behaviour):
+    """Install a stub yt_dlp whose download calls `behaviour(opts)`."""
+    import sys
+    import types
+    seen = []
+
+    class FakeYoutubeDL:
+        def __init__(self, opts):
+            self.opts = opts
+        def __enter__(self):
+            return self
+        def __exit__(self, *exc):
+            return False
+        def extract_info(self, url, download):
+            seen.append(self.opts["extractor_args"].get("youtube"))
+            behaviour(self.opts)
+
+    mod = types.ModuleType("yt_dlp")
+    mod.YoutubeDL = FakeYoutubeDL
+    mod.version = types.SimpleNamespace(__version__="2026.08.19")
+    monkeypatch.setitem(sys.modules, "yt_dlp", mod)
+    return seen
+
+def test_download_falls_through_to_the_next_client(tmp_path, monkeypatch):
+    def behaviour(opts):
+        client = opts["extractor_args"].get("youtube")
+        if client is None:
+            raise RuntimeError("HTTP Error 403: Forbidden")
+        (tmp_path / "source_video.webm").write_bytes(b"v")
+    seen = _fake_yt_dlp(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 == [None, {"player_client": ["tv"]}]
+    assert any("403" in m for m in logged)
+
+def test_download_reports_version_and_clients_when_all_fail(tmp_path, monkeypatch):
+    import pytest
+    from songtrace.pipeline import PipelineError
+    def behaviour(opts):
+        raise RuntimeError("HTTP Error 403: Forbidden")
+    _fake_yt_dlp(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 "2026.08.19" in str(e.value)
+    assert "default, tv" in str(e.value)
+    assert "403" in str(e.value)
+
+def test_download_with_no_configured_clients_still_tries_default(tmp_path, monkeypatch):
+    seen = _fake_yt_dlp(
+        monkeypatch, lambda opts: (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 == [None]
+
+def test_ydl_log_drops_debug_lines():
+    logged = []
+    log = acquire._YdlLog(logged.append)
+    log.debug("[debug] internal detail")
+    log.debug("[youtube] Extracting URL")
+    log.warning("nsig extraction failed")
+    log.error("HTTP Error 403: Forbidden")
+    assert logged == ["[youtube] Extracting URL", "nsig extraction failed",
+                      "HTTP Error 403: Forbidden"]

tests/test_align.py

@@ -0,0 +1,162 @@
+import json
+
+import pytest
+
+from songtrace.config import Config
+from songtrace.lyrics import normalize_lyrics
+from songtrace.pipeline import Ctx, PipelineError, SongInput
+from songtrace.stages import align
+
+LYRICS = "she's got a smile\nthat it seems to me"
+
+
+def make_ctx(tmp_path, lyrics=LYRICS, transcript=None):
+    work = tmp_path / "w"
+    work.mkdir(exist_ok=True)
+    if transcript is None:
+        transcript = [{"start": 1.0, "end": 4.0,
+                       "text": "she's got a smile that it seems to me"}]
+    (work / "transcript.json").write_text(json.dumps(transcript))
+    inp = SongInput(title="T", artist="A", lyrics_text=lyrics)
+    return Ctx(inp=inp, cfg=Config(whisper_url="u", device="cpu"),
+               work=work, log=lambda m: None)
+
+
+def fake_align(vocals_path, start, end, tokens, device):
+    step = (end - start) / len(tokens)
+    return [(start + i * step, start + (i + 1) * step, 0.9)
+            for i in range(len(tokens))]
+
+
+def run_align(tmp_path, monkeypatch, lyrics=LYRICS, transcript=None):
+    monkeypatch.setattr(align, "force_align_section", fake_align)
+    ctx = make_ctx(tmp_path, lyrics, transcript)
+    align.align_stage().run(ctx)
+    return json.loads((ctx.work / "alignment.json").read_text())
+
+
+def check_invariants(rows, lyrics):
+    ref = normalize_lyrics(lyrics)
+    assert [r["word"] for r in rows] == ref.words
+    assert [r["line_index"] for r in rows] == ref.word_line_index
+    prev = 0.0
+    for r in rows:
+        assert r["start"] is not None and r["end"] is not None
+        assert prev <= r["start"] <= r["end"]
+        prev = r["end"]
+
+
+def test_alignment_covers_all_words(tmp_path, monkeypatch):
+    rows = run_align(tmp_path, monkeypatch)
+    assert [r["word"] for r in rows] == \
+        ["she's", "got", "a", "smile", "that", "it", "seems", "to", "me"]
+    assert rows[0]["line_index"] == 0 and rows[-1]["line_index"] == 1
+    starts = [r["start"] for r in rows]
+    assert starts == sorted(starts)
+
+
+def test_mismatched_lyrics_fail(tmp_path, monkeypatch):
+    monkeypatch.setattr(align, "force_align_section", fake_align)
+    ctx = make_ctx(tmp_path, lyrics="totally unrelated nonsense words banana")
+    with pytest.raises(PipelineError, match="similarity"):
+        align.align_stage().run(ctx)
+
+
+def test_unalignable_word_interpolated(tmp_path, monkeypatch):
+    lyrics = "she's got ... smile\nthat it seems to me"
+    rows = run_align(tmp_path, monkeypatch, lyrics)
+    dots = next(r for r in rows if r["word"] == "...")
+    assert dots["score"] == 0.0
+    prev = next(r for r in rows if r["word"] == "got")
+    nxt = next(r for r in rows if r["word"] == "smile")
+    assert prev["end"] <= dots["start"] <= dots["end"] <= nxt["start"] + 1e-6
+    check_invariants(rows, lyrics)
+
+
+def test_leading_unalignable_word(tmp_path, monkeypatch):
+    lyrics = "... got a smile\nthat it seems to me"
+    rows = run_align(tmp_path, monkeypatch, lyrics)
+    assert rows[0]["word"] == "..." and rows[0]["score"] == 0.0
+    assert rows[0]["start"] >= 0.0
+    assert rows[0]["end"] <= rows[1]["start"] + 1e-6
+    check_invariants(rows, lyrics)
+
+
+def test_trailing_unalignable_word(tmp_path, monkeypatch):
+    lyrics = "she's got a smile\nthat it seems to ..."
+    rows = run_align(tmp_path, monkeypatch, lyrics)
+    assert rows[-1]["word"] == "..." and rows[-1]["score"] == 0.0
+    assert rows[-2]["end"] <= rows[-1]["start"]
+    check_invariants(rows, lyrics)
+
+
+def test_consecutive_unalignable_words(tmp_path, monkeypatch):
+    lyrics = "she's got ... ... smile\nthat it seems to me"
+    rows = run_align(tmp_path, monkeypatch, lyrics)
+    dots = [r for r in rows if r["word"] == "..."]
+    assert len(dots) == 2
+    assert all(r["score"] == 0.0 for r in dots)
+    assert dots[0]["end"] <= dots[1]["start"]
+    check_invariants(rows, lyrics)
+
+
+def test_all_words_unalignable_in_section(tmp_path, monkeypatch):
+    lyrics = "... ...\n... ..."
+    transcript = [{"start": 1.0, "end": 4.0, "text": "..."}]
+    monkeypatch.setattr(align, "force_align_section", fake_align)
+    ctx = make_ctx(tmp_path, lyrics, transcript)
+    align.align_stage().run(ctx)
+    rows = json.loads((ctx.work / "alignment.json").read_text())
+    check_invariants(rows, lyrics)
+    assert all(r["score"] == 0.0 for r in rows)
+    assert rows[-1]["end"] <= 5.0
+
+
+def test_non_chronological_transcript(tmp_path, monkeypatch):
+    transcript = [{"start": 30.0, "end": 33.0, "text": "that it seems to me"},
+                  {"start": 1.0, "end": 4.0, "text": "she's got a smile"}]
+    rows = run_align(tmp_path, monkeypatch, transcript=transcript)
+    check_invariants(rows, LYRICS)
+    assert len({r["start"] for r in rows}) == len(rows)
+    assert rows[3]["end"] <= 5.0
+    assert rows[4]["start"] >= 29.0
+
+
+def test_inverted_section_does_not_collapse(tmp_path, monkeypatch):
+    monkeypatch.setattr(align, "vocal_sections", lambda segments: [(29.0, 5.0)])
+    rows = run_align(tmp_path, monkeypatch)
+    check_invariants(rows, LYRICS)
+    assert len({r["start"] for r in rows}) == len(rows)
+    assert rows[0]["start"] >= 5.0 and rows[-1]["end"] <= 29.0
+
+
+def test_words_split_across_sections(tmp_path, monkeypatch):
+    transcript = [{"start": 1.0, "end": 4.0, "text": "she's got a smile"},
+                  {"start": 30.0, "end": 33.0, "text": "that it seems to me"}]
+    rows = run_align(tmp_path, monkeypatch, transcript=transcript)
+    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 = "she's got a smile\nthat it seems to me\nbut then it fades away"
+    transcript = [{"start": 1.0, "end": 4.0,
+                   "text": "she's got a smile that it seems to me"}]
+    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 = "she's got a smile\nthat it seems to me"
+    transcript = [
+        {"start": 1.0, "end": 4.0, "text": "she's got a smile that it seems to me"},
+        {"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]]

