Copy page

VoiceHub 0.3 release readiness

This is the authoritative release-candidate checklist for VoiceHub 0.3.0. It separates locally reproducible checks, cross-platform CI, real-checkpoint evidence, and maintainer-controlled publication. A pending hardware or external gate is never counted as a pass.

Candidate report

Updated 2026-08-04. The source version is 0.3.0. PyPI still serves 0.1.6, so installing voicehub from PyPI does not yet provide the current repository contract.

Gate	Current evidence	Status
Source, docs, benchmark, and PyPI candidate alignment	Local check identified 0.3.0 as a new candidate over PyPI 0.1.6	Passed
Full CPU-safe suite	Merge commit 566db682 passed the supported Linux, macOS, and Windows test matrix in CI run 30749398917. Fresh exact-current macOS environments passed 2,582 tests and 4,657 subtests with 15 skips and 35 warnings on Python 3.10.19 in 226.26 seconds and Python 3.11.15 in 212.37 seconds. The Python 3.12.12 full-dependency environment executed three additional default-runtime paths and passed 2,585 tests and 4,795 subtests with 12 skips and 35 warnings in 181.72 seconds. All seven asset/oracle paths pass separately on this worktree; only three Triton and two CUDA-extension paths remain locally unpassed	Passed remotely for merge commit 566db682 and locally on exact-current macOS Python 3.10-3.12; exact-current Linux, Windows, and remote default-runtime CI remain pending
Formatting and lint	PR 71 exact head 8ea5e94 passed CI lint and pre-commit.ci; merge commit 566db682 passed CI lint. The exact current candidate set now passes every configured hook over 4,675 tracked and candidate untracked files, excluding the protected uv.lock, with identical before/after SHA-256 manifests	Passed locally for the exact current worktree and remotely for merge commit 566db682; exact-current-worktree remote lint remains pending
Documentation	Merge commit 566db682 passed the strict build and deployment in Docs run 30749398929; the current shell passed the strict eleven-language build, the eleven-route DOM validator, and a pinned Playwright 1.62.0 plus Axe 4.12.1 validator. Its 60 route/viewport/palette cases protect geometry, state, screenshot-derived signatures, automated detectable accessibility rules, and complete native sequential focus. Six cases each enforce the current Home, Installation, Quickstart, Pipeline, Auto Classes, SpeechT5, Trainer, Optimization, Contribution, and Models API route-specific structure and content inventories. Six activate Home page copy, six activate Installation first-code copy, six activate Installation page copy, six activate every Quickstart tab set by keyboard, six activate Quickstart page copy, six activate Pipeline code copy, and six each activate Auto Classes, SpeechT5, Trainer, Optimization, Contribution, and Models API page copy at every viewport and palette, verify exact clipboard text, visible success/idle states where required, and focus retention, and rerun Axe after activation. Twenty pointer and 20 Enter cases activate the final TOC target on every representative route in both palettes, require the exact hash, CSS target, one settled active link, visible heading alignment, zero overflow, keyboard focus retention where applicable, and rerun Axe. Forty desktop/tablet Ctrl+K cases and 20 mobile pointer cases open and close search on every representative route in both palettes, protect focus and ARIA/inert/tab-order state, rerun Axe while open, restore focus to the visible breakpoint-specific control after Escape, and allow native Tab to continue without delayed restoration. Thirty Enter and 30 pointer cases open and close the version menu across every route, viewport, and palette, require three exact destinations, one current item, visible in-viewport geometry, focus and ARIA state, zero overflow, and an open-state Axe pass, then restore focus to the summary after Escape. Twenty ArrowDown cases switch every representative desktop/tablet route to Turkish and 20 pointer/semantic cases switch them to Arabic; all 40 require the exact 11-locale inventory and production-base destinations, selected target, LTR/RTL direction, palette preservation, zero overflow, and a localized-route Axe pass. Twenty Enter cases and 20 pointer cases switch every representative desktop/tablet route to the opposite palette; all 40 require exact toggle labels and targets, route and locale stability, focus transfer to the newly visible toggle, preserved geometry, zero overflow, and a post-switch Axe pass. Twenty Enter and 20 pointer cases focus the exact source repository link on every representative desktop/tablet route, validate its name, target, geometry, outline, overflow, and Axe result, and perform deterministic browser navigation. Thirty settled interaction cases cover open search/results/empty, version, branch, and drawer states. The latest matrix traversed 4,745 visible focus stops, rejected inactive branches, reached BODY, and returned to the skip link in every base case; it recorded 348 keyboard cases in total. Auto Classes uses its exact API ancestry, while all 68 generated model guides use their exact Base classes ancestry. Thirty-six nested-branch cases activate and restore all nine visible Base classes/model disclosures on SpeechT5 at desktop and tablet widths in both palettes, retain the sticky rail through document scroll, require unique page-subsection landmark names, and rerun Axe. Thirty-two root-branch cases activate and restore every top-level disclosure, and two separate cases retain mobile drawer activation evidence. Sixty page-action cases cover all ten representative routes at all three viewports and both palettes: 30 Enter and 30 pointer paths perform 60 exact edit navigations, 114 exact previous/next navigations, and 60 Back to top activations with focus, palette, route, overflow, and Axe checks	Passed remotely for merge commit 566db682 and locally for the tested current slice; the uncommitted worktree is not covered remotely, automated Axe coverage is not a complete WCAG or manual audit, and direct raw-pixel equivalence with intentionally different upstream content, branding, and palette is not claimed
Wheel, sdist, and editable installs	Merge commit 566db682 passed Package CI run 30749398914. A fresh current-slice probe produced a 57,194,978-byte wheel and a 55,468,722-byte source distribution and passed wheel, sdist, and editable installs with 68 models, 81 provenance manifests, 193 compliance files, required package data, zero dependency violations, and no eager PyTorch import. Earlier worktree fingerprint builds are retained below as historical evidence	Passed remotely for merge commit 566db682 and locally for the tested current slice; exact tagged-workflow artifact hashes remain pending
Python 3.10–3.12 on Linux, macOS, and Windows	Merge commit 566db682 passed all nine version/platform jobs, both runtime smokes, default runtime, training, and lint in CI run 30749398917. Fresh exact-current environments pass the complete suite on macOS 26.5.2 arm64 with Python 3.10.19, 3.11.15, and 3.12.12, PyTorch 2.8.0, Transformers 5.14.1, and pytest 9.1.1; the Python 3.12 run also activates the full default runtime	Passed remotely for merge commit 566db682 and locally for the exact-current macOS worktree on Python 3.10-3.12; exact-current remote Linux, macOS, and Windows execution remains pending
Canonical AI guidance	Merge commit 566db682 passed the canonical guidance contract in every supported platform matrix job; the current tree's merged guidance change passed 13 local tests and 11 subtests	Passed cross-platform for merge commit 566db682 and locally for the recorded tree
Released-checkpoint TTS, ASR, and VAD evidence	Dated RTX 4090 JSON and guide; see matrix below	Passed for the listed representatives
Pinned small release assets	The official ESPNet configuration plus SenseVoice and SpeechBrain tokenizers at immutable revisions matched declared sizes, file fingerprints, extracted tokens, and published encoding vectors on the exact current worktree	Passed locally for the exact current worktree; Package CI and the tagged release build repeat all three opt-in online gates
TEN-VAD checkpoint oracle	The official 315,449-byte ONNX graph at immutable revision 22a3bcd4509d0faaa8eef4881e8af5f39c178950 matched its pinned digest, converted to native Safetensors, and matched ONNX Runtime across 25 recurrent steps on the exact current worktree	Passed locally with pinned ONNX Runtime 1.22.1; Package CI and the tagged build repeat the isolated development oracle
NVIDIA QuartzNet checkpoint conversion	The official 70,993,538-byte NGC stt_en_quartznet15x5 1.0.0rc1 archive matched its pinned digest and exact converted tensor fingerprint on the exact current worktree	Passed locally; Package CI and the tagged build repeat the isolated conversion without redistributing the NGC artifact
WeNet GigaSpeech checkpoint and tokenizer	The immutable openspeech/wenet-models mirror at revision 90acd57d17169a15d5ceab462c6e7db3bd003921 supplied the exact 503,845,602-byte upstream archive with pinned SHA-256 061ccfa51d64ebe7ea091a5a13ae31e37d9c36f4eface5c7bafc80bd4a06b26e; restricted conversion and published tokenizer vectors passed on the exact current worktree	Passed locally; Package CI and the tagged build now repeat both isolated gates without redistributing the pickle archive
Tokenless publication workflow	Source contract test verifies separate build/publish jobs, protected environment, and job-scoped OIDC	Passed locally; tagged run pending
Protected pypi environment and PyPI publisher	GitHub's environment inventory currently contains only github-pages; PyPI publisher settings require maintainer access	Pending maintainer configuration
Git tag, GitHub release, and PyPI publication	No local/remote v0.3.0 tag or matching GitHub release exists; publication requires explicit maintainer approval	Pending
Shared diffusion-serving dispatch	Native and vLLM-Omni model sets are derived from registered architecture features; runtime extension and fail-closed tests pass	Passed; the registry-wide shared-layer policy covers the remaining provider-branch boundary
Shared task-factory defaults	TTS, ASR, and VAD no-argument defaults are unique ModelSpec.default_for_task declarations; modern and compatibility factories follow live metadata and contain no registered-model literals	Passed; the registry-wide shared-layer policy covers the remaining provider-branch boundary
Package-root public API inventory	A generated reference resolves all 261 unique voicehub.__all__ exports, records canonical source and line, kind, version-stable signature or constant/type-alias marker, summary, and lazy state, and fails on duplicate, unresolved, undocumented, source-less, or stale entries. The exact focused contract passes on Python 3.10-3.12, and a fresh subprocess confirms package-root import remains PyTorch-free	Passed locally; exact-current remote platform execution remains pending
Model contribution completion gate	An activated package-local manifest now supplies the lazy ModelSpec, aliases, and honest inference-only training profile without editing either central catalog; inactive scaffolds stay undiscovered and duplicate or richer unsupported declarations fail explicitly	Passed for representative zero-central-edit, TTS/ASR/VAD, extension, mismatch, lazy-import, and no-mutation fixtures
Model-page source provenance	The generator resolves source records from model packages and every lazy native-architecture component without importing an implementation; 58 pages link existing SOURCE.json files and the remaining 10 explicitly report that no integration-specific record is bundled. All 68 pages use one registry-derived uppercase-first display label, the same nine-section detail contract, verified paper and upstream GitHub metadata, and direct links to their resolved configuration and model facade source while retaining canonical keys	Passed; registry-derived documentation tests reject missing files, false provenance links, unresolved facade source, missing upstream GitHub references, incomplete paper metadata, fewer than four parsed examples, absent limitations or evidence boundaries, stale generated pages, stale display names, and regressions across representative TTS, ASR, and VAD architectures
Checkpoint documentation provenance	A shared declarative contract distinguishes real Hugging Face repositories, external archives, verified mirrors, and local-only inputs; page, index, gallery, quickstart, and notebook generation consume the same metadata without provider-name branches	Passed; 59 real Hub notebooks remain, WeNet records its immutable byte-identical mirror without presenting the external archive as a native Hugging Face model repository, and integrations without a default no longer claim a registry default
Native dependency boundary	The policy derives 755 seed files from the stable core boundary, all 383 immediate model-package Python facades, literal architecture component references, and three runtime-generated vendored roots; the 1,304-file fixed-point closure has zero violations and does not import PyTorch	Passed; a new model facade joins the default audit without a central provider-list edit
Shared provider independence	A source-only AST policy checks all 202 shared Python files against 68 canonical model types and 102 live aliases; declarative metadata and model-local code remain valid, while comparisons and every supported condition form fail	Passed with zero shared behavior violations; runtime extensions join without a central list edit
Universal optimization lifecycle	The registry-derived CPU-safe contract executes all six public passes across all 68 models: 408 pairs cover application, validation, manifest reporting, strict JSON serialization, semantic/state preservation, restoration, and cleanup without silent skips. Pass-specific suites separately exercise configured and compiled application paths	Passed locally for all 408 pairs; three Triton and two compiled CUDA-extension checks remain hardware-limited and unpassed

Layered verification matrix

Contract coverage and checkpoint coverage answer different questions and stay separate in release reports.