tests/test_audio_io.py

@@ -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)

tests/test_cli.py

@@ -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

tests/test_cli_make.py

@@ -0,0 +1,116 @@
+import os
+
+import pytest
+
+import songtrace.cli as cli
+from songtrace.config import Config
+from songtrace.pipeline import PipelineError, PipelineResult
+
+
+def setup_cfg(monkeypatch, tmp_path):
+    monkeypatch.setattr(cli, "load_config",
+                        lambda path=None: Config(whisper_url="u"))
+
+def test_make_success(tmp_path, monkeypatch, capsys):
+    setup_cfg(monkeypatch, tmp_path)
+    lyrics = tmp_path / "l.txt"
+    lyrics.write_text("la la")
+    captured = {}
+    def fake_run(inp, cfg, log=print, force_from=None, stages=None):
+        captured["inp"] = inp
+        captured["force_from"] = force_from
+        return PipelineResult(output_dir=tmp_path / "out", warnings=["w1"])
+    monkeypatch.setattr(cli, "run_pipeline", fake_run)
+    rc = cli.main(["make", "--title", "T", "--artist", "A",
+                   "--lyrics", str(lyrics), "--youtube", "http://y",
+                   "--force-from", "align"])
+    assert rc == 0
+    assert captured["inp"].lyrics_text == "la la"
+    assert captured["force_from"] == "align"
+    out = capsys.readouterr().out
+    assert "w1" in out and str(tmp_path / "out") in out
+
+def test_make_pipeline_error(tmp_path, monkeypatch, capsys):
+    setup_cfg(monkeypatch, tmp_path)
+    lyrics = tmp_path / "l.txt"
+    lyrics.write_text("x")
+    def boom(*a, **kw):
+        raise PipelineError("lyrics don't match")
+    monkeypatch.setattr(cli, "run_pipeline", boom)
+    rc = cli.main(["make", "--title", "T", "--artist", "A",
+                   "--lyrics", str(lyrics), "--youtube", "http://y"])
+    assert rc == 1
+    assert "lyrics don't match" in capsys.readouterr().err
+
+def test_make_requires_one_source(tmp_path, monkeypatch, capsys):
+    setup_cfg(monkeypatch, tmp_path)
+    lyrics = tmp_path / "l.txt"
+    lyrics.write_text("x")
+    rc = cli.main(["make", "--title", "T", "--artist", "A",
+                   "--lyrics", str(lyrics)])
+    assert rc != 0
+
+def test_make_rejects_zero_bpm(tmp_path, monkeypatch, capsys):
+    setup_cfg(monkeypatch, tmp_path)
+    lyrics = tmp_path / "l.txt"
+    lyrics.write_text("x")
+    try:
+        cli.main(["make", "--title", "T", "--artist", "A",
+                  "--lyrics", str(lyrics), "--youtube", "http://y",
+                  "--bpm", "0"])
+        assert False, "expected SystemExit"
+    except SystemExit as e:
+        assert e.code != 0
+    assert "--bpm" in capsys.readouterr().err
+
+def test_make_rejects_negative_bpm(tmp_path, monkeypatch, capsys):
+    setup_cfg(monkeypatch, tmp_path)
+    lyrics = tmp_path / "l.txt"
+    lyrics.write_text("x")
+    try:
+        cli.main(["make", "--title", "T", "--artist", "A",
+                  "--lyrics", str(lyrics), "--youtube", "http://y",
+                  "--bpm", "-120"])
+        assert False, "expected SystemExit"
+    except SystemExit as e:
+        assert e.code != 0
+    assert "--bpm" in capsys.readouterr().err
+
+def test_make_reports_missing_lyrics_file(tmp_path, monkeypatch, capsys):
+    setup_cfg(monkeypatch, tmp_path)
+    def boom(*a, **kw):
+        raise AssertionError("run_pipeline must not run with an unreadable lyrics file")
+    monkeypatch.setattr(cli, "run_pipeline", boom)
+    rc = cli.main(["make", "--title", "T", "--artist", "A",
+                   "--lyrics", str(tmp_path / "missing.txt"), "--youtube", "http://y"])
+    assert rc != 0
+    assert "--lyrics" in capsys.readouterr().err
+
+@pytest.mark.skipif(os.geteuid() == 0, reason="root ignores file permissions")
+def test_make_reports_unreadable_lyrics_file(tmp_path, monkeypatch, capsys):
+    setup_cfg(monkeypatch, tmp_path)
+    lyrics = tmp_path / "l.txt"
+    lyrics.write_text("x")
+    lyrics.chmod(0o000)
+    def boom(*a, **kw):
+        raise AssertionError("run_pipeline must not run with an unreadable lyrics file")
+    monkeypatch.setattr(cli, "run_pipeline", boom)
+    try:
+        rc = cli.main(["make", "--title", "T", "--artist", "A",
+                       "--lyrics", str(lyrics), "--youtube", "http://y"])
+    finally:
+        lyrics.chmod(0o644)
+    assert rc != 0
+    assert "--lyrics" in capsys.readouterr().err
+
+def test_make_reports_undecodable_lyrics_file(tmp_path, monkeypatch, capsys):
+    setup_cfg(monkeypatch, tmp_path)
+    lyrics = tmp_path / "l.txt"
+    lyrics.write_bytes(b"\xff\xfe\x00bad")
+    def boom(*a, **kw):
+        raise AssertionError("run_pipeline must not run with undecodable lyrics")
+    monkeypatch.setattr(cli, "run_pipeline", boom)
+    rc = cli.main(["make", "--title", "T", "--artist", "A",
+                   "--lyrics", str(lyrics), "--youtube", "http://y"])
+    assert rc != 0
+    assert "--lyrics" in capsys.readouterr().err

tests/test_config.py

@@ -0,0 +1,40 @@
+import pytest
+from pathlib import Path
+from songtrace.config import Config, load_config
+
+
+def test_load_minimal(tmp_path):
+    cfg_file = tmp_path / "config.toml"
+    cfg_file.write_text('whisper_url = "http://w:9000/v1"\n')
+    cfg = load_config(cfg_file)
+    assert cfg.whisper_url == "http://w:9000/v1"
+    assert cfg.demucs_model == "htdemucs_ft"
+
+
+def test_paths_and_overrides(tmp_path):
+    cfg_file = tmp_path / "config.toml"
+    cfg_file.write_text(
+        'whisper_url = "http://w/v1"\noutput_dir = "/srv/songs"\nport = 9999\n'
+    )
+    cfg = load_config(cfg_file)
+    assert cfg.output_dir == Path("/srv/songs")
+    assert cfg.port == 9999
+
+
+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("SONGTRACE_CONFIG", str(cfg_file))
+    assert load_config().whisper_url == "http://env/v1"
+
+
+def test_unknown_key_rejected(tmp_path):
+    cfg_file = tmp_path / "c.toml"
+    cfg_file.write_text('whisper_url = "x"\nwhipser_model = "typo"\n')
+    with pytest.raises(ValueError, match="whipser_model"):
+        load_config(cfg_file)
+
+
+def test_missing_file(tmp_path):
+    with pytest.raises(FileNotFoundError):
+        load_config(tmp_path / "nope.toml")

tests/test_e2e.py

@@ -0,0 +1,30 @@
+import os
+import shutil
+
+import pytest
+
+requires_full = pytest.mark.skipif(
+    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 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["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 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