Layer	Scope	Evidence	Interpretation
Registry and lazy construction	All 34 TTS, 23 ASR, and 11 VAD integrations	Full pytest suite plus the TTS and ASR/VAD registry audits	Proves discovery, configuration, lazy allocation, and normalized contract behavior; it does not prove every public weight file
Native graph and checkpoint shape	Every registered family	Model-specific CPU/meta tests and immutable inventory tests	Proves the implemented graph and declared checkpoint namespace under deterministic fixtures
Package surface	Every registered runtime	scripts/check_distribution.py and Package CI	Proves wheel, sdist, editable install, package data, import coverage, and declared dependencies
TTS real checkpoint	vits with facebook/mms-tts-eng@c71de0f	tts_vits_rtx4090_2026-07-31.json	Complete tokenizer, acoustic graph, flow, and vocoder path on an RTX 4090
ASR real checkpoints	Moonshine tiny, Wav2Vec2 base, and Whisper tiny through five adapters	asr_vad_rtx4090_2026-07-31.json	Complete transcription paths with deterministic text and single-sample WER; not a corpus accuracy claim
ASR checkpoint conversion	NeMo QuartzNet15x5 character CTC	Pinned NGC 1.0.0rc1 archive, source/config/weights digests, tensor inventory, and strict native reload	Proves the official 639-tensor namespace converts to the implemented graph; it is not a corpus accuracy claim
VAD real checkpoints or algorithms	Silero, Sherpa/Silero, WebRTC, and Auditok	Same ASR/VAD JSON evidence	Complete segmentation paths with boundary/score comparisons
Hardware-limited remainder	Gated, restricted, multi-gigabyte, or unavailable checkpoints	Provider documentation and explicit coverage records in benchmark JSON	Pending by design; no claim that every public checkpoint was downloaded

The readable methodology and results are in the RTX 4090 speech benchmark. Performance numbers are checkpoint- and machine-specific and are not release thresholds on other hardware.

Local release gates

Run from a clean checkout with Python 3.12 after installing .[test,training,docs]:

python scripts/check_release.py
python -m pytest -q
pre-commit run --all-files
mkdocs build --strict --clean
python scripts/check_documentation_dom.py site
python scripts/check_documentation_visual.py site
python scripts/check_distribution.py

Before the final capability-driven serving slice, the local working tree passed the full suite in independent macOS environments on Python 3.10.19 (173.39 seconds), Python 3.11.15 (154.82 seconds), and Python 3.12.12 (106.98 seconds). The Python 3.10 and 3.11 runs created isolated temporary environments, installed .[test] with uv's CPU PyTorch backend, and did not use or modify the repository lock file. Each run reported 2,400 passed, 25 skipped, 3,287 subtests, and 35 warnings.

The current tree then reran the complete Python 3.12 suite after replacing the diffusion-serving model allowlist with architecture capability discovery. It reported 2,401 passed, 25 skipped, 3,287 subtests, and 35 warnings in 102.48 seconds. The 16 changed-contract tests separately passed under Python 3.10.19, 3.11.15, and 3.12.12. Skipped paths are recorded as unverified, not passed. Candidate CI separately passed the supported Python versions on Linux, macOS, and Windows for commit e2bfb4a; that remote result does not prove later uncommitted changes.

The next provider-branch slice moved ASR and VAD task defaults into unique ModelSpec.default_for_task declarations. The current Python 3.12 tree then reported 2,405 passed, 25 skipped, 3,287 subtests, and 35 warnings in 97.82 seconds. Its five changed registry/auto tests passed under Python 3.10.19, 3.11.15, and 3.12.12 without loading a heavy backend. An attempted broader Python 3.10 run reused the Python 3.12 environment and failed on the expected compiled-PyTorch ABI mismatch; it is invalid cross-version evidence and is not counted as a pass.

The following compatibility slice declared the existing Orpheus TTS default in the same registry metadata and removed the provider literal from AutoInferenceModel. The final Python 3.12 tree reported 2,406 passed, 25 skipped, 3,289 subtests, and 35 warnings in 103.26 seconds. Its three changed, dependency-light default-contract tests passed under Python 3.10.19, 3.11.15, and 3.12.12. The runtime replacement regression proves both the modern TTS factory and the compatibility factory follow a new process-local default without editing either shared factory.

The next contribution slice replaced the scaffold checker's quoted-name search with import-free AST validation of built-in ModelSpec, alias, and training profile declarations. A quoted comment can no longer satisfy discovery; wrong lazy modules, classes, config paths, checkpoints, tasks, aliases, and training tasks produce separate errors. TTS, ASR, and VAD fixtures pass, and a completed separately distributed extension still passes without either central catalog file. The final Python 3.12 tree reported 2,411 passed, 25 skipped, 3,303 subtests, and 35 warnings in 95.12 seconds. Its five changed pure-Python contract tests passed under Python 3.10.19, 3.11.15, and 3.12.12. The first focused run exposed a shadowed checkpoint variable, and the first broad docs run exposed a non-compiling alias example; both failed runs are excluded from passing evidence, and both exact regressions passed after correction.

The following contribution slice added the read-only catalog command. It derives the built-in ModelSpec, alias entries, task enum, and honest inference-only training profile from model-integration.json, labels each central insertion point, and never imports VoiceHub or a model backend. The same output completed the checker fixtures for TTS, ASR, and VAD and remained byte-for-byte deterministic without changing their file inventories. Its three pure-Python tests passed on Python 3.10.19, 3.11.15, and 3.12.12. The focused test's first fake-catalog assembly had invalid indentation, and the first selected pre-commit run reformatted the new docstring; neither failed run is counted. The corrected current tree reported 2,414 passed, 25 skipped, 3,307 subtests, and 35 warnings in 107.99 seconds.

The complete current executable and test tree then passed the same full suite in independent macOS environments on Python 3.10.19 (166.51 seconds) and Python 3.11.15 (153.60 seconds). Each direct uv pip install -e ".[test]" used the CPU PyTorch backend in a new temporary virtual environment and did not read or modify the repository lock file. Together with the Python 3.12.12 run, all three supported interpreters reported exactly 2,414 passed, 25 skipped, 3,307 subtests, and 35 warnings. This strengthens supported-version evidence; it does not substitute for Linux or Windows execution of the later local tree.

The next architecture-policy slice removed 230 model/provider paths from a 253-entry manually maintained native-runtime seed list. The audited boundary is now derived from 23 stable core roots, every immediate Python file in each model package, literal lazy component references in architecture registrations, and only three narrow source roots needed by runtime-generated vendored imports. The resulting 755 seeds reach a 1,304-file fixed-point closure with zero external dependency violations and no eager PyTorch import. A temporary future model facade importing transformers fails the default audit without any central policy edit, while MeloTTS and OpenVoice regressions prove that active internal closures stay covered and dormant vendored frontends stay excluded. The pure policy probe reported the same 755 seeds, 1,304 closure files, zero violations, and no PyTorch import on Python 3.10.19, 3.11.15, and 3.12.12. The current Python 3.12.12 tree then reported 2,416 passed, 25 skipped, 3,307 subtests, and 35 warnings in 113.71 seconds. Two initial focused runs exposed missing literal architecture references and direct-seed/closure test semantics; the corrected tests pass, and those failed runs are not counted as release evidence.

The following shared-architecture slice unified the boundary previously covered by four partial provider-branch checks into one registry-wide policy. It scans all 202 Python files outside model-local models/ and architectures/ roots against 68 canonical model types and 102 live aliases. Declarative catalogs, licensing, provenance, and training metadata remain valid, while provider literals in comparisons, if and conditional expressions, loop conditions, assertions, comprehension filters, and match cases or guards fail with file and line evidence. A runtime-registered model and alias join the default audit without editing a policy list. The complete related slice reported 141 passes and 178 subtests; the same synthetic if and match violations were detected on Python 3.10.19, 3.11.15, and 3.12.12. The current Python 3.12.12 tree then reported 2,420 passed, 25 skipped, 3,307 subtests, and 35 warnings in 113.34 seconds. The first selected pre-commit run let docformatter update the new policy file and exited nonzero; the formatted source passed the second run and the focused four-test suite, so the first run is not counted as passing evidence.

The next contribution slice made model-integration.json the package-local source of truth for a completed inference-only built-in. New scaffolds are inactive by default and therefore cannot enter either registry while their runtime, checkpoint revision, or evidence is incomplete. Once explicitly activated, one strict source-only parse requires every package facade, IMPLEMENTATION_STATUS = "ready", an immutable checkpoint revision, and bundled license text before it derives the lazy ModelSpec, aliases, capabilities, components, checkpoint, and task; the same manifest produces an honest inference-only ModelTrainingSpec. No model package or PyTorch module is imported, and neither voicehub/models/registry.py nor voicehub/training/specs.py needs a model entry. A richer training claim must use an explicit profile, while simultaneous manifest and legacy central declarations fail as three actionable duplicate errors. The focused scaffold suite reported 19 passes and 35 subtests, and the related registry/training slice reported 116 passes and 330 subtests. The same temporary model and training profile were discovered without importing PyTorch or the model package on Python 3.10.19, 3.11.15, and 3.12.12. The current Python 3.12.12 tree then reported 2,424 passed, 25 skipped, 3,310 subtests, and 35 warnings in 114.23 seconds. The final activation-gate rerun reported the same counts in 111.58 seconds. The first documentation run exposed its obsolete requirement for a central alias mapping, and the first two selected pre-commit runs applied formatter changes and exposed one continuation-indent error. The corrected documentation contract and formatted source pass; none of those failed runs is counted as passing evidence.

The following CPU-safe evidence slice audited every skip reason. Seven tests in the old mock-Transformers VAD class were permanently skipped because the native Wav2Vec2 provider had replaced that runtime. The redundant mock loader, windowing, training, and serialization fixtures were removed; four still-valid speech-label, checkpoint-stride, and invalid-training-input edge contracts were moved into the executable native provider suite. The two focused native VAD files reported 21 passes and 20 subtests, and the related VAD, Wav2Vec2, registry, task, and inference slice reported 137 passes and 235 subtests. The complete Python 3.12.12 tree then reported 2,428 passed, 18 skipped, 3,310 subtests, and 35 warnings in 112.27 seconds. Every remaining skip names its missing dependency, dedicated CI job, CUDA host, checkpoint, or release asset; none is counted as passing evidence. The changed test sources also parsed on Python 3.10.19, 3.11.15, and 3.12.12; syntax parsing is not reported as runtime coverage on the two interpreters without installed test dependencies.

The next default-runtime evidence slice corrected the active codec import inventory. Its Higgs entry still pointed at a dormant vendored tokenizer that required vector_quantize_pytorch, even though the registered higgstts wrapper executes VoiceHub's native PyTorch-only tokenizer. The gate now imports the active native Higgs module beside the DAC, FishTTS, and Irodori codec paths while explicitly rejecting audiotools, loguru, and vector_quantize_pytorch. It executes under the installed test environment without a dependency skip. The focused codec file reported six passes; the related codec, Higgs, FishTTS, Llasa, dependency-policy, and optimization slice reported 130 passes, 52 subtests, and four warnings. The complete Python 3.12.12 tree then reported 2,429 passed, 17 skipped, 3,310 subtests, and 35 warnings in 114.45 seconds. The changed test source parsed on Python 3.10.19, 3.11.15, and 3.12.12; parsing alone is not runtime evidence on the two interpreters without installed PyTorch.