tests/test_grid.py

@@ -0,0 +1,259 @@
+import json
+
+import numpy as np
+import pytest
+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,
+)
+
+
+def note(start, duration, pitch=0, text="la"):
+    return {"start": start, "duration": duration, "pitch": pitch, "text": text,
+            "low_confidence": False}
+
+
+def assert_ordered(body):
+    """Beats never go backwards and no two notes share a beat."""
+    cursor = None
+    for item in body:
+        if cursor is not None:
+            assert item.beat >= cursor
+        cursor = item.beat + (0 if isinstance(item, LineBreak) else item.length)
+    notes = [i for i in body if isinstance(i, Note)]
+    for a, b in zip(notes, notes[1:]):
+        assert a.beat + a.length <= b.beat
+    assert all(n.length >= 1 for n in notes)
+
+
+def test_scale_bpm():
+    assert scale_bpm(128.0) == 256.0
+    assert scale_bpm(64.0) == 256.0
+    assert scale_bpm(250.0) == 250.0
+    assert scale_bpm(450.0) == 225.0
+
+
+@pytest.mark.parametrize("bpm", [0.0, -120.0, float("inf"), float("nan")])
+def test_scale_bpm_rejects_unusable_tempo(bpm):
+    with pytest.raises(ValueError):
+        scale_bpm(bpm)
+
+
+def test_to_ultrastar_rejects_empty_notes():
+    with pytest.raises(ValueError):
+        to_ultrastar([], [], 240.0, [])
+
+
+def test_to_ultrastar_conversion():
+    notes = [note(2.0, 0.5, 3, "la"), note(2.6, 0.25, 5, " la")]
+    song = to_ultrastar(notes, breaks=[2.55], file_bpm=240.0,
+                        headers=[("TITLE", "T")])
+    # spb = 60/(4*240) = 0.0625s; gap = 2000ms
+    assert get_header(song, "GAP") == "2000"
+    assert get_header(song, "BPM") == "240"
+    assert song.headers[0] == ("TITLE", "T")
+    n0, br, n1 = song.body
+    assert n0 == Note(kind=":", beat=0, length=8, pitch=3, text="la")
+    assert br == LineBreak(beat=9)          # (2.55-2.0)/0.0625 = 8.8 -> 9
+    assert n1.beat == 10 and n1.length == 4
+
+
+def test_min_length_one():
+    song = to_ultrastar([note(1.0, 0.001, 0, "x")], [], 240.0, [])
+    assert song.body[0].length == 1
+
+
+def test_fractional_bpm_header():
+    song = to_ultrastar([note(0.0, 1.0)], [], 232.5, [])
+    assert get_header(song, "BPM") == "232,50"
+
+
+def test_beat_cursor_separates_notes_quantised_onto_one_beat():
+    notes = [note(1.0, 0.01, 0, "a"), note(1.01, 0.01, 1, "b"),
+             note(1.02, 0.01, 2, "c")]
+    song = to_ultrastar(notes, [], 240.0, [])   # all three round to beat 0
+    assert [n.beat for n in song.body] == [0, 1, 2]
+    assert [n.length for n in song.body] == [1, 1, 1]
+    assert_ordered(song.body)
+
+
+def test_beat_cursor_preserves_onsets_that_do_not_collide():
+    notes = [note(0.0, 0.5), note(1.0, 0.5), note(2.0, 0.5)]
+    song = to_ultrastar(notes, [], 240.0, [])
+    assert [n.beat for n in song.body] == [0, 16, 32]
+
+
+def test_line_break_pushed_past_the_preceding_note():
+    notes = [note(0.0, 0.56), note(0.57, 0.1)]
+    song = to_ultrastar(notes, [0.51], 240.0, [])   # break rounds to beat 8
+    n0, br, n1 = song.body
+    assert n0.beat + n0.length == 9
+    assert br == LineBreak(beat=9)
+    assert n1.beat >= br.beat
+    assert_ordered(song.body)
+
+
+def test_line_break_on_a_note_start_precedes_that_note():
+    """No gap across a line boundary puts the break exactly on the next onset."""
+    notes = [note(0.0, 0.5, 0, "end"), note(0.5, 0.5, 2, "New")]
+    song = to_ultrastar(notes, [0.5], 240.0, [])
+    assert song.body[1] == LineBreak(beat=8)
+    assert [n.text for n in song.body if isinstance(n, Note)] == ["end", "New"]
+    assert_ordered(song.body)
+
+
+def test_line_break_not_after_the_following_note():
+    notes = [note(0.0, 0.02, 0, "a"), note(0.03, 0.02, 1, "b")]
+    song = to_ultrastar(notes, [0.025], 240.0, [])
+    br = song.body[1]
+    assert isinstance(br, LineBreak)
+    assert br.beat <= song.body[2].beat
+    assert_ordered(song.body)
+
+
+def sung_words(rng):
+    """Words cut into equal syllable slices, the way build_notes emits them."""
+    notes, breaks, t = [], [], 1.7
+    for i in range(60):
+        word = float(rng.uniform(0.05, 0.6))
+        sylls = int(rng.integers(1, 4))
+        for k in range(sylls):
+            notes.append(note(t + k * word / sylls, word / sylls,
+                              int(rng.integers(-12, 12))))
+        t += word + float(rng.uniform(0.0, 0.4))
+        if i % 5 == 4:
+            breaks.append(t - 0.05)
+    return notes, breaks
+
+
+@pytest.mark.parametrize("bpm", [200.0, 300.0, 399.0])
+def test_no_overlap_on_realistic_input(bpm):
+    rng = np.random.default_rng(7)
+    moved = 0
+    for _ in range(20):
+        notes, breaks = sung_words(rng)
+        song = to_ultrastar(notes, breaks, bpm, [])
+        assert len(song.body) == len(notes) + len(breaks)
+        assert song.body[0].beat == 0
+        assert_ordered(song.body)
+        spb = 60 / (4 * bpm)
+        emitted = [n for n in song.body if isinstance(n, Note)]
+        moved += sum(1 for n, src in zip(emitted, notes)
+                     if n.beat != round((src["start"] - notes[0]["start"]) / spb))
+    assert moved > 0        # otherwise the fuzz never exercises the beat cursor
+
+
+def write_work(work, video="source_video.webm"):
+    work.mkdir()
+    (work / "notes_sec.json").write_text(json.dumps({
+        "notes": [note(1.0, 0.5, 0, "la"), note(2.0, 0.5, 2, "la")],
+        "breaks": [1.75],
+        "warnings": [],
+    }))
+    (work / "meta.json").write_text(json.dumps(
+        {"video_file": video, "thumbnail": "source_video.jpg"}))
+
+
+def make_ctx(work, **kwargs):
+    inp = SongInput(title="Song", artist="Band", lyrics_text="x", **kwargs)
+    return Ctx(inp=inp, cfg=Config(whisper_url="u", creator="me"), work=work,
+               log=lambda m: None)
+
+
+def test_grid_stage_writes_song_txt(tmp_path, monkeypatch):
+    work = tmp_path / "w"
+    write_work(work)
+    monkeypatch.setattr(grid_stage_mod.grid, "detect_bpm", lambda p: 120.0)
+    ctx = make_ctx(work, language="English", genre="Pop", year=1999)
+    stage = grid_stage_mod.grid_stage()
+    assert stage.outputs(ctx) == [work / "song.txt"]
+    stage.run(ctx)
+    song = parse_ultrastar((work / "song.txt").read_text(encoding="utf-8"))
+    assert [k for k, _ in song.headers] == [
+        "TITLE", "ARTIST", "LANGUAGE", "GENRE", "YEAR", "CREATOR",
+        "MP3", "COVER", "VIDEO", "BPM", "GAP"]
+    assert get_header(song, "BPM") == "240"        # 120 doubled into range
+    assert get_header(song, "GAP") == "1000"
+    assert get_header(song, "MP3") == "Band - Song.mp3"
+    assert get_header(song, "COVER") == "Band - Song.jpg"
+    assert get_header(song, "VIDEO") == "Band - Song.webm"
+    assert [n.text for n in song.body if isinstance(n, Note)] == ["la", "la"]
+    assert_ordered(song.body)
+
+
+def test_grid_stage_omits_optional_headers(tmp_path, monkeypatch):
+    work = tmp_path / "w"
+    write_work(work, video="source_video.mp4")
+    monkeypatch.setattr(grid_stage_mod.grid, "detect_bpm", lambda p: 210.0)
+    ctx = make_ctx(work)
+    grid_stage_mod.grid_stage().run(ctx)
+    song = parse_ultrastar((work / "song.txt").read_text(encoding="utf-8"))
+    assert [k for k, _ in song.headers] == [
+        "TITLE", "ARTIST", "CREATOR", "MP3", "COVER", "VIDEO", "BPM", "GAP"]
+    assert get_header(song, "VIDEO") == "Band - Song.mp4"
+    assert get_header(song, "CREATOR") == "me"
+
+
+def test_grid_stage_uses_bpm_override(tmp_path, monkeypatch):
+    work = tmp_path / "w"
+    write_work(work)
+
+    def boom(path):
+        raise AssertionError("detect_bpm must not run with an override")
+
+    monkeypatch.setattr(grid_stage_mod.grid, "detect_bpm", boom)
+    ctx = make_ctx(work, bpm_override=333.0)
+    grid_stage_mod.grid_stage().run(ctx)
+    song = parse_ultrastar((work / "song.txt").read_text(encoding="utf-8"))
+    assert get_header(song, "BPM") == "333"
+
+
+def test_grid_stage_honours_zero_bpm_override(tmp_path, monkeypatch):
+    """bpm_override=0.0 must not be mistaken for 'not provided' (falsy-zero bug)."""
+    work = tmp_path / "w"
+    write_work(work)
+
+    def boom(path):
+        raise AssertionError("detect_bpm must not run with an explicit override")
+
+    monkeypatch.setattr(grid_stage_mod.grid, "detect_bpm", boom)
+    captured = {}
+
+    def fake_to_ultrastar(notes, breaks, file_bpm, headers):
+        captured["bpm"] = file_bpm
+        return to_ultrastar(notes, breaks, 240.0, headers)
+
+    monkeypatch.setattr(grid_stage_mod.grid, "to_ultrastar", fake_to_ultrastar)
+    ctx = make_ctx(work, bpm_override=0.0)
+    grid_stage_mod.grid_stage().run(ctx)
+    assert captured["bpm"] == 0.0
+
+
+def test_grid_stage_reports_unusable_tempo_as_pipeline_error(tmp_path, monkeypatch):
+    work = tmp_path / "w"
+    write_work(work)
+    monkeypatch.setattr(grid_stage_mod.grid, "detect_bpm", lambda p: 0.0)
+    ctx = make_ctx(work)
+    with pytest.raises(PipelineError, match="tempo"):
+        grid_stage_mod.grid_stage().run(ctx)
+
+
+def test_grid_stage_reports_no_notes_as_pipeline_error(tmp_path, monkeypatch):
+    work = tmp_path / "w"
+    work.mkdir()
+    (work / "notes_sec.json").write_text(json.dumps(
+        {"notes": [], "breaks": [], "warnings": []}))
+    (work / "meta.json").write_text(json.dumps(
+        {"video_file": "source_video.webm", "thumbnail": "source_video.jpg"}))
+    monkeypatch.setattr(grid_stage_mod.grid, "detect_bpm", lambda p: 120.0)
+    ctx = make_ctx(work)
+    with pytest.raises(PipelineError, match="no notes"):
+        grid_stage_mod.grid_stage().run(ctx)
+
+
+def test_grid_stage_registered():
+    names = [s.name for s in all_stages()]
+    assert names[-1] == "package" and names[-2] == "grid"

tests/test_lyrics.py

@@ -0,0 +1,40 @@
+from songtrace.lyrics import normalize_lyrics, alignment_token
+
+RAW = """[Verse 1]
+She's got a smile
+
+[Chorus]
+Whoa, sweet child o' mine
+Take me down to 7 streets
+"""
+
+
+def test_normalize_strips_sections_and_blanks():
+    ref = normalize_lyrics(RAW)
+    assert ref.lines == [
+        ["She's", "got", "a", "smile"],
+        ["Whoa,", "sweet", "child", "o'", "mine"],
+        ["Take", "me", "down", "to", "7", "streets"],
+    ]
+    assert ref.words[0] == "She's"
+    assert ref.word_line_index == [0, 0, 0, 0, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2]
+
+
+def test_alignment_token():
+    assert alignment_token("Whoa,") == "whoa"
+    assert alignment_token("o'") == "o'"
+    assert alignment_token("7") == "seven"
+    assert alignment_token("...") == ""
+
+
+def test_alignment_token_no_internal_whitespace():
+    assert alignment_token("21") == "twentyone"
+    assert alignment_token("7th") == "seventh"
+
+
+def test_alignment_token_hyphenated_word():
+    assert alignment_token("well-known") == "wellknown"
+
+
+def test_alignment_token_curly_apostrophe():
+    assert alignment_token("don’t") == "don't"