The following default-runtime slice removed the last two optional-dependency skips. They exercised dormant MeloTTS Japanese and OuteTTS GGUF provider files, while the registered built-ins use VoiceHub's native MeloTTS and OuteTTS architectures. Their replacements import each native frontend, runtime, and public wrapper in fresh processes while rejecting MeloTTS's legacy MeCab, pykakasi, unidic_lite, and Transformers frontends and OuteTTS's legacy llama_cpp, loguru, polars, and Transformers backends. The formatted focused compatibility file reported seven passes. The related native provider, inference, lifecycle, compile-target, and dependency-policy slice reported 144 passes and 61 subtests. The complete Python 3.12.12 tree then reported 2,431 passed, 15 skipped, 3,310 subtests, and 35 warnings in 114.57 seconds. Every remaining skip now names a dedicated CI job, CUDA or toolkit requirement, checkpoint, or release asset; none is counted as passing evidence. The first selected pre-commit run let YAPF rewrite the new helper and exited nonzero; the formatted focused suite and second hook run passed, so the first run is not counted. The test source parsed on Python 3.10.19, 3.11.15, and 3.12.12; parsing alone is not runtime evidence on interpreters without installed dependencies.

The next release-asset slice exercised the remaining ESPNet token-list gate against the official Hugging Face repository at immutable revision bc6bbd771cec698f070640ee677a66719181f0a2. The downloaded configuration was 82,131 bytes and matched SHA-256 16351b9bf79631d1df0a4645a858dc330c40434cf03470408c9c8fd446b6ea19; its extracted token list matched SHA-256 48ec6eedbee6a22e2a9b51adeb425af3c39db23128086c015240f591601a3ea3. The opt-in test passed through VoiceHub's dependency-free Hub transport, while the same test remained explicitly skipped in the default offline run. Package CI and the tagged release build now require this online gate. The related ESPNet, Hub transport, release-readiness, and packaging slice reported 56 passes, one offline skip, and 81 subtests. An earlier unqualified python invocation could not start pytest in the local shell and is not counted as a pass. The first selected pre-commit run let YAPF format the changed test and exited nonzero; the formatted tests and second hook run passed. The complete current Python 3.12.12 tree then reported 2,431 passed, 15 skipped, 3,310 subtests, and 35 warnings in 112.86 seconds. The online ESPNet pass remains separate from those offline suite counts.

The following tagged-runtime slice closed a release-workflow discrepancy. Main CI already ran the three opt-in default-runtime tests on Ubuntu and in macOS and Windows smoke jobs, but both full-suite steps in the tagged release workflow left them skipped. The tagged nine-job Python/platform matrix and the dependent build job now set VOICEHUB_FULL_RUNTIME_TEST=1; the workflow source contract requires both declarations. The focused file reported five passes and 138 subtests. The release-equivalent complete Python 3.12.12 macOS run then reported 2,434 passed, 12 skipped, 3,448 subtests, and 35 warnings in 110.63 seconds. Those three default-runtime paths are locally verified, while remote execution of the changed tagged matrix is still pending. The separate ESPNet online gate accounts for one of the 12 remaining default-run skips, leaving 11 hardware, checkpoint, tokenizer, or ONNX paths unverified.

The next release-asset slice added the official SenseVoiceSmall tokenizer at immutable revision 3847d57b6bdf2dd8875cb1508d2af43d80a16bf7 to the same opt-in online gate as ESPNet. VoiceHub's Hub transport downloaded the 377,341-byte file, which matched SHA-256 aa87f86064c3730d799ddf7af3c04659151102cba548bce325cf06ba4da4e6a8. The native tokenizer reproduced the pinned English text IDs, control-token labels, and semantic language, emotion, and event values. Package CI and the tagged build now run both release-asset tests through one plural gate. Their focused online run reported two passes; the same SenseVoice test remained an explicit skip offline. The related SenseVoice, ESPNet, Hub, release, and packaging slice reported 71 passes, two offline skips, and 81 subtests. The first selected pre-commit run let YAPF format the SenseVoice test and exited nonzero; the formatted online tests and second hook run passed. With the default-runtime and both release-asset gates enabled, the complete Python 3.12.12 macOS suite reported 2,436 passed, 10 skipped, 3,448 subtests, and 35 warnings in 112.97 seconds.

The following real-checkpoint slice exercised the official TEN-VAD ONNX graph from immutable source revision 22a3bcd4509d0faaa8eef4881e8af5f39c178950. The 315,449-byte source matched SHA-256 e10b98a0cab1c98e847fbdda14cb3d45a38336d47535a3f63a0fb6c4e0f4cdf4 before VoiceHub's standard-library ONNX reader converted it to native Safetensors. The differential oracle used pinned ONNX Runtime 1.22.1 for 25 deterministic recurrent feature steps; native speech probability and four LSTM states stayed within the existing 2e-7 and 2e-6 absolute-error thresholds. The test now passes the expected official source digest into conversion and requires the exported metadata to report an official source. Package CI and the tagged build download the exact raw file, install ONNX Runtime only for this isolated development gate, and leave the wheel runtime dependency-free. The focused oracle passed and remained explicitly skipped without its source environment variable. The related VAD, checkpoint, release, and packaging slice reported 64 passes, one offline skip, and 115 subtests. With every local opt-in gate enabled, the complete Python 3.12.12 macOS suite reported 2,437 passed, 9 skipped, 3,448 subtests, and 35 warnings in 112.68 seconds.

The next real-checkpoint slice exercised NVIDIA's official stt_en_quartznet15x5 archive from NGC release 1.0.0rc1. The 70,993,538-byte source matched SHA-256 1b9b7b87a9277e6fef164d8f99d1226f0511af154423bbf919b920421ac9602f. VoiceHub's restricted, weights_only converter then verified the embedded configuration and weight digests, all 639 tensor names and shapes, the 19,018,554-value state count, and the native Safetensors strict reload. Package CI and the tagged build now download the exact NGC version and run only this isolated conversion gate; the NVIDIA-governed artifact remains outside the distribution. The focused oracle passed, and the related NeMo, release, packaging, and distribution slice reported 32 passes and 76 subtests. The first all-opt-in run reported 2,437 passes and 9 skips because its temporary TEN-VAD source was no longer present; that run is not evidence for an eight-skip state. After redownloading and fingerprinting the pinned TEN-VAD graph, the complete Python 3.12.12 macOS suite reported 2,438 passed, 8 skipped, 3,448 subtests, and 35 warnings in 114.03 seconds. The remaining paths are five CUDA/Triton or CUDA-toolkit checks, one SpeechBrain tokenizer gate, and the WeNet checkpoint and tokenizer gates.

The following documentation-provenance slice removed a false negative from the model-page generator. It previously searched a registry architecture alias as if it were a package directory and therefore claimed that 38 integrations had no bundled source record. Source discovery now follows the lazy module paths already declared by each native ArchitectureSpec, checks both package-root and source/ layouts without importing an implementation, and preserves model-local precedence. Twenty-eight regenerated pages now point to real manifests; 58 of 68 pages have an existing source link and the remaining 10 honestly report no integration-specific record. The focused documentation contract reported 3 passes and 142 subtests, including exact TTS, ASR, and VAD examples, path existence, deterministic 68-page regeneration, and a fresh process that imports none of NeMo, Safetensors, SentencePiece, PyTorch, or Transformers. The related documentation, registry, optimization, scaffold, packaging-metadata, and distribution-contract suite reported 102 passes and 1,839 subtests; strict multilingual documentation and release-version alignment also passed. The first clean distribution run passed the wheel probe but its sdist probe exited nonzero without reproducible stderr and is not counted as a pass. The exact sdist install and probe then passed in a preserved clean environment, and a fresh complete distribution run passed wheel, sdist, and editable installs with 68 models, 81 provenance manifests, 193 compliance files, zero dependency violations, and no eager PyTorch import.

The same evidence audit rechecked WeNet before selecting this bounded slice. The pinned upstream README still lists the 20210728 GigaSpeech U2++ archive, but the official HTTP artifact endpoint returned 404 on 2026-08-02, its HTTPS variant failed certificate-hostname validation, and the apparent wenet/gigaspeech-u2pp-conformer Hugging Face page returned 404. The 503,845,602-byte archive is therefore recorded as currently inaccessible, not passed; its checkpoint and tokenizer opt-in tests remain explicit unverified gates.

The next checkpoint-documentation slice removed the resulting false public claim. A shared CheckpointDocumentation projection now derives provider, authoritative URL, availability status, quickstart input, and limitation note from architecture metadata. It is model-independent and preserves the lazy boundary. The WeNet package declares an external archive, links the immutable upstream README, records the 2026-08-02 HTTP 404 and certificate-hostname failure in SOURCE.json, and shows only a local VoiceHub-native artifact in its quickstart. Its false Hugging Face link, Colab badge, gallery row, and generator-owned notebook were removed. The same generic path corrected styletts2 and vad_transformers from "registry default" to "no default". The focused page/notebook contract reported 4 passes and 127 subtests. The related WeNet, architecture, registry, optimization, scaffold, provider-independence, documentation, and packaging suite reported 140 passes, 2 explicitly unverified WeNet skips, and 1,906 subtests. The first selected pre-commit run applied YAPF and is not counted as passing; the formatted second run passed. Strict multilingual documentation, release alignment, and fresh wheel, sdist, and editable probes also passed.

The following release-asset slice exercised SpeechBrain's published tokenizer.ckpt from immutable model revision 979a53a7a3f6c9291c02c040fd8ebfb2471cf8a3. VoiceHub's dependency-free Hub transport downloaded 253,217 bytes with SHA-256 37a6cba34cd520b33fd83612d5efc8ba7e351166541eb2726642bb3032234d31. The native SentencePiece implementation reproduced the pinned HELLO WORLD encoding and decoding vectors. The common release-asset opt-in now covers this test beside ESPNet and SenseVoice, and both Package CI and the tagged release build require the three-test gate. The focused SpeechBrain/workflow contract reported 3 passes, the combined online release-asset gate reported 3 passes, and the related SpeechBrain, release, packaging, registry, documentation, optimization, and scaffold suite reported 129 passes and 1,838 subtests. After redownloading and fingerprinting the TEN-VAD and QuartzNet artifacts, the full opt-in Python 3.12.12 macOS suite reported 2,442 passes, 7 skips, 3,521 subtests, and 35 warnings in 115.78 seconds. The remaining skips are five CUDA/Triton or CUDA-toolkit gates plus the inaccessible WeNet checkpoint and tokenizer; none is counted as passed. Strict multilingual documentation, release-version alignment, and fresh wheel, source-distribution, and editable probes also passed.

The next supported-version evidence slice installed the current working tree from source into fresh temporary CPU environments with direct uv pip install commands; the repository lock-file hash remained unchanged. The full default-offline suite reported 2,434 passes, 15 explicit skips, 3,383 subtests, and 35 warnings on both Python 3.10.19 in 179.77 seconds and Python 3.11.15 in 166.56 seconds. Together with the current Python 3.12.12 full opt-in result, the complete tree has now executed on every supported interpreter on macOS. This does not substitute for Linux or Windows execution of the uncommitted tree; the latest remote nine-job matrix remains tied to f2d6332.

The following formatting-and-lint evidence slice ran the complete repository pre-commit configuration against the current working tree. End-of-file, trailing-whitespace, case-conflict, private-key, AWS-credential, pyupgrade, isort, YAPF, Markdown formatting, Flake8, and docformatter hooks all passed without modifying a file. This replaces older partial-hook evidence; it does not substitute for the pending Linux and Windows matrix.

The release workflow first repeats the complete test suite on Linux, macOS, and Windows with Python 3.10, 3.11, and 3.12. Its build job cannot begin unless all nine tagged-source jobs pass. It then repeats the remaining local gates, checks that tag v0.3.0 points at the checked-out commit, verifies that 0.3.0 is not already on PyPI, builds one wheel/sdist pair, checks their embedded metadata and size, and transfers those exact artifacts to a separate publish job.

Fresh clean-install builds passed embedded name, version, and size checks and remained below the 100 MB publication limit. Gzip timestamps and evidence-only documentation edits can change archive bytes, so release sizes and hashes are recorded from the exact workflow artifacts rather than copied from a local build.