tests/test_notes.py

@@ -0,0 +1,349 @@
+import json
+
+import numpy as np
+import pytest
+from songtrace.config import Config
+from songtrace.notes import (
+    MIN_NOTE_SEC, TimedNote, build_notes, hz_to_ultrastar, segment_span,
+    syllabify,
+)
+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
+    times = np.arange(0, duration, 0.01)
+    return times, np.full_like(times, hz), np.full_like(times, 0.9)
+
+
+def assert_no_overlap(notes):
+    for a, b in zip(notes, notes[1:]):
+        assert a.start < b.start
+        assert a.start + a.duration <= b.start
+    assert all(n.duration > 0 for n in notes)
+
+
+def test_hz_to_ultrastar():
+    assert hz_to_ultrastar(261.63) == 0        # C4
+    assert hz_to_ultrastar(440.0) == 9         # A4
+    assert hz_to_ultrastar(220.0) == -3        # A3
+
+
+def test_hz_to_ultrastar_rejects_non_positive():
+    with pytest.raises(ValueError):
+        hz_to_ultrastar(0.0)
+    with pytest.raises(ValueError):
+        hz_to_ultrastar(-100.0)
+
+
+def test_segment_span_single_flat():
+    times, f0, conf = flat_pitch()
+    segs = segment_span(times, f0, conf, 1.0, 2.0)
+    assert len(segs) == 1
+    s, e, midi = segs[0]
+    assert abs(midi - 60.0) < 0.1
+    assert abs(s - 1.0) < 0.02 and abs(e - 2.0) < 0.02
+
+
+def test_segment_span_splits_on_jump():
+    times = np.arange(0, 2.0, 0.01)
+    f0 = np.where(times < 1.0, 261.63, 329.63)   # C4 -> E4
+    conf = np.full_like(times, 0.9)
+    segs = segment_span(times, f0, conf, 0.0, 2.0)
+    assert len(segs) == 2
+    assert abs(segs[0][2] - 60.0) < 0.2 and abs(segs[1][2] - 64.0) < 0.2
+    assert abs(segs[0][1] - segs[1][0]) < 1e-6      # contiguous
+
+
+def test_segment_span_masks_low_conf():
+    times, f0, conf = flat_pitch()
+    conf[:] = 0.1
+    assert segment_span(times, f0, conf, 0.0, 1.0) == []
+
+
+def test_segment_span_ignores_brief_pitch_glitch():
+    times, f0, conf = flat_pitch(duration=1.0)
+    f0[50:52] = 523.25                          # 2-frame octave blip
+    conf = np.full_like(times, 0.9)
+    segs = segment_span(times, f0, conf, 0.0, 1.0)
+    assert len(segs) == 1
+
+
+def test_segment_span_drops_sub_minimum_segments():
+    times, f0, conf = flat_pitch(duration=1.0)
+    assert segment_span(times, f0, conf, 0.5, 0.52) == []
+
+
+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"]
+
+
+def test_syllabify_preserves_word_characters():
+    assert "".join(syllabify("well-known", "English")) == "well-known"
+    assert "".join(syllabify("beautiful", "English")) == "beautiful"
+
+
+def test_syllabify_language_without_pyphen_dict_is_not_mishandled_as_english():
+    # Japanese is a known language (e.g. to whisper_client) but has no pyphen
+    # dictionary; it must not silently fall back to English hyphenation rules.
+    assert syllabify("beautiful", "Japanese") == ["beautiful"]
+    assert syllabify("beautiful", "English") == ["beau", "ti", "ful"]
+
+
+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": "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[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
+    assert warnings == []
+
+
+def test_build_notes_melisma_continuation():
+    times = np.arange(0, 3.0, 0.01)
+    f0 = np.where(times < 1.0, 261.63, 329.63)
+    conf = np.full_like(times, 0.9)
+    rows = [{"word": "mine", "start": 0.2, "end": 2.8, "score": 0.9, "line_index": 0}]
+    notes, _, _ = build_notes(rows, times, f0, conf, "English")
+    assert [n.text for n in notes] == ["mine", "~"]
+    assert notes[0].pitch != notes[1].pitch
+
+
+def test_build_notes_unvoiced_word_flagged():
+    times, f0, conf = flat_pitch()
+    conf[:] = 0.0
+    rows = [{"word": "hey", "start": 1.0, "end": 1.4, "score": 0.9, "line_index": 0}]
+    notes, _, warnings = build_notes(rows, times, f0, conf, "English")
+    assert notes[0].low_confidence and len(warnings) == 1
+    assert warnings[0] == "1 low-confidence note near 0:01–0:01"
+
+
+def test_build_notes_low_score_word_flagged():
+    times, f0, conf = flat_pitch()
+    rows = [{"word": "hey", "start": 1.0, "end": 1.4, "score": 0.1, "line_index": 0}]
+    notes, _, warnings = build_notes(rows, times, f0, conf, "English")
+    assert all(n.low_confidence for n in notes) and len(warnings) == 1
+
+
+def test_build_notes_unvoiced_word_inherits_previous_pitch():
+    times, f0, conf = flat_pitch(hz=440.0)      # A4 -> 9, not the fallback 0
+    conf[300:] = 0.0
+    rows = [
+        {"word": "one", "start": 1.0, "end": 2.0, "score": 0.9, "line_index": 0},
+        {"word": "two", "start": 4.0, "end": 4.5, "score": 0.9, "line_index": 0},
+    ]
+    notes, _, _ = build_notes(rows, times, f0, conf, "English")
+    assert notes[1].pitch == notes[0].pitch == 9
+    assert notes[1].low_confidence and not notes[0].low_confidence
+
+
+def test_build_notes_unvoiced_first_word_uses_zero_pitch():
+    times, f0, conf = flat_pitch()
+    conf[:] = 0.0
+    rows = [{"word": "hey", "start": 1.0, "end": 1.4, "score": 0.9, "line_index": 0}]
+    notes, _, _ = build_notes(rows, times, f0, conf, "English")
+    assert notes[0].pitch == 0
+
+
+def test_build_notes_multisyllable_word_emits_syllables_only():
+    times, f0, conf = flat_pitch()
+    rows = [
+        {"word": "reminds", "start": 1.0, "end": 1.8, "score": 0.9, "line_index": 0},
+        {"word": "reminds", "start": 2.0, "end": 2.8, "score": 0.9, "line_index": 0},
+    ]
+    notes, _, _ = build_notes(rows, times, f0, conf, "English")
+    assert [n.text for n in notes] == ["re", "minds", " re", "minds"]
+
+
+def test_build_notes_more_syllables_than_segments():
+    times, f0, conf = flat_pitch()
+    rows = [{"word": "beautiful", "start": 1.0, "end": 1.9, "score": 0.9,
+             "line_index": 0}]
+    notes, _, _ = build_notes(rows, times, f0, conf, "English")
+    assert [n.text for n in notes] == ["beau", "ti", "ful"]
+    assert all(abs(n.duration - notes[0].duration) < 1e-6 for n in notes)
+    for a, b in zip(notes, notes[1:]):
+        assert abs(a.start + a.duration - b.start) < 1e-6
+
+
+def test_build_notes_subdivides_the_widest_segment_first():
+    times = np.arange(0, 2.0, 0.01)
+    f0 = np.where(times < 1.1, 329.63, 261.63)   # 0.1s E4 then 0.9s C4 in-word
+    conf = np.full_like(times, 0.9)
+    rows = [{"word": "beautiful", "start": 1.0, "end": 2.0, "score": 0.9,
+             "line_index": 0}]
+    notes, _, _ = build_notes(rows, times, f0, conf, "English")
+    assert [n.text for n in notes] == ["beau", "ti", "ful"]
+    assert [n.pitch for n in notes] == [4, 0, 0]      # the 0.9s segment split
+    assert abs(notes[0].duration - 0.1) < 1e-6
+    assert abs(notes[1].duration - notes[2].duration) < 1e-6
+
+
+def test_build_notes_zero_duration_row_gets_minimum_length():
+    times, f0, conf = flat_pitch()
+    rows = [
+        {"word": "...", "start": 1.0, "end": 1.0, "score": 0.0, "line_index": 0},
+        {"word": "go", "start": 1.0, "end": 1.5, "score": 0.9, "line_index": 0},
+    ]
+    notes, _, _ = build_notes(rows, times, f0, conf, "English")
+    assert notes[0].duration >= MIN_NOTE_SEC
+    assert_no_overlap(notes)
+
+
+def test_build_notes_repeated_zero_duration_rows_stay_separate():
+    times, f0, conf = flat_pitch()
+    rows = [{"word": w, "start": 1.0, "end": 1.0, "score": 0.0, "line_index": 0}
+            for w in ("...", "...", "...", "...")]
+    rows.append({"word": "go", "start": 1.0, "end": 1.5, "score": 0.9,
+                 "line_index": 0})
+    notes, _, _ = build_notes(rows, times, f0, conf, "English")
+    assert len(notes) == 5
+    assert_no_overlap(notes)
+
+
+def test_build_notes_short_multisyllable_word_does_not_overlap():
+    times, f0, conf = flat_pitch()
+    rows = [
+        {"word": "reminds", "start": 8.92, "end": 8.99, "score": 0.9,
+         "line_index": 0},
+        {"word": "me", "start": 8.99, "end": 9.20, "score": 0.9, "line_index": 0},
+    ]
+    notes, _, _ = build_notes(rows, times, f0, conf, "English")
+    assert [n.text for n in notes] == ["re", "minds", " me"]
+    assert_no_overlap(notes)
+
+
+def test_build_notes_never_overlaps_on_dense_input():
+    rng = np.random.default_rng(11)
+    times = np.arange(0, 40.0, 0.01)
+    words = ["She's", "got", "a", "beautiful", "smile", "reminds", "me",
+             "everything", "I", "lost", "yesterday", "oh"]
+    for _ in range(8):
+        f0 = 220.0 * 2 ** (rng.integers(0, 12, len(times)) // 3 / 12.0)
+        conf = np.where(rng.random(len(times)) < 0.9, 0.9, 0.1)
+        rows, t = [], 0.5
+        for i in range(80):
+            dur = float(rng.uniform(0.06, 0.55))
+            rows.append({"word": words[i % len(words)], "start": t,
+                         "end": t + dur, "score": 0.9, "line_index": i // 6})
+            t += dur + float(rng.uniform(0.0, 0.15))
+        notes, _, _ = build_notes(rows, times, f0, conf, "English")
+        assert_no_overlap(notes)
+
+
+def test_build_notes_starts_are_non_decreasing():
+    times = np.arange(0, 12.0, 0.01)
+    f0 = 220.0 * 2 ** (np.floor(times * 2) % 5 / 12.0)   # stepped melody
+    conf = np.where((times % 3.0) < 2.6, 0.9, 0.1)       # unvoiced gaps
+    words = ["She's", "got", "a", "beautiful", "smile", "that",
+             "reminds", "me", "of", "everything", "I", "lost"]
+    rows = []
+    for i, w in enumerate(words):
+        dur = 0.8 if i % 3 else 0.07       # every third word sung very fast
+        rows.append({"word": w, "start": 0.5 + i * 0.9, "end": 0.5 + i * 0.9 + dur,
+                     "score": 0.9 if i % 5 else 0.2, "line_index": i // 4})
+    notes, breaks, warnings = build_notes(rows, times, f0, conf, "English")
+    assert len(notes) >= len(words)
+    starts = [n.start for n in notes]
+    assert starts == sorted(starts)
+    assert breaks == sorted(breaks) and len(breaks) == 2
+    assert_no_overlap(notes)
+    assert warnings and all("low-confidence note" in w for w in warnings)
+
+
+def test_build_notes_line_initial_after_break_has_no_space():
+    times, f0, conf = flat_pitch()
+    rows = [
+        {"word": "one", "start": 1.0, "end": 1.4, "score": 0.9, "line_index": 0},
+        {"word": "two", "start": 2.0, "end": 2.4, "score": 0.9, "line_index": 1},
+        {"word": "three", "start": 3.0, "end": 3.4, "score": 0.9, "line_index": 1},
+    ]
+    notes, breaks, _ = build_notes(rows, times, f0, conf, "English")
+    assert [n.text for n in notes] == ["one", "two", " three"]
+    assert breaks == [1.7]
+
+
+def test_notes_stage_writes_json(tmp_path):
+    work = tmp_path / "w"
+    work.mkdir()
+    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": "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"),
+              cfg=Config(whisper_url="u"), work=work, log=lambda m: None)
+    stage = notes_stage_mod.notes_stage()
+    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 data["breaks"] == [2.25]
+    assert data["warnings"] == []
+    assert TimedNote(**data["notes"][0]).pitch == 0
+
+
+def test_notes_stage_registered():
+    assert "notes" in [s.name for s in all_stages()]
+
+
+def test_build_notes_skips_notes_within_missing_span():
+    times, f0, conf = flat_pitch()
+    rows = [
+        {"word": "keep", "start": 1.0, "end": 1.4, "score": 0.9, "line_index": 0},
+        {"word": "skip", "start": 5.0, "end": 5.4, "score": 0.9, "line_index": 1},
+    ]
+    notes, _, _ = build_notes(rows, times, f0, conf, "English",
+                              missing_spans=[(4.5, 6.0)])
+    assert [n.text for n in notes] == ["keep"]
+
+
+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_build_notes_skipped_row_mid_line_does_not_corrupt_spacing_or_breaks():
+    # "gone" falls inside a missing span and is skipped; "away" is the same
+    # line's next word and must still get its leading space, and the break
+    # before the next line must use the last *emitted* word's end (3.4, from
+    # "away"), not the skipped row's end (2.4).
+    times, f0, conf = flat_pitch()
+    rows = [
+        {"word": "keep", "start": 1.0, "end": 1.4, "score": 0.9, "line_index": 0},
+        {"word": "gone", "start": 2.0, "end": 2.4, "score": 0.9, "line_index": 0},
+        {"word": "away", "start": 3.0, "end": 3.4, "score": 0.9, "line_index": 0},
+        {"word": "next", "start": 5.0, "end": 5.4, "score": 0.9, "line_index": 1},
+    ]
+    notes, breaks, _ = build_notes(rows, times, f0, conf, "English",
+                                   missing_spans=[(1.9, 2.5)])
+    assert [n.text for n in notes] == ["keep", " away", "next"]
+    assert breaks == [(3.4 + 5.0) / 2]
+
+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