The current repository-wide pre-commit gate initially normalized imports in 15 files and applied YAPF and docformatter changes. That first formatter run exited with failure and is not counted as a pass. The second pre-commit run --all-files completed every hook successfully. A raw AST fingerprint changed because docformatter edits docstring constants, so it is not used as behavior evidence. The formatted Python 3.12 tree instead reran the complete suite and reported 2,400 passed, 25 skipped, 3,287 subtests, and 35 warnings. The later diffusion-serving slice passed its selected hooks after formatter changes and the Python 3.12 full suite reported 2,401 passed with the same 25 skips, 3,287 subtests, and 35 warnings. The task-default slice also required one formatter pass before its selected hooks succeeded; its full-suite count was 2,405. The compatibility-default slice also required one isort pass before its selected hooks succeeded; the second selected run passed and the final full-suite count was 2,406. The scaffold-catalog checker slice passed its selected hooks without a formatter rewrite, and its full-suite count was 2,411. The catalog-renderer slice required one docformatter rewrite before its selected hooks passed; its final full-suite count is 2,414. The native dependency-boundary slice required one YAPF/docformatter rewrite before its selected hooks passed; the corrected focused tests reported 19 passes, and its final full-suite count is 2,416. The shared provider-independence slice required one docformatter rewrite before its selected hooks passed; its final full-suite count is 2,420. The manifest discovery slice required two formatter runs before its selected hooks passed; its final full-suite count is 2,424. The later VAD skip-audit slice passed its selected hooks without a rewrite and reduced the current unverified count from 25 to 18; its final full-suite count is 2,428. The default-runtime codec slice also passed its selected hooks without a rewrite and reduced the current unverified count to 17; its final full-suite count is 2,429. The native MeloTTS/OuteTTS dependency slice required one YAPF rewrite before its selected hooks passed and reduced the current unverified count to 15; its final full-suite count is 2,431.

The diffusion-serving resolver no longer contains a provider-name allowlist. Its native and vLLM-Omni verified model snapshots are derived from diffusion-serving-native and diffusion-serving-vllm-omni features declared beside each architecture integration. A runtime-registration regression proves that a new native model can join the resolver without editing shared serving code; unverified pairings and VibeVoice's incomplete high-level path still fail closed.

The shared auto factories no longer embed registered TTS, ASR, or VAD model names. The former AutoInferenceModel Orpheus default is now the TTS registry declaration, preserving its no-argument behavior while sharing the same policy with AutoModelForTextToSpeech. The registry enforces at most one default_for_task declaration per task and exposes the selection through get_default_model_spec(). A runtime extension regression proves that the no-argument factory follows replaced registry metadata without changing voicehub/auto.py; missing and ambiguous defaults fail explicitly.

PR 68 was merged as main commit 679d5bd before its required matrix was green. PR 69 head b6de7b5 subsequently repaired its canonical AI skill frontmatter and normalized materialized root-guidance pointers. CI run 30741892984 then passed every Linux and macOS Python job, both runtime smokes, default runtime, training, and lint. Windows 3.10, 3.11, and 3.12 each failed only the same two scaffold-checker assertions: actionable diagnostics rendered temporary repository-relative paths with native \ separators while the public cross-platform contract expected stable / separators. Each Windows job otherwise reported 2,437 passes, 16 explicit skips, 3,392 subtests, and 35 warnings.

The bounded correction centralizes repository-relative display through PurePath.as_posix() and includes a dependency-free PureWindowsPath regression. The focused scaffold file reported 20 passes and 35 subtests; the related registry, model-page, optimization, release, and AI-guidance slice reported 113 passes and 1,849 subtests; selected pre-commit hooks, all 68 generated model pages, all 59 notebooks, and release alignment also passed. The complete Python 3.12.12 suite reported 2,439 passes, 15 explicit skips, 3,394 subtests, and 35 warnings in 115.67 seconds. That correction is now PR 69 head 3a3e224; CI run 30742766090 passed every Linux, macOS, and Windows Python 3.10-3.12 job, both runtime smokes, default runtime, training, and lint.

The following documentation-parity slices record the official Transformers main commit and toctree fingerprint, map representative routes, and restore the left navigation and right table of contents on all eleven localized home sources. Rendered checks at 1440 x 900, 1024 x 768, and 390 x 844 verified the desktop shell, mobile collapse, both VoiceHub palettes at all three viewports, and the absence of horizontal overflow. The tablet slice then matched the reference region behavior at 1024 x 768: a persistent 270-pixel left navigation, hidden right table of contents, 739-pixel content region inside the 1,009-pixel main region, and no redundant drawer button. The mobile drawer and desktop three-column shell remained structurally unchanged. The drawer backdrop now excludes the 242-pixel panel itself, leaving a 133 x 844 pointer target at 390 x 844. A real backdrop click in both LTR and RTL layouts and Escape in the LTR layout cleared the drawer checkbox and returned the panel off-canvas with no horizontal overflow. Initial probes against the full-width backdrop and the preexisting Escape behavior did not close the drawer; those failed checks are not counted. The final 30 documentation contracts reported 1,132 subtests, the related registry and optimization slice reported 70 passes and 310 subtests, all 68 generated model pages and 59 notebooks were current, and the strict multilingual build completed. Release alignment, 13 AI-guidance and release-readiness tests with 11 subtests, and selected hooks also passed. Fresh wheel, source-distribution, and editable probes passed with the inventory recorded above. The earlier shell slice's first selected hook run let YAPF format its regression and exited nonzero; the formatted regression and second hook run passed, so the failed run is not counted as passing evidence.

The distribution gate inventories every SOURCE.json, license, licence, NOTICE, and COPYING file under the installed package. Each of the 81 source manifests must contain a pinned revision/release and explicit license metadata; all 193 compliance files plus the project-level Apache-2.0 license must survive both wheel and source-distribution builds.

PR 69 exact head 3a3e224 passed every required job in CI run 30742766090. Interim main head 6a75eda subsequently passed Package CI and the strict documentation build in runs 30745818302 and 30745818292. The documentation-parity branch derives from 6a75eda. PR 71 exact head 8ea5e94 passed all nine platform/version jobs, both runtime smokes, default runtime, training, lint, strict documentation, package build, and pre-commit.ci. The documentation deployment job was skipped for the pull request and is not counted as a pass. PR 71 was merged externally at 2026-08-02 13:11 UTC as 566db6822d47a335c720efb9ea66d7bcb22a1a82. That exact merge commit passed all nine platform/version jobs, both runtime smokes, default runtime, training, lint, the strict documentation build and deployment, and Package CI in runs 30749398917, 30749398929, and 30749398914. The current uncommitted navigation and header worktree remains outside that remote evidence.

The next left-navigation state slice maps the Transformers and VoiceHub Installation routes. Exact rendered checks at 1440 x 900, 1024 x 768, and 390 x 844 now cover one visible active item, the expanded current branch, visible keyboard focus, and zero horizontal overflow. VoiceHub rendered its 219 x 34-pixel desktop item and 212 x 34-pixel tablet item in both light and dark themes. Its opened mobile drawer measured 242 pixels with a 133-pixel backdrop and exposed one 251 x 48-pixel active row in both themes. The mapped Transformers active item measured about 218 x 31 pixels at desktop and tablet widths and 323 x 31 pixels in its opened mobile navigation; all three reference states exposed visible keyboard focus without horizontal overflow. The focused source regression first failed before the state styles existed. A rendered keyboard probe then exposed Material's focus-visible compatibility class; the first selector did not draw an outline, and that failed probe is not counted. The compatibility selector and focused regression now pass; the exact viewport rerun's three focused navigation and drawer tests also pass. The documentation file reported 30 passes and 1,132 subtests, the related registry and universal-optimization files reported 46 passes and 236 subtests, all 68 model pages and 59 model notebooks were current, release alignment found all five benchmark files, the strict multilingual documentation and fresh wheel, source-distribution, and editable probes passed. The complete Python 3.12.12 suite reported 2,444 passes, 15 explicit skips, 3,406 subtests, and 35 warnings in 116.03 seconds. The first selected hook run let YAPF format the new regression and exited nonzero; the formatted second run passed, so the first is not counted. The reference mobile light state and the remaining representative page matrix remain pending and are not reported as passed.

The next bounded documentation slice replaced the five product tabs with the Goal's eight-section hierarchy: Get started, Base classes, Inference, Training, Quantization and optimization, Ecosystem integrations, Resources, and API. The generated 68-model block now remains nested under Base classes > Models, while all existing documentation routes remain unchanged. The source contract first failed against the five-section configuration and then passed after the navigation and all ten localized label inventories were updated. Rendered checks at 1440 x 900, 1024 x 768, and 390 x 844 verified exact order, visible keyboard focus, both VoiceHub themes, and zero horizontal overflow. These later uncommitted worktree changes are not covered by PR 71 exact-head CI and must not be treated as cross-platform or exact-head release evidence. On the current Python 3.12.12 worktree, 43 documentation, release, and guidance tests passed with 1,143 subtests; the related registry and universal-optimization files passed 46 tests with 236 subtests; all 68 model pages and 59 notebooks were current; release alignment and the strict eleven-language build passed; and the wheel, source-distribution, and editable probes passed. The complete suite reported 2,444 passes, 15 explicit skips, 3,406 subtests, and 35 warnings in 114.71 seconds. The first selected hook run let YAPF format the regression and exited nonzero; the formatted regression and second hook run passed, so the failed hook run is not counted.

The next bounded header slice adds the missing version control and renders the six product controls in the reference order: product, search, version, language, theme, and source. The version menu distinguishes the current main documentation and 0.3.0 release candidate from the published PyPI 0.1.6 package. Rendered desktop, tablet, mobile LTR, and mobile RTL checks exercised pointer dismissal, Enter and Space activation, Escape dismissal with focus restoration, aria-expanded, a two-pixel focus outline, both VoiceHub themes, and zero horizontal overflow. Initial source, keyboard, focus-compatibility, mobile-label, and RTL-overflow probes each failed before their corresponding implementation or correction; none is counted as passed. This evidence covers control order and version interaction only. The reference places its controls in the documentation rail at desktop and tablet widths and includes a Hugging Face corporate banner, while VoiceHub retains a Material top header. Exact placement and geometry, the corporate-banner decision, remaining focus order, other expanded control states, the reference mobile light render, and the representative page matrix remain pending. These uncommitted changes are not covered by merge-commit CI. The focused header regression and 31-test documentation file passed; the combined documentation, release, and guidance slice reported 44 passes and 1,143 subtests. The related registry and universal optimization slice reported 46 passes and 236 subtests. All 68 model pages and 59 notebooks were current, release alignment found all five benchmark files, the strict eleven-language build completed, and the fresh wheel, source-distribution, and editable probes passed. The complete Python 3.12.12 suite reported 2,445 passes, 15 explicit skips, 3,406 subtests, and 35 warnings in 111.61 seconds. Every applicable selected hook passed; the Markdown-formatting hook found no matching files and is not counted as a pass.

The following bounded search-interaction slice matches the reference expanded dialog's outer bounds at all three checked viewports: 500 x 72 pixels at x = 470, y = 64 on desktop; 500 x 72 pixels at x = 262, y = 64 on tablet; and 358 x 72 pixels at x = 16, y = 64 on mobile. VoiceHub's named trigger opens by pointer, Enter, Space, or Command/Ctrl+K, updates aria-expanded, focuses the query field, locks body scrolling, and closes with Escape while restoring a visible focus indicator. Light and dark desktop, tablet, and mobile states plus English LTR and Arabic RTL mobile containment had zero overflow. The reference Command+K and Escape paths were also exercised, and both reference mobile themes were captured. Its checked mobile icon has no accessible name, so that upstream limitation is not counted as a pass.

The initial 688 x 48-pixel local dialog, the first constrained 625 x 92/100- pixel correction, the first Escape path without visible focus restoration, and the first mobile state with 77 pixels of header overflow all failed and are not passing evidence. After correction, the focused search contract passed, the 32-test documentation file and combined documentation/release/guidance slice reported 45 passes and 1,143 subtests, and the registry and universal- optimization slice reported 46 passes and 236 subtests. All 68 model pages and 59 notebooks were current, release alignment found all five benchmark files, the strict eleven-language build completed, and fresh wheel, source- distribution, and editable probes passed with 68 models, 81 provenance manifests, and 193 compliance files. The complete Python 3.12.12 suite reported 2,446 passes, 15 explicit skips, 3,406 subtests, and 35 warnings in 114.70 seconds. Every applicable selected hook passed; the Markdown-formatting hook found no matching files and is not counted as a pass. These uncommitted changes remain outside merge-commit CI, and the collapsed reference-rail placement, corporate-banner decision, other expanded header controls, and remaining representative page matrix are still pending.

The next bounded language-control slice replaces Material's 40 x 40-pixel icon and 125 x 200-pixel custom list with the reference's compact native select. VoiceHub and Transformers now expose the same 48 x 26-pixel control dimensions at 1440 x 900 and 1024 x 768; their x/y coordinates still differ because the reference uses its documentation rail while VoiceHub retains the known top- header placement. Both sites hide the control at 390 x 844. VoiceHub preserves only its eleven built locales, labels the control accessibly, displays the current uppercase locale code, and selected TR through the control to load a Turkish document with the Turkish option still selected.

Pointer, Enter, and Space activation followed by Escape kept the current value and focus on the select on both sites. An outside-pointer probe transferred local focus to the target control, the local two-pixel focus indicator was visible, and light/dark desktop and tablet plus English LTR and Arabic RTL mobile states had zero overflow. Native option popups are browser chrome and are not claimed as DOM-geometry or screenshot evidence. The focused source test first failed because the override did not exist. The first rendered native-select probe then exposed white text on white; its scoped regression failed before the color correction. None of those failed checks is counted as passing evidence.

After correction, the focused language contract and 33-test documentation file passed. The combined documentation/release/guidance slice reported 46 passes and 1,143 subtests; registry and universal optimization reported 46 passes and 236 subtests; 68 model pages and 59 notebooks were current; release alignment found all five benchmark files; and the strict eleven-language build passed. Fresh wheel, source-distribution, and editable probes passed with 68 models, 81 provenance manifests, 193 compliance files, zero dependency violations, and no eager PyTorch import. The complete Python 3.12.12 suite reported 2,447 passes, 15 explicit skips, 3,406 subtests, and 35 warnings in 106.30 seconds. Every applicable selected hook passed; the Markdown-formatting hook found no matching files and is not counted. At that point, the uncommitted worktree was outside merge-commit CI, and reference-rail placement, the corporate-banner decision, theme/source exact responsive evidence, and the representative page matrix were left for later slices.

The following bounded header-tail slice replaces Material's 40 x 40-pixel theme control and 234 x 48-pixel repository block with the current reference's compact forms. At the executed 1280 x 720 viewport, VoiceHub rendered a 34 x 24-pixel theme button, a 55 x 16-pixel source link, and a 12-pixel gap, matching the measured reference geometry. The theme switched in both directions through pointer, Enter, and Space while restoring focus to the newly visible control with a two-pixel outline; both palettes retained zero overflow. The source link gained an accessible name and pointer activation reached the declared GitHub repository.

The focused source contract first failed because neither override existed. The first combined documentation run then exposed a stale single-selector language assertion after the mobile visibility rule was shared. An earlier browser-driven Enter probe on the native source link did not navigate, and the first embedded exact-viewport attempt was blocked by URL security policy. None of those failed checks is counted.

The final bounded correction adds an explicit Enter path to the named source link. Its focused regression first failed before the handler existed and then passed after correction. Browser evidence verified Enter navigation to the declared repository. Fresh 1440 x 900 and 1024 x 768 renders retained the 34 x 24-pixel theme button, 55 x 16-pixel source link, and 12-pixel gap in both palettes. Fresh 390 x 844 light and dark renders hid language, theme, and source with zero rendered width or height. Every state had zero document and header overflow; keyboard theme switching restored focus with a two-pixel outline and two-pixel offset. A separate attempt to move focus from the theme button with a synthetic Tab key did not move focus and is not counted as focus-order evidence.

After correction, the focused regression and 34-test documentation file passed. The combined documentation/release/guidance slice reported 47 passes and 1,143 subtests; the complete Python 3.12.12 suite reported 2,448 passes, 15 explicit skips, 3,406 subtests, and 35 warnings in 117.70 seconds. The strict eleven-language build completed; 68 model pages, 59 notebooks, and release alignment passed. Fresh wheel, source-distribution, and editable probes passed with 68 models, 81 provenance manifests, 193 compliance files, zero dependency violations, and no eager PyTorch import; the wheel measured 57,186,791 bytes and the source distribution 55,427,491 bytes. The worktree remains outside exact-commit remote CI.

The next bounded shell slice moves the six documentation controls out of the Material top-header layout and into the reference's desktop and tablet rail. The final exact-viewport renders use a 65-pixel global VoiceHub brand row, a 270-pixel left rail, and a 270 x 128-pixel rail-control block. At 1440 x 900 the main regions measure 270/900/270 pixels; at 1024 x 768 they measure 270/754 pixels with the right table of contents hidden. The product, collapsed search, and utility regions begin at x = 16 and y = 77, 112, and 150. The utility row preserves the reference's 80 x 26 version, 48 x 26 language, 34 x 24 theme, and 55 x 16 source controls and 12-pixel final gap. All eight root navigation sections are visible without horizontal tabs. The 1024-pixel Arabic route mirrors the rail to x = 754 and leaves the 754-pixel content region at x = 0. Light and dark desktop geometry is identical and every checked state has zero horizontal overflow.

The 390 x 844 shell retains the existing 64-pixel mobile header and working 242-pixel drawer. Moving search into the shared control wrapper initially made the opened mobile search region collapse to zero size; that failed probe is not counted. The scoped mobile-open rule was corrected and the rerun restored the expected 358 x 72 region at x = 16, y = 64, with Escape dismissal and zero overflow. Version Enter/Escape state, Command/Ctrl+K search focus and dismissal, theme switching, mobile drawer pointer/Escape behavior, LTR/RTL mirroring, and both palettes were exercised. VoiceHub has no Hugging Face corporate ecosystem, so the matched global row intentionally contains VoiceHub branding rather than invented corporate links. Complete sequential focus order and the eight pending representative page pairs remain explicit gates.

The focused shell contract and 35-test documentation file passed with 1,132 subtests. The registry and universal-optimization slice reported 46 passes and 236 subtests; all 68 model pages and 59 notebooks were current; release alignment found all five benchmark files; the strict eleven-language build and fresh wheel, source-distribution, and editable probes completed. These uncommitted changes remain outside merge-commit CI and are not cross-platform evidence. The complete Python 3.12.12 suite then reported 2,449 passes, 15 explicit skips, 3,406 subtests, and 35 warnings in 122.45 seconds.

The next bounded representative-page slice maps the Transformers and VoiceHub Quickstart routes. The source regression first failed against the First generation navigation label and the page's earlier task-example-only structure. After correction, VoiceHub exposes the mapped setup, lazy discovery, pretrained-model, inference, trainer, and next-step hierarchy, plus a page-copy action. Exact browser measurements match the reference article at 804 pixels wide on 1440 x 900, 658 pixels on 1024 x 768, and 342 pixels with a 24-pixel gutter on 390 x 844. The title, section headings, body type, desktop code block, and mobile 12/366-pixel code bounds also match. VoiceHub light and dark renders at all three viewports had zero horizontal overflow, and pointer copy reported Copied.

The first desktop gutter correction still resolved to Material's earlier 24-pixel physical margin because the theme selector had higher specificity; that intermediate render is not counted. The later equal-specificity override produced the measured 48-pixel gutter. The first copy-control render wrapped its label and was likewise replaced before final evidence. A locator-driven Enter/Tab probe did not expose copied or focus-visible state, so page-action keyboard activation and sequential focus remain pending rather than passed. An initial ad hoc optimization inventory also imported a nonexistent list_optimization_passes symbol and failed; the corrected registry-object query reported six passes and 408 model/pass pairs, while the failed command is not evidence.

After correction, the focused contract passed; the documentation file reported 36 passes and 1,133 subtests; and the registry and universal-optimization slice reported 46 passes and 236 subtests. All 68 model pages and 59 model notebooks were current, release alignment found all five benchmark files and 68 documented providers, and the strict eleven-language build passed. Fresh wheel, source-distribution, and editable probes passed with 68 models, 81 provenance manifests, 193 compliance files, zero dependency violations, and no eager PyTorch import; the wheel measured 57,186,791 bytes and the source distribution 55,428,271 bytes. The complete Python 3.12.12 suite reported 2,450 passes, 15 explicit skips, 3,407 subtests, and 35 warnings in 121.32 seconds. Every applicable selected hook passed; the Markdown-formatting hook found no matching files and is not counted. These uncommitted changes remain outside exact-commit remote CI. Home, Installation, and Quickstart remain partial at their stated interaction gates, and seven representative page pairs remain unexamined.

A follow-up iteration selected only the Quickstart page-copy keyboard boundary. The first focused source regression failed because the action had no explicit keyboard contract. The first correction intercepted keydown and disabled the button while copying; the rendered Enter probe then reported Copy failed and lost focus, so neither result was counted. A second regression rejected the disabled state and required an aria-busy lifecycle, a selection fallback, and focus restoration. Its rendered Enter probe preserved the two-pixel focus treatment but still reported Copy failed because both clipboard methods were rejected from the intercepted key event. The final bounded correction retains the native button's one keyboard-generated click, avoids preventDefault, and invokes the copy routine only once while restoring keyboard focus.

The focused source contract now passes. Pointer activation replaced a clipboard sentinel with the complete 5,252-character Quickstart article, aria-busy returned to false, and the exact 804/658/342-pixel article geometry remained unchanged at 1440 x 900, 1024 x 900, and 390 x 844. The 390-pixel render kept the 12/366-pixel code bounds; light and dark renders had zero horizontal overflow. Locator-driven Enter and Space both left the button focused with the solid two-pixel outline and two-pixel offset, but that driver did not dispatch the native button click; the seeded clipboard remained unchanged. Two bounded Chrome attempts could not navigate to the localhost preview. Native keyboard activation and sequential focus are therefore precise pending gates and are not reported as passed.

After this correction, the documentation file reported 36 passes and 1,133 subtests; the registry and universal-optimization slice reported 46 passes and 236 subtests; all 68 model pages and 59 model notebooks were current; and release alignment found all five benchmark files and 68 documented providers. The strict eleven-language build and selected applicable pre-commit hooks passed. Fresh wheel, source-distribution, and editable probes passed with 68 models, 81 provenance manifests, 193 compliance files, zero dependency violations, and no eager PyTorch import; the wheel measured 57,186,791 bytes and the source distribution 55,428,842 bytes. The complete Python 3.12.12 suite reported 2,450 passes, 15 explicit skips, 3,407 subtests, and 35 warnings in 113.65 seconds. These uncommitted changes remain outside exact-commit remote CI.

The next bounded representative-page slice maps the Transformers Pipeline tutorial to VoiceHub's speech tasks. The focused regression first failed at import time because VoiceHub did not export a pipeline function or task adapters. The implemented dependency-light contract now selects the registered TTS, ASR, or VAD auto factory, preserves normalized outputs, retains lazy loading, wraps configured model objects without changing runtime state, and rejects cross-task or incomplete models. It deliberately does not report a universal vectorized batch or chunking path.

The VoiceHub guide now follows the reference's task, parameter, device, batch, task-specific parameter, chunking, large-input, and large-model hierarchy while keeping speech-specific examples and limitations. Both articles measured 804 pixels from x = 318 at 1440 x 900, 658 pixels from x = 318 at 1024 x 768, and 342 pixels from x = 24 at 390 x 844. VoiceHub light and dark states retained those bounds with zero document overflow. The active Pipeline navigation, complete desktop TOC, mobile drawer, Tasks anchor, Command+K/Escape search, and pointer page copy passed; the copied article contained 6,742 characters.