tests/test_package.py

@@ -0,0 +1,187 @@
+import json
+from pathlib import Path
+
+import pytest
+from PIL import Image
+
+from songtrace.config import Config
+from songtrace.pipeline import (
+    STAGE_ORDER, Ctx, PipelineError, SongInput, run_pipeline, song_slug,
+)
+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")
+
+
+def make_ctx(tmp_path, cover=None):
+    work = tmp_path / "work"
+    work.mkdir()
+    (work / "song.txt").write_text(SONG_TXT)
+    (work / "audio.wav").write_bytes(b"w")
+    (work / "source_video.webm").write_bytes(b"v")
+    (work / "meta.json").write_text(json.dumps(
+        {"video_file": "source_video.webm", "thumbnail": None}))
+    inp = SongInput(title="T", artist="A", lyrics_text="x", cover_path=cover)
+    cfg = Config(whisper_url="u", output_dir=tmp_path / "out", work_dir=tmp_path)
+    return Ctx(inp=inp, cfg=cfg, work=work, log=lambda m: None)
+
+
+def make_cover(tmp_path):
+    p = tmp_path / "c.png"
+    Image.new("RGB", (640, 480), "red").save(p)
+    return p
+
+
+def make_pipeline_inputs(tmp_path, cover=None):
+    """Set up inputs at the exact work-dir path run_pipeline() would use."""
+    cfg = Config(whisper_url="u", output_dir=tmp_path / "out",
+                 work_dir=tmp_path / "work_root")
+    inp = SongInput(title="T", artist="A", lyrics_text="x", cover_path=cover)
+    work = cfg.work_dir / song_slug(inp)
+    work.mkdir(parents=True)
+    (work / "song.txt").write_text(SONG_TXT)
+    (work / "audio.wav").write_bytes(b"w")
+    (work / "source_video.webm").write_bytes(b"v")
+    (work / "meta.json").write_text(json.dumps(
+        {"video_file": "source_video.webm", "thumbnail": 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"))
+    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"
+    img = Image.open(out / "A - T.jpg")
+    assert img.width == img.height
+
+
+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"))
+    ctx = make_ctx(tmp_path, cover=make_cover(tmp_path))
+    package.package_stage().run(ctx)
+    out = tmp_path / "out" / "A - T"
+    song = parse_ultrastar((out / "A - T.txt").read_text())
+    assert get_header(song, "VIDEO") == "A - T.mp4"
+    # VIDEO keeps its original slot, ahead of BPM/GAP
+    keys = [k for k, _ in song.headers]
+    assert keys.index("VIDEO") < keys.index("BPM") < keys.index("GAP")
+
+
+def test_package_no_cover_drops_header(tmp_path, monkeypatch):
+    monkeypatch.setattr(package, "ffmpeg", lambda *a: Path(a[-1]).write_bytes(b"m"))
+    monkeypatch.setattr(package, "ffprobe_codec", lambda p: ("webm", "vp9"))
+    ctx = make_ctx(tmp_path, cover=None)
+    package.package_stage().run(ctx)
+    song = parse_ultrastar((tmp_path / "out" / "A - T" / "A - T.txt").read_text())
+    assert get_header(song, "COVER") is None
+
+
+def test_package_uses_work_thumbnail_when_no_cover_path(tmp_path, monkeypatch):
+    monkeypatch.setattr(package, "ffmpeg", lambda *a: Path(a[-1]).write_bytes(b"m"))
+    monkeypatch.setattr(package, "ffprobe_codec", lambda p: ("webm", "vp9"))
+    ctx = make_ctx(tmp_path, cover=None)
+    make_cover(tmp_path).rename(ctx.work / "thumb.png")
+    meta = json.loads((ctx.work / "meta.json").read_text())
+    meta["thumbnail"] = "thumb.png"
+    (ctx.work / "meta.json").write_text(json.dumps(meta))
+    package.package_stage().run(ctx)
+    out = tmp_path / "out" / "A - T"
+    assert (out / "A - T.jpg").exists()
+    song = parse_ultrastar((out / "A - T.txt").read_text())
+    assert get_header(song, "COVER") == "A - T.jpg"
+
+
+def test_package_outputs_done_marker():
+    ctx_outputs = package.package_stage().outputs
+    cfg = Config(whisper_url="u", output_dir=Path("out"), work_dir=Path("work"))
+    inp = SongInput(title="T", artist="A", lyrics_text="x")
+    ctx = Ctx(inp=inp, cfg=cfg, work=Path("work"), log=lambda m: None)
+    assert ctx_outputs(ctx) == [Path("work/package.done")]
+
+
+def test_package_corrupt_cover_degrades_to_no_cover(tmp_path, monkeypatch):
+    monkeypatch.setattr(package, "ffmpeg", lambda *a: Path(a[-1]).write_bytes(b"m"))
+    monkeypatch.setattr(package, "ffprobe_codec", lambda p: ("webm", "vp9"))
+    bad_cover = tmp_path / "bad.jpg"
+    bad_cover.write_bytes(b"not an image")
+    ctx = make_ctx(tmp_path, cover=bad_cover)
+    package.package_stage().run(ctx)
+    out = tmp_path / "out" / "A - T"
+    assert not (out / "A - T.jpg").exists()
+    song = parse_ultrastar((out / "A - T.txt").read_text())
+    assert get_header(song, "COVER") is None
+
+
+def test_package_partial_failure_is_not_cached(tmp_path, monkeypatch):
+    monkeypatch.setattr(package, "ffmpeg", lambda *a: Path(a[-1]).write_bytes(b"m"))
+
+    def boom(p):
+        raise PipelineError("ffprobe failed")
+
+    monkeypatch.setattr(package, "ffprobe_codec", boom)
+    ctx = make_ctx(tmp_path, cover=make_cover(tmp_path))
+    with pytest.raises(PipelineError):
+        package.package_stage().run(ctx)
+    out = tmp_path / "out" / "A - T"
+    assert (out / "A - T.txt").exists()
+    assert (out / "A - T.mp3").exists()
+    assert not (out / "A - T.webm").exists()
+    assert not package.package_stage().outputs(ctx)[0].exists()
+
+
+def test_package_success_is_cached_by_pipeline(tmp_path, monkeypatch):
+    monkeypatch.setattr(package, "ffmpeg", lambda *a: Path(a[-1]).write_bytes(b"m"))
+    monkeypatch.setattr(package, "ffprobe_codec", lambda p: ("webm", "vp9"))
+    inp, cfg = make_pipeline_inputs(tmp_path, cover=make_cover(tmp_path))
+    stage = package.package_stage()
+    run_pipeline(inp, cfg, log=lambda m: None, stages=[stage])
+    logs = []
+    run_pipeline(inp, cfg, log=logs.append, stages=[stage])
+    assert any("[package] cached, skipping" in m for m in logs)
+
+
+def test_package_retry_after_partial_failure_completes_folder(tmp_path, monkeypatch):
+    calls = {"n": 0}
+
+    def flaky_ffprobe(p):
+        calls["n"] += 1
+        if calls["n"] == 1:
+            raise PipelineError("ffprobe failed")
+        return ("webm", "vp9")
+
+    monkeypatch.setattr(package, "ffmpeg", lambda *a: Path(a[-1]).write_bytes(b"m"))
+    monkeypatch.setattr(package, "ffprobe_codec", flaky_ffprobe)
+    inp, cfg = make_pipeline_inputs(tmp_path, cover=make_cover(tmp_path))
+    stage = package.package_stage()
+    with pytest.raises(PipelineError):
+        run_pipeline(inp, cfg, log=lambda m: None, stages=[stage])
+    out = cfg.output_dir / "A - T"
+    assert (out / "A - T.txt").exists() and (out / "A - T.mp3").exists()
+    assert not (out / "A - T.webm").exists()
+
+    run_pipeline(inp, cfg, log=lambda m: None, stages=[stage])
+    assert (out / "A - T.webm").exists()
+    assert (out / "A - T.jpg").exists()
+
+
+def test_slash_in_title_is_stripped_from_folder_and_filenames(tmp_path, monkeypatch):
+    monkeypatch.setattr(package, "ffmpeg", lambda *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))
+    ctx.inp.title = "T/2"
+    package.package_stage().run(ctx)
+    out = tmp_path / "out" / "A - T2"
+    assert (out / "A - T2.txt").exists()
+    assert (out / "A - T2.mp3").exists()
+
+
+def test_all_stages_matches_stage_order():
+    assert [s.name for s in all_stages()] == STAGE_ORDER

tests/test_pipeline.py

@@ -0,0 +1,55 @@
+import json
+from pathlib import Path
+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",
+                  work_dir=tmp_path / "work")
+
+def make_inp():
+    return SongInput(title="T/est", artist="A", lyrics_text="la la")
+
+def fake_stage(name, calls):
+    def outputs(ctx):
+        return [ctx.work / f"{name}.done"]
+    def run(ctx):
+        calls.append(name)
+        (ctx.work / f"{name}.done").write_text("x")
+    return Stage(name=name, outputs=outputs, run=run)
+
+def test_slug_strips_slash():
+    assert song_slug(make_inp()) == "A - Test"
+
+def test_runs_all_stages_then_skips_cached(tmp_path):
+    cfg, inp, calls = make_cfg(tmp_path), make_inp(), []
+    stages = [fake_stage("s1", calls), fake_stage("s2", calls)]
+    run_pipeline(inp, cfg, log=lambda m: None, stages=stages)
+    assert calls == ["s1", "s2"]
+    run_pipeline(inp, cfg, log=lambda m: None, stages=stages)
+    assert calls == ["s1", "s2"]
+
+def test_force_from_reruns_tail(tmp_path):
+    cfg, inp, calls = make_cfg(tmp_path), make_inp(), []
+    stages = [fake_stage("s1", calls), fake_stage("s2", calls), fake_stage("s3", calls)]
+    run_pipeline(inp, cfg, log=lambda m: None, stages=stages)
+    calls.clear()
+    run_pipeline(inp, cfg, log=lambda m: None, stages=stages, force_from="s2")
+    assert calls == ["s2", "s3"]
+
+def test_warnings_surfaced(tmp_path):
+    cfg, inp = make_cfg(tmp_path), make_inp()
+    def run(ctx):
+        (ctx.work / "notes_sec.json").write_text(
+            json.dumps({"notes": [], "breaks": [], "warnings": ["w1"]}))
+        (ctx.work / "n.done").write_text("x")
+    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"]

tests/test_pitch.py

@@ -0,0 +1,19 @@
+import numpy as np
+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):
+    work = tmp_path / "w"
+    work.mkdir()
+    times = np.arange(0, 1.0, 0.01)
+    f0 = np.full_like(times, 220.0)
+    conf = np.full_like(times, 0.9)
+    monkeypatch.setattr(pitch, "track_pitch", lambda p, m, d: (times, f0, conf))
+    ctx = Ctx(inp=SongInput(title="T", artist="A", lyrics_text="x"),
+              cfg=Config(whisper_url="u"), work=work, log=lambda m: None)
+    pitch.pitch_stage().run(ctx)
+    data = np.load(work / "pitch.npz")
+    assert np.allclose(data["f0"], 220.0)
+    assert data["times"].shape == data["conf"].shape