The synthetic Enter page-action probe restored named-button focus and the focus-visible class but did not receive clipboard permission. Pointer copy in the same harness did receive permission and reported Copied. Native keyboard clipboard activation is therefore still inaccessible, and complete sequential focus is pending; neither is counted as passed. An initial click selected a hidden duplicate Tasks link and timed out, and the first reference geometry selector targeted the page shell rather than its article. The corrected visible link and article selector produced the recorded evidence; the failed probes are not evidence.

After correction, the focused pipeline/documentation slice reported 9 passes and 13 subtests. The registry, model-page, import, and documentation slice reported 118 passes, 3 explicit default-runtime skips, and 1,459 subtests. The optimization slice reported 168 passes and 759 subtests. The strict eleven-language build passed, and fresh wheel, source-distribution, and editable probes reported 68 models, 81 provenance manifests, 193 compliance files, zero dependency violations, and no eager PyTorch import. The wheel was 57,189,063 bytes and the source distribution was 55,432,769 bytes. The complete Python 3.12.12 suite reported 2,458 passes, 15 explicit skips, 3,417 subtests, and 35 warnings in 115.32 seconds. The skips remain assigned to three default-runtime CI jobs, five CUDA/Triton/toolkit paths, and seven inaccessible release assets, checkpoints, or oracles. Every selected pre-commit hook passed, including Markdown formatting. These uncommitted changes remain outside exact-commit remote CI.

The next bounded shared-shell slice verifies the right table of contents on the mapped Pipeline pair. The current official and local pages both use a 270-pixel desktop rail at x = 1,170, keep their link geometry stable while the article scrolls, hide the rail at 1,024 and 390 pixels, and retain zero horizontal overflow. VoiceHub's tracking contract selected exactly one of Tasks, Parameters, or Large models after the corresponding heading crossed the fixed-header threshold. Its earlier inherited filled pill was heavier than the reference, so the smallest correction scopes the secondary active state to a transparent background, no box shadow, and the existing modern VoiceHub color token in both light and dark themes.

The focused regression first failed because that scoped secondary state did not exist, then passed after the correction. The first runner invocation used an unavailable python command and is not evidence; the same test was executed with the repository virtual environment for both the failing and passing runs. Pointer anchors placed headings below the fixed shell, and focused TOC links rendered a solid two-pixel outline with a two-pixel offset. Locator-driven Enter and a native browser keypress both left the target link focused without navigating, so keyboard anchor activation remains inaccessible and is not counted as passed.

After the correction, the complete documentation contract file reported 38 passes and 1,135 subtests, and the strict eleven-language site build passed. Every applicable selected pre-commit hook passed; the Markdown-formatting hook found no matching files and is not counted. This slice changes no registry, runtime, optimization, or packaging contract, so the immediately preceding successful evidence for those surfaces remains applicable. The working tree is still outside exact-commit remote CI.

The next bounded shared-shell slice closes the left documentation rail's desktop and tablet scroll-offset mismatch. The official Pipeline product label and navigation region move upward by the 65-pixel global row as the article scrolls. VoiceHub now applies the same bounded transition: at 1,440 x 900 its product controls move from y = 65 to y = 0, its product label from y = 77 to y = 12, its primary rail from y = 65 to y = 0, and its independent navigation region from y = 193 / 707 pixels to y = 128 / 772 pixels. At 1,024 x 768, the scrolled primary rail has a 640-pixel navigation region below the 128-pixel control block and the secondary rail remains hidden. Light and dark renders at both widths retain the same geometry and zero horizontal overflow. The 390 x 844 shell retains its fixed 64-pixel mobile header and working 242-pixel drawer without overflow. Keyboard focus on the desktop search input renders a solid two-pixel indigo form outline with a two-pixel offset.

The focused source regression first failed because the shell did not track the article offset. Intermediate header-transform and header-position attempts changed fixed descendants or obscured the rail controls and are not counted. Material's inline primary-rail dimensions and tablet stacking state required the final narrowly scoped overrides. The resulting focused regression passes; the documentation contract reports 39 passes and 1,135 subtests; JavaScript syntax, whitespace validation, the strict eleven-language build, and every applicable selected pre-commit hook pass. Markdown formatting found no matching files and is not counted. One earlier full documentation run failed only because its source assertion still required the superseded fixed top offset; the assertion was updated to the tracked contract before the passing complete run. This documentation-only slice does not change registry, model, optimization, runtime, or package contents, so the immediately preceding successful evidence for those surfaces remains applicable. The worktree is still outside exact-commit remote CI.

The next bounded representative-page slice maps the Transformers Auto Classes reference to VoiceHub's model index. The refreshed inventories still contain 68 models (34 TTS, 23 ASR, and 11 VAD), 102 aliases, 68 generated pages, 59 generated notebooks, six public optimization passes, and 408 model/pass pairs. The first source regression failed because the index still began with Model guides. A new immutable ModelSpec.display_name projection now derives the presentation label from the public class name without changing any canonical key. The generator uses it consistently in every model-page heading, index link, and navigation label and exposes a compact AutoConfig, AutoProcessor, task-specific auto-model, discovery, and registered-model hierarchy.

The local Auto Classes page rendered in both VoiceHub palettes at 1,440 x 900, 1,024 x 768, and 390 x 844. Its article measured 804, 658, and 342 pixels; the desktop left and right rails measured 270 pixels; the right rail collapsed at the smaller widths; and every state retained one active Auto Classes item and zero document overflow. The four mobile tables scroll inside their own wrappers. Keyboard focus on the active desktop item rendered a solid two-pixel outline and inset cue. The official reference was rendered at its available 1,280 x 720 viewport and exposed the mapped auto-class hierarchy with zero overflow. Its product banner, community panel, NLP families, and exhaustive generated API inventory are intentional product/content differences.

The focused index and guide checks reported two passes and 136 subtests, and the all-task display-name regression reported one pass and 68 subtests. The broader registry, documentation, and optimization slice reported 141 passes and 1,597 subtests. The complete documentation file reported 40 passes and 1,204 subtests; the strict eleven-language build passed; and candidate release alignment found five benchmark files and 68 documented providers. The full Python 3.12.12 suite reported 2,462 passes, 15 explicit skips, 3,554 subtests, and 35 warnings in 126.54 seconds. Fresh wheel, source-distribution, and editable probes reported 68 models, 81 provenance manifests, 193 compliance files, zero dependency violations, and no eager PyTorch import. The wheel was 57,189,200 bytes and the source distribution was 55,434,554 bytes. Every applicable selected pre-commit hook passed after YAPF's first invocation reformatted the changed Python files; that modifying invocation is not counted as a pass. Markdown formatting found no matching files and is not counted.

One combined focused-test command named three nonexistent test files and collected no tests; the corrected command used the repository's actual registry and optimization filenames and produced the 141-pass result above. An ad hoc inventory that attempted to load MkDocs' Python-specific YAML tag with safe_load also failed and is not evidence. Complete sequential focus is still unverified: a locator Tab probe did not advance, and the native desktop fallback was safety-blocked. These uncommitted changes remain outside exact-head remote CI, so neither that interaction gate nor cross-platform coverage for this worktree is reported as passed.

The next bounded representative-page slice maps the official Transformers SpeechT5 model detail to VoiceHub's generated model guides. The refreshed inventories remain 68 models (34 TTS, 23 ASR, and 11 VAD), 102 aliases, 68 generated pages, 59 generated notebooks, six public optimization passes, and 408 model/pass pairs. The first focused SpeechT5 regression failed because the old guide did not provide the mapped Usage, configuration, processing, limitations, and source-linked API hierarchy. The generator now applies one nine-section detail contract to every registered model without importing its implementation. It resolves paper and upstream GitHub metadata alongside the stable configuration and model facades, requires at least four parseable Python examples, separates checkpoint defaults from execution evidence, and states optional-dependency, hardware, optimization-failure, and real-checkpoint limitations explicitly.

The local SpeechT5 page rendered with 804-, 658-, and 342-pixel articles at 1,440 x 900, 1,024 x 768, and 390 x 844. Every state had zero document overflow; all eight desktop tables fit their wrappers, and all eight mobile tables scrolled inside their wrappers. The default palette passed all three widths and the dark palette passed desktop. The desktop primary and secondary rails were visible, the right table of contents exposed the complete generated hierarchy, the smaller widths collapsed the right rail, and the mobile drawer showed one visible active SpeechT5 item. Pointer copy reported Copied, the edit action targeted the SpeechT5 documentation source, both public-facade links targeted their resolved local source paths, and desktop active-link focus rendered a solid two-pixel outline with a two-pixel offset. The official page was rendered at the controller's available 1,280 x 720 viewport and exposed the mapped configuration, processor, task-model, and vocoder API hierarchy, expandable parameters, copy controls, and zero overflow.

The complete documentation file reported 41 passes and 1,272 subtests. The corrected registry, latest-ASR, speech-task, and universal-optimization slice reported 101 passes and 427 subtests. All 68 model pages and all 59 model notebooks were generator-current, the strict eleven-language build passed, and every applicable selected pre-commit hook passed after YAPF's first invocation reformatted the changed Python files; that modifying invocation is not counted as a pass. The first post-implementation representative run expected double quotes while the generator correctly emitted Python repr with single quotes, and an intermediate all-page run used an obsolete 210-line bound; both failed assertions were corrected before the complete passing run. One broader command named nonexistent test_latest_asr_models.py and test_speech_task.py files and collected no tests; the corrected command used test_latest_asr_registry.py and test_speech_task_registry.py and produced the 101-pass result. Two attempted mobile dark-theme clicks exceeded the browser selector deadline and are not evidence.

Exact reference renders at the three target viewports, dark local tablet and mobile renders, and complete sequential focus remain pending and are not counted as passed. This documentation-generation slice does not change package contents or model runtime behavior, so distribution probes and the full suite were not rerun; their immediately preceding evidence remains applicable. The worktree remains outside exact-commit remote CI.

The next bounded representative-page slice maps the official Transformers Trainer overview to VoiceHub. The pinned upstream snapshot remains commit b3a36037d3feb22e3f0174b3dd4248fcc0f0f722 with toctree SHA-256 f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a; the 30-line Trainer source has SHA-256 e7c5368c1223c2b195321468fcd1ac64c5cda52a1e44d4adf854a6473b6c9ee0. The refreshed repository inventories remain 68 models (34 TTS, 23 ASR, and 11 VAD), 102 aliases, 68 complete model pages, 59 generated notebooks, six public optimization passes, 408 model/pass pairs, five benchmark records, and eight model-contribution steps.

The first focused regression failed because Training still linked its only Get started item directly to the long fine-tuning guide. VoiceHub now exposes one nested Trainer overview followed by Fine-tuning, matching the reference mental model without discarding the speech-specific workflow. The concise overview documents the shared Trainer and TrainingArguments loop, the model-owned objective boundary, fail-closed support validation, and four speech-relevant next steps. It contains the same title and single Next steps heading depth as the reference and adds no copied upstream prose, generic loss, or unsupported training claim.

The local overview rendered at 1,440 x 900, 1,024 x 768, and 390 x 844 with 804-, 658-, and 342-pixel articles. Light and dark states had identical geometry, one visible active overview item, the expected desktop TOC and responsive rail/drawer behavior, and zero document overflow at every width. Pointer copy reported Copied; the edit action targeted the new source; the next footer link targeted Fine-tuning; and focused navigation rendered a solid two-pixel outline with a two-pixel offset. The official page rendered at the controller's available 1,280 x 720 viewport with a 644-pixel article, one active overview item, one Next steps heading, copy and source actions, and zero document overflow. Its main-version banner, community panel, and Trainer video remain intentional product-content differences.

The complete documentation file reported 42 passes and 1,279 subtests. The focused Trainer, training-contract, adapter, runtime, and speech-training slice reported 127 passes and 180 subtests; the registry and universal-optimization slice reported 47 passes and 304 subtests. All 68 model pages and all 59 model notebooks were generator-current, release alignment found five benchmark files and 68 documented providers, the strict eleven-language build passed, and all applicable selected pre-commit hooks passed without modifying a file. Markdown formatting found no matching files and is not counted.

The first post-implementation focused rerun used a line-sensitive prose assertion and failed on normal Markdown wrapping; the corrected regression normalizes whitespace before checking the contract. The first strict preview rebuild also rejected a nonexistent #trainingarguments API anchor; the link now targets the existing #training-arguments anchor and the complete strict build passes. A screenshot attempt waited on a drawer button after the responsive state had changed and timed out; the corrected desktop capture is the visual evidence. An exact reference viewport override remained fixed at 1,280 x 720, two reference theme-control clicks did not change the palette, and the local short-page TOC click did not update the URL hash. Exact responsive reference renders, reference dark mode, that anchor interaction, and complete sequential focus remain pending and are not counted as passed.

This documentation-only navigation slice does not change model, training, optimization, or package runtime code. The full suite and distribution probes were therefore not rerun; their immediately preceding evidence remains applicable. The untracked uv.lock stayed unchanged, and the worktree remains outside exact-commit remote CI.

The next bounded representative-page slice replaces the stale Transformers perf_infer_gpu_one mapping with the current official optimization_overview. The pinned upstream commit remains b3a36037d3feb22e3f0174b3dd4248fcc0f0f722; the current overview source has SHA-256 19622667a7299f258f5c9a72940c9f26492619636f35d9bd592701c02745b620, and the toctree remains f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a. The refreshed local inventories remain 68 models (34 TTS, 23 ASR, and 11 VAD), 102 aliases, 68 complete model pages, 59 generated notebooks, six public optimization passes, 408 model/pass pairs, five benchmark records, and eight model-contribution steps.

The first focused regression failed because the fifth navigation group still linked its Overview label to the detailed TTS workflow. VoiceHub now exposes a concise Optimization overview before that preserved workflow and the existing codec, diffusion, and VITS-family guides. Its table covers compile, flash-attention-4, custom-kernels, codec-kernels, diffusion-cache, and diffusion-sampling exactly once. A single copyable helper demonstrates discovery, application, manifest reporting, and restoration. The prose keeps support evidence honest: compatibility is decided by runtime validation, not registry discovery; failed reversible passes roll back; no registry-wide public quantization pass exists; parallelism is a training or serving topology; and continuous batching belongs to the serving scheduler.

The local page rendered at 1,440 x 900, 1,024 x 768, and 390 x 844 with 804-, 658-, and 342-pixel articles. Light and dark states retained the same geometry, one active Overview item, the expected desktop TOC and responsive rail/drawer behavior, internal mobile table scrolling, copy/edit actions, previous/next navigation, and zero document overflow. The official page also rendered at all three target viewports with its current heading hierarchy, active Overview, desktop TOC, responsive rail, technique table, version banner, community panel, copy action, and zero document overflow. Reference dark mode and complete sequential focus remain pending and are not counted as passed.

The corrected focused regression passed with six subtests. The complete documentation file reported 43 passes and 1,292 subtests. Registry and all documented public optimization suites reported 138 passes and 839 subtests; one PyTorch decomposition warning was emitted and is not reported as a pass or failure. All 68 model pages and 59 model notebooks remained generator-current, release alignment found five benchmark files and 68 documented providers, and the strict eleven-language documentation build passed. The earlier failing focused run remains a development failure, not evidence.

This navigation and documentation slice does not change package contents or model runtime behavior, so the full suite and distribution probes were not rerun; their immediately preceding evidence remains applicable. Exact-commit remote CI, reference dark mode, complete sequential focus, protected publisher configuration, and publication remain open. The untracked uv.lock stayed unchanged.

The next bounded representative-page slice corrects the stale Transformers contribution mapping. The pinned upstream commit remains b3a36037d3feb22e3f0174b3dd4248fcc0f0f722; the current modular guide has SHA-256 a52af1ee1754e6e658535f071a727fac31a057a3a848c28c18873a5ff371dd96, the legacy contribution guide has SHA-256 4d7e7066deeefde340c3e0460eae540f343cdfe642d690a600cba0a90441cb03, and the pinned toctree remains f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a. The current upstream hierarchy places the modular guide under Base classes, Models, Contribute and labels the former route Legacy model contribution.

VoiceHub now mirrors that nested contribution group and uses the generic Add a model label in every configured locale. The guide maps the upstream reduced-boilerplate and standalone-file outcome to speech-specific composition: provider configuration, runtime, task wrapper, conversion, audio, codec, and streaming boundaries stay explicit instead of being generated through model inheritance. Its eight cards and matching file table cover Create, Audit, Configure, Wrap, Register, Support, Test, and Document, including optional owned architecture, provenance and legal files, the manifest, training and optimization factories, CPU-safe tests, generated provider documentation, and navigation.

The local page rendered at 1,440 x 900, 1,024 x 768, and 390 x 844 with 804-, 658-, and 342-pixel articles. Light and dark states retained zero document overflow. The process renders as two 392-pixel desktop columns, two 319-pixel tablet columns, and one 342-pixel mobile column. The mobile drawer exposes the active Contribute and Add a model path, and all three wide tables scroll inside their 374-pixel wrappers. The official modular guide rendered at the same three viewports with its current active route, desktop right TOC, version banner, community panel, copy action, table, internally scrollable mobile code, and zero document overflow. Reference dark mode and complete sequential focus remain pending and are not counted as passed.

The initial focused regression failed on the legacy route and seven-card process. The stylesheet assertion then failed before the Material list-rule specificity fix, and the first corrected mobile render exposed a second specificity mismatch before the one-column override was aligned. Two exploratory inventory commands also failed because they assumed nonexistent optimization and model-spec attributes. None of these failed checks is counted as passed.

After correction, the focused contribution, navigation, and stylesheet slice reported three passes and 38 subtests. The complete documentation contract reported 44 passes and 1,310 subtests; the model scaffold contract reported 20 passes and 35 subtests; and the registry plus public optimization slice reported 76 passes and 786 subtests. All 68 model pages and 59 model notebooks remained generator-current, release alignment found five benchmark files and 68 documented providers, and the strict eleven-language documentation build passed. The full suite and distribution probes were not rerun for this navigation, prose, test, and CSS slice; their preceding evidence remains applicable. Exact-commit remote CI, reference dark mode, complete sequential focus, protected publisher configuration, publication approval, and the API reference representative-page mapping remain open. The untracked uv.lock stayed unchanged.

The next bounded representative-page slice maps the current Transformers Models API. The pinned upstream commit remains b3a36037d3feb22e3f0174b3dd4248fcc0f0f722; the 44-line Models source has SHA-256 a4899c758b5d621075b2e2f39f0aa79671010c88b08d1508f7af5731375c2871, and the toctree remains f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a. All 24 current upstream Main Classes entries now map to a VoiceHub route or an explicit unsupported or out-of-domain record.

VoiceHub now exposes API > Main Classes > Models while retaining the existing Full API reference. The representative page documents the shared PreTrainedSpeechModel marker; the TTS, audio, ASR, and VAD task bases; lazy loading, saving, and training transitions; normalized outputs; public source links; and portable versus provider-native artifacts. It records that no public push_to_hub() contract exists rather than treating local saving as a registry-wide sharing pass.

The local page rendered at 1,440 x 900, 1,024 x 768, and 390 x 844 with 804-, 658-, and 342-pixel articles in both VoiceHub palettes. It retained the active nested Models route, the desktop TOC, responsive drawer, contained code, four internally scrolling mobile tables, adjacent page navigation, and zero document overflow. Pointer copy wrote the complete 5,106-character page and all three declared source targets resolved. The official page rendered at the same three viewports with its current Models, PreTrainedModel, ModuleUtilsMixin, and Pushing to the Hub hierarchy, generated signature controls, responsive navigation, internally scrollable mobile code, and zero document overflow.

Three official theme-control clicks did not change the reference palette. The initial local Model outputs TOC click reached y = 68 but left the preceding link active; that failed probe is not a pass. The later shared-shell correction aligns the target and observer thresholds and closes this pointer transition. A local Enter copy probe retained the two-pixel focus outline without changing the clipboard. Reference dark mode, native keyboard copy activation, and complete sequential focus are therefore pending and are not counted as passed.

The initial focused regression failed because API was still a flat entry and the Models page did not exist. After correction it reported one pass and 50 subtests. The complete documentation contract reported 45 passes and 1,379 subtests; the base API, speech-core, and registry slice reported 49 passes and 216 subtests; and the public optimization slice reported 59 passes and 593 subtests. All 68 model pages and 59 model notebooks remained generator-current, release alignment found five benchmark files and 68 documented providers, and the strict eleven-language build passed. An exploratory optimization inventory failed after treating its registry as iterable; the corrected public list() query reported six passes and 408 model/pass pairs. The failed query is not evidence.

The full suite and distribution probes were not rerun for this navigation, documentation, and test slice; their preceding evidence remains applicable. Exact-commit remote CI, reference dark mode, native keyboard copy activation, complete sequential focus, protected publisher configuration, publication approval, and the other representative-page interaction gates remain open. The untracked uv.lock stayed unchanged.

The next bounded shared-shell slice closes the Models API pointer TOC transition without introducing a competing observer. Measurement showed that Material's default hash target stopped at y = 68, while its existing navigation tracker switched to the new heading only at approximately y = 64. The shared target margin now derives from the 65-pixel VoiceHub header and Material's own heading offset, resolving to 64 pixels. The first focused regression failed because that target contract did not exist; the corrected regression passes.

At 1,440 x 900, all four Models API TOC links retained their requested hashes, positioned their headings between y = 63.8 and y = 64.1, and became the sole active link in both VoiceHub palettes with zero document overflow. The mapped Pipeline route preserved the same result for Tasks, Parameters, and Large models, and manual scroll states at the threshold retained one matching active link. At 1,024 x 768 and 390 x 844, the right rail remained hidden with zero width, zero height, and zero document overflow. The official Models reference retained its current hash navigation and zero overflow at the pinned upstream commit. Its inaccessible dark palette remains unpassed.

The complete documentation contract reported 45 passes and 1,379 subtests; the registry slice reported 52 passes and 316 subtests; and the public optimization slice reported 161 passes, 751 subtests, and four warnings. The 13 release and distribution-source contracts passed. All 68 model pages and 59 model notebooks remained generator-current, release alignment found five benchmark files and 68 documented providers, and the strict eleven-language build passed. Fresh wheel, source-distribution, and editable probes reported 68 models, 81 provenance manifests, 193 compliance files, zero dependency violations, and no eager PyTorch import; the wheel measured 57,189,200 bytes and the source distribution 55,439,250 bytes. This CSS and regression slice did not rerun the complete Python suite. Exact-commit remote CI, native keyboard activation, complete sequential focus, reference dark mode, protected publisher configuration, and publication approval remain open.

The next bounded evidence slice closes the responsive SpeechT5 model-detail comparison shared by all 68 generated provider pages. The pinned Transformers commit remains b3a36037d3feb22e3f0174b3dd4248fcc0f0f722; the current 87-line SpeechT5 source has SHA-256 71bba8a2921cf637383fb8d6f2fd66df9cd95deb59118b9f49e1362485c27eb5, and the toctree remains f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a.

The local and official pages rendered at 1,440 x 900, 1,024 x 768, and 390 x 844 with matching 804-, 658-, and 342-pixel article widths. Both palettes were verified on both sites at all three sizes. VoiceHub retained its active SpeechT5 navigation, thirteen-link desktop TOC, collapsed responsive right rail, source and edit targets, eight contained tables, and zero document overflow. Its mobile drawer exposed Base classes, Models, Text to speech, and SpeechT5 in both palettes. Public API and SpeechT5ForTextToSpeech pointer transitions kept their requested hashes, landed at y = 64, and left exactly one matching active TOC link.