tests/test_rebuild.py

@@ -0,0 +1,126 @@
+import dataclasses
+from pathlib import Path
+
+import pytest
+
+from songtrace.config import Config
+from songtrace.pipeline import STAGE_ORDER, PipelineError, SongInput
+from songtrace.rebuild import (FIELD_STAGE, changed_fields, plan,
+                                   relocate_work, superseded_output)
+
+
+def song(**kw):
+    base = dict(title="T", artist="A", lyrics_text="la", youtube_url="http://y")
+    return SongInput(**{**base, **kw})
+
+
+def cfg_for(tmp_path):
+    return Config(whisper_url="u", work_dir=tmp_path / "work",
+                  output_dir=tmp_path / "out")
+
+
+def test_every_song_input_field_is_mapped():
+    """A field added to SongInput without a FIELD_STAGE entry would make edits to
+    it invalidate nothing, silently building from stale intermediates."""
+    actual = {f.name for f in dataclasses.fields(SongInput)}
+    assert actual - set(FIELD_STAGE) == set()
+    assert set(FIELD_STAGE) - actual == set()
+
+
+def test_every_mapped_stage_is_a_real_stage():
+    assert set(FIELD_STAGE.values()) <= set(STAGE_ORDER)
+
+
+def test_no_change_means_no_force_from():
+    assert plan(song(), song()) is None
+
+
+@pytest.mark.parametrize("field,value,stage", [
+    ("youtube_url", "http://other", "acquire"),
+    ("video_path", Path("/tmp/v.mp4"), "acquire"),
+    ("language", "de", "transcribe"),
+    ("lyrics_text", "different words", "align"),
+    ("title", "Other", "grid"),
+    ("artist", "Other", "grid"),
+    ("genre", "rock", "grid"),
+    ("year", 1999, "grid"),
+    ("bpm_override", 123.0, "grid"),
+    ("cover_path", Path("/tmp/c.jpg"), "package"),
+])
+def test_single_field_maps_to_its_stage(field, value, stage):
+    assert plan(song(), song(**{field: value})) == stage
+
+
+def test_earliest_stage_wins_when_several_change():
+    new = song(lyrics_text="different", year=1999, language="de")
+    assert plan(song(), new) == "transcribe"
+
+
+def test_changed_fields_lists_what_moved():
+    assert sorted(changed_fields(song(), song(year=1999, genre="rock"))) == \
+        ["genre", "year"]
+
+
+def test_slug_fields_invalidate_no_later_than_the_stage_that_writes_the_slug():
+    """song.txt embeds the slug in #MP3/#COVER/#VIDEO and is written by `grid`,
+    so a title/artist edit must rerun grid or earlier for relocate_work to be safe."""
+    for field in ("title", "artist"):
+        assert STAGE_ORDER.index(FIELD_STAGE[field]) <= STAGE_ORDER.index("grid")
+
+
+def test_relocate_moves_the_work_dir_when_the_slug_changes(tmp_path):
+    cfg = cfg_for(tmp_path)
+    src = cfg.work_dir / "A - T"
+    src.mkdir(parents=True)
+    (src / "audio.wav").write_bytes(b"x")
+    assert relocate_work(song(), song(title="Fixed"), cfg) is True
+    assert not src.exists()
+    assert (cfg.work_dir / "A - Fixed" / "audio.wav").read_bytes() == b"x"
+
+
+def test_relocate_is_a_noop_when_the_slug_is_unchanged(tmp_path):
+    cfg = cfg_for(tmp_path)
+    src = cfg.work_dir / "A - T"
+    src.mkdir(parents=True)
+    assert relocate_work(song(), song(year=1999), cfg) is False
+    assert src.exists()
+
+
+def test_relocate_is_a_noop_when_there_is_nothing_to_move(tmp_path):
+    cfg = cfg_for(tmp_path)
+    assert relocate_work(song(), song(title="Fixed"), cfg) is False
+
+
+def test_relocate_refuses_to_clobber_an_existing_work_dir(tmp_path):
+    cfg = cfg_for(tmp_path)
+    (cfg.work_dir / "A - T").mkdir(parents=True)
+    (cfg.work_dir / "A - Fixed").mkdir(parents=True)
+    with pytest.raises(PipelineError, match="already exists"):
+        relocate_work(song(), song(title="Fixed"), cfg)
+    assert (cfg.work_dir / "A - T").exists()
+
+
+def test_superseded_output_names_the_replaced_folder(tmp_path):
+    root = tmp_path / "out"
+    assert superseded_output(str(root / "A - Old"), str(root / "A - New"), root) == \
+        (root / "A - Old").resolve()
+
+
+def test_superseded_output_is_none_when_the_slug_did_not_change(tmp_path):
+    root = tmp_path / "out"
+    assert superseded_output(str(root / "A - T"), str(root / "A - T"), root) is None
+
+
+def test_superseded_output_is_none_without_a_previous_build(tmp_path):
+    root = tmp_path / "out"
+    assert superseded_output(None, str(root / "A - T"), root) is None
+
+
+def test_superseded_output_refuses_a_path_outside_the_output_root(tmp_path):
+    root = tmp_path / "out"
+    assert superseded_output(str(tmp_path / "elsewhere"), str(root / "A - T"), root) is None
+
+
+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

tests/test_separate.py

@@ -0,0 +1,26 @@
+import numpy as np
+import soundfile as sf
+import torch
+from songtrace.config import Config
+from songtrace.pipeline import Ctx, SongInput
+from songtrace.stages import separate
+
+
+def test_separate_writes_stems(tmp_path, monkeypatch):
+    work = tmp_path / "w"
+    work.mkdir()
+    (work / "audio.wav").write_bytes(b"")
+    stems = {
+        "vocals": torch.ones(2, 100) * 0.5,
+        "drums": torch.ones(2, 100) * 0.1,
+        "bass": torch.ones(2, 100) * 0.1,
+        "other": torch.ones(2, 100) * 0.1,
+    }
+    monkeypatch.setattr(separate, "separate_audio", lambda p, m, d: stems)
+    ctx = Ctx(inp=SongInput(title="T", artist="A", lyrics_text="x"),
+              cfg=Config(whisper_url="u"), work=work, log=lambda m: None)
+    separate.separate_stage().run(ctx)
+    vocals, sr = sf.read(work / "vocals.wav")
+    assert sr == 44100 and vocals.shape == (100, 2)
+    instrumental, _ = sf.read(work / "instrumental.wav")
+    assert np.allclose(instrumental, 0.3, atol=1e-4)

tests/test_transcribe.py

@@ -0,0 +1,66 @@
+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)
+
+def test_vocal_sections_merges_close_segments():
+    sections = vocal_sections([seg(2, 5), seg(6, 9), seg(30, 33)])
+    assert sections == [(1.0, 10.0), (29.0, 34.0)]
+
+def test_vocal_sections_handles_out_of_order_segments():
+    sections = vocal_sections([seg(30, 33), seg(1, 4)])
+    assert sections == [(0.0, 5.0), (29.0, 34.0)]
+
+def test_vocal_sections_keeps_contained_segment():
+    sections = vocal_sections([seg(1, 10), seg(3, 5)])
+    assert sections == [(0.0, 11.0)]
+
+def test_similarity_high_for_matching_lyrics():
+    ref = normalize_lyrics("she's got a smile\nthat it seems to me")
+    segs = [seg(0, 2, "She's got a smile"), seg(2, 4, "that it seems to me")]
+    assert lyrics_similarity(segs, ref) > 0.9
+
+def test_similarity_low_for_wrong_lyrics():
+    ref = normalize_lyrics("completely different words here entirely")
+    segs = [seg(0, 2, "she's got a smile that it seems to me")]
+    assert lyrics_similarity(segs, ref) < 0.4
+
+def test_classify_lyrics_match_flags_extra_reference_line():
+    ref = normalize_lyrics("she's got a smile\nthat it seems to me\nbut then it fades")
+    segs = [seg(1, 4, "she's got a smile that it seems to me")]
+    dropped, missing = classify_lyrics_match(segs, ref)
+    assert dropped == [2]
+    assert missing == []
+
+def test_classify_lyrics_match_flags_unmatched_singing():
+    ref = normalize_lyrics("she's got a smile\nthat it seems to me")
+    segs = [seg(1, 4, "she's got a smile that it seems to me"),
+            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("she's got a smile\nthat it seems to me")
+    # line 1 heard with one word garbled ("it" -> "bit"): still a replace/equal
+    # match, so the line stays matched rather than dropped
+    segs = [seg(1, 4, "she's got a smile that bit seems to me")]
+    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 matched line keeps its notes.
+    ref = normalize_lyrics("she's got a smile\nthat it seems to me")
+    segs = [seg(1, 4, "she's got a smile that it seems to me"),
+            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

tests/test_ultrastar.py

@@ -0,0 +1,38 @@
+from pathlib import Path
+from songtrace.ultrastar import (
+    Note, LineBreak, parse_ultrastar, serialize_ultrastar,
+    get_header, format_bpm, parse_bpm,
+)
+
+FIXTURES = Path(__file__).parent / "fixtures"
+
+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"
+    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")
+    assert any(isinstance(x, LineBreak) and x.beat == 95 for x in song.body)
+    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"))
+    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"))
+    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"):
+        song = parse_ultrastar(load(name))
+        again = parse_ultrastar(serialize_ultrastar(song))
+        assert again == song
+
+def test_format_bpm():
+    assert format_bpm(256.0) == "256"
+    assert format_bpm(217.98) == "217,98"

tests/test_whisper.py

@@ -0,0 +1,60 @@
+import httpx
+from pathlib import Path
+from songtrace.config import Config
+from songtrace.whisper_client import TranscriptSegment, transcribe
+
+def test_transcribe_parses_segments(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["auth"] = request.headers.get("authorization")
+        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/v1", whisper_api_key="sekrit")
+    segs = transcribe(cfg, audio, client=client)
+    assert segs == [TranscriptSegment(start=0.5, end=2.0, text="hello world")]
+    assert seen["url"] == "http://w:9000/v1/audio/transcriptions"
+    assert seen["auth"] == "Bearer sekrit"
+
+
+def test_transcribe_maps_language_to_iso_code(tmp_path):
+    audio = tmp_path / "v.wav"
+    audio.write_bytes(b"RIFF")
+    seen = {}
+
+    def handler(request: httpx.Request) -> httpx.Response:
+        seen["body"] = request.read().decode("utf-8", errors="replace")
+        return httpx.Response(200, json={"segments": []})
+
+    client = httpx.Client(transport=httpx.MockTransport(handler))
+    cfg = Config(whisper_url="http://w:9000/v1")
+    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"]