The official page retained its active SpeechT5 route, generated API hierarchy, page-copy action, and twelve collapsed signature controls. Activating the first 60-parameter control exposed its full signature. The previously unverified reference dark state is now closed for this page pair: the visible theme menu produced verified dark desktop, tablet, and mobile renders. The official mobile page itself measured a 594-pixel document scroll width and a 570-pixel article scroll width because long generated code and API spans are not contained. VoiceHub intentionally retains internal overflow at 390 pixels instead of copying that upstream defect.

The first focused pytest invocation used the wrong test-class name and collected no tests; the corrected SpeechT5 contract passed. The first inventory invocation used an unavailable python shim; a later exploratory query assumed a nonexistent ModelSpec.model_id attribute and a nonexistent notebook-script name. The corrected virtual-environment inventory reports 68 models (34 TTS, 23 ASR, and 11 VAD), 102 aliases, no invalid display names, 68 provider pages, no missing or orphaned navigation entries, six public optimization passes, and 408 model/pass pairs. The page and notebook generators report 68 and 59 current artifacts, and release alignment finds five benchmark files and 68 documented providers. None of the failed exploratory commands is counted as evidence.

A direct mobile theme-label click timed out because that label is intentionally hidden at the mobile breakpoint. The verified path used the visible tablet theme control and then returned to the exact mobile viewport; this does not establish native mobile keyboard access. After the evidence update, the focused SpeechT5 contract reported one pass; the complete documentation contract reported 45 passes and 1,379 subtests; and the registry contract reported 17 passes and 193 subtests. The 68 model pages and 59 model notebooks remained generator-current, release alignment found five benchmark files and 68 documented providers, and the strict eleven-language documentation build passed.

No source, stylesheet, registry, runtime, or package change was needed. The public optimization suites, distribution probes, and complete Python suite were therefore not rerun; their immediately preceding evidence remains applicable. Native keyboard activation, complete sequential focus, exact-commit remote CI, protected publisher configuration, and publication approval remain open. Other representative pages still retain their own unverified reference-theme states. The untracked uv.lock stayed unchanged.

The next bounded evidence slice closes the exact responsive Auto Classes model index comparison that fronts all 68 registered integrations. The pinned Transformers commit remains b3a36037d3feb22e3f0174b3dd4248fcc0f0f722; the current 301-line Auto Classes source has SHA-256 557f5836c0722fef6a484c46805dfab0eb69a387b028a914b132350edf09f167, and the toctree remains f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a.

The local and official pages rendered at 1,440 x 900, 1,024 x 768, and 390 x 844 with matching 804-, 658-, and 342-pixel articles. Light and dark were verified on both sites at all three sizes. VoiceHub retained the active Auto Classes route, eight-link desktop TOC, collapsed responsive right rail, four contained tables, source/edit controls, Quickstart/Bark adjacent navigation, and zero document overflow. Its mobile drawer exposed Base classes, Models, and Auto Classes in both palettes. Pointer page copy replaced a sentinel with the complete 6,200-character article. Registered models and Voice activity detection pointer transitions kept their requested hashes, landed at y = 64, and left exactly one matching active TOC link.

The official page retained its active Auto Classes route, generated configuration/processor/model hierarchy, desktop TOC, page-copy control, responsive navigation, and zero document overflow. Its #transformers.AutoConfig pointer transition retained the hash and placed the target at y = 36. The community panel, product banner, non-speech class families, and 324,041-pixel generated API article are upstream content differences that VoiceHub does not fabricate. The official copy button did not replace a clipboard sentinel in this harness; that failed interaction is not a pass.

The focused model-index contract reported one pass and 68 subtests before and after the evidence edit. The complete documentation contract reported 45 passes and 1,379 subtests, while the registry contract reported 17 passes and 193 subtests. All 68 model pages and 59 model notebooks remained generator-current, release alignment found five benchmark files and 68 documented providers, and the strict eleven-language documentation build passed.

No source, stylesheet, generator, registry, runtime, or package change was needed. The public optimization suites, distribution probes, and complete Python suite were therefore not rerun; their immediately preceding evidence remains applicable. Native keyboard activation, complete sequential focus, exact-commit remote CI, protected publisher configuration, and publication approval remain open. The untracked uv.lock stayed unchanged.

The next bounded release slice closes the exact responsive and reference-theme Trainer comparison. The pinned Transformers commit remains b3a36037d3feb22e3f0174b3dd4248fcc0f0f722; its 30-line Trainer source has SHA-256 e7c5368c1223c2b195321468fcd1ac64c5cda52a1e44d4adf854a6473b6c9ee0, and the toctree remains f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a.

The local and official Trainer pages rendered at 1,440 x 900, 1,024 x 768, and 390 x 844 with matching 804-, 658-, and 342-pixel articles. Light and dark were verified on both sites at all three sizes. Both retained zero document overflow, the concise Trainer and Next steps hierarchy, one active Trainer overview route when navigation was visible, responsive navigation, page actions, and a Fine-tuning destination. The local mobile drawer and official mobile documentation menu retained their active Trainer overview entries in both palettes.

Local page copy replaced a sentinel with the complete 1,257-character article. The official copy action left its 25-character sentinel unchanged, so it is recorded as failed rather than passed. The official #next-steps pointer transition retained its hash in both palettes but bottom-clamped its short-page target at y = 417 without an active TOC item. The first VoiceHub pointer probe also scrolled but then lost its hash and active item when Material's observer reclassified the short page. A bounded header-control repair now waits for the pointer scroll to settle and preserves the requested hash and sole active TOC link without intercepting navigation. The corrected local transition retains #next-steps in both palettes, marks exactly one matching active link, and bottom-clamps at y = 329. Direct anchored loading remains valid.

The first focused pytest invocation used the wrong test-class name and collected no tests. The corrected pre-change Trainer contract passed, and the post-change Trainer plus TOC contracts reported two passes. Exact responsive and dark evidence did not reveal any local geometry or overflow change. Native keyboard activation and complete sequential focus remain inaccessible in the current driver and are not counted as passed. The official copy failure, exact-commit remote CI, protected publisher configuration, and publication approval remain open.

The complete documentation contract then reported 45 passes and 1,379 subtests; the registry contract reported 17 passes and 193 subtests. All 68 model pages and 59 model notebooks remained generator-current, release alignment found five benchmark files and 68 documented providers, and the strict eleven-language documentation build passed. The selected code, credential, whitespace, and file-integrity pre-commit hooks passed; the Markdown hook had no matching files and is not counted as executed evidence. Public optimization suites, distribution probes, packaging checks, and the complete Python suite were not rerun because the change is confined to the documentation shell and its contract test. Their immediately preceding evidence remains applicable. The untracked uv.lock stayed unchanged.

The next bounded evidence slice closes the exact responsive and reference-dark Optimization comparison. The pinned Transformers commit remains b3a36037d3feb22e3f0174b3dd4248fcc0f0f722; its 178-line Optimization source has SHA-256 19622667a7299f258f5c9a72940c9f26492619636f35d9bd592701c02745b620, and the toctree remains f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a.

The local and official pages rendered at 1,440 x 900, 1,024 x 768, and 390 x 844 with matching 804-, 658-, and 342-pixel articles. Light and dark were verified on both sites at all three sizes. Both retained zero document overflow, an active Overview route when navigation was visible, desktop right TOCs, collapsed responsive rails, page actions, and mobile documentation navigation. The local mobile technique table remained contained by a 374-pixel wrapper with a 752-pixel internal scroll width.

Local page copy replaced a sentinel with the complete 3,730-character article. Its Compilation and Diffusion sampling pointer transitions retained their hashes, landed between y = 64.1 and y = 64.2, and left exactly one matching TOC link active in both palettes. The official Compilation and Caching transitions retained their hashes at y = 35.9 and y = 36.2. The official page-copy control left its 21-character sentinel unchanged, so it remains failed rather than passed. Native keyboard activation and complete sequential focus remain unverified and are not counted as passed. No source, stylesheet, generator, registry, optimization runtime, or package change was needed.

The focused Optimization contract reported one pass and six subtests. The complete documentation contract reported 45 passes and 1,379 subtests, while the registry and all documented public optimization suites reported 178 passes and 944 subtests. One platform-specific PyTorch decomposition warning and three weight-normalization deprecation warnings remain warnings rather than passes or failures. All 68 model pages and 59 model notebooks remained generator-current, release alignment found five benchmark files and 68 documented providers, and the strict eleven-language documentation build passed. The selected file-integrity, whitespace, credential, and case-conflict pre-commit hooks passed; code and Markdown hooks had no matching files and are not counted as executed evidence. Distribution probes and the complete Python suite were not rerun for this evidence-only change. Exact-commit remote CI, protected publisher configuration, publication approval, native keyboard activation, and complete sequential focus remain open. The untracked uv.lock stayed unchanged.

The next bounded evidence slice closes the exact responsive and reference-dark Contribution comparison. The pinned Transformers commit remains b3a36037d3feb22e3f0174b3dd4248fcc0f0f722; its 500-line modular guide has SHA-256 a52af1ee1754e6e658535f071a727fac31a057a3a848c28c18873a5ff371dd96. The current registry-derived contribution inventory remains eight ordered steps: Create, Audit, Configure, Wrap, Register, Support, Test, and Document.

The local and official contribution pages rendered at 1,440 x 900, 1,024 x 768, and 390 x 844 with matching 804-, 658-, and 342-pixel articles. Light and dark were verified on both sites at all three sizes, and every render retained zero document overflow. The local process retained two 392-pixel desktop columns, two 319-pixel tablet columns, and one 342-pixel mobile column. The mobile drawer preserved the active Contribute and Add a model path, and all three local tables remained contained by their 374-pixel scrolling wrappers. The official mobile documentation menu preserved its active modular guide.

Local pointer copy replaced a sentinel with the complete 18,197-character article. Its Register once transition retained #5-register-once, placed the heading at y = 64, and left exactly one matching right-TOC link active. The official copy action replaced its sentinel with the complete 24,443-character Markdown source. Its Generate the modeling files transition retained #generate-the-modeling-files, focused the requested link, and placed the target at y = 36; the official TOC did not expose a separate active-link state. Native keyboard activation and complete sequential focus remain unverified and are not counted as passed. No source, stylesheet, generator, registry, optimization runtime, or package change was needed.

The focused contribution contract passed before the evidence edit with 18 subtests. The post-edit contribution and model-scaffold slice reported 21 passes and 53 subtests. The complete documentation contract reported 45 passes and 1,379 subtests. Registry and all documented public optimization suites reported 305 passes and 1,049 subtests. Five native-kernel checks were skipped: three require VOICEHUB_TEST_TRITON_KERNELS=1 on a Triton CUDA host, and two require VOICEHUB_TEST_CUDA_EXTENSIONS=1 on a CUDA-toolkit host. They remain hardware-limited gates and are not counted as passes. One platform-specific PyTorch decomposition warning and seven weight-normalization deprecation warnings remain warnings rather than passes or failures. All 68 model pages and 59 model notebooks remained generator-current, release alignment found five benchmark files and 68 documented providers, and the strict eleven-language documentation build passed. The selected file-integrity, whitespace, credential, and case-conflict pre-commit hooks passed; code and Markdown hooks had no matching files and are not counted as executed evidence. Distribution probes and the complete Python suite were not rerun for this evidence-only change. Exact-commit remote CI, protected publisher configuration, publication approval, native keyboard activation, complete sequential focus, and the five native-kernel hardware gates remain open. The untracked uv.lock stayed unchanged.

The next bounded evidence slice closes the exact responsive and reference-dark Models API comparison. The pinned Transformers commit remains b3a36037d3feb22e3f0174b3dd4248fcc0f0f722; its 44-line Models source retains SHA-256 a4899c758b5d621075b2e2f39f0aa79671010c88b08d1508f7af5731375c2871, and the current Main Classes inventory remains 24 entries with mapped or explicitly non-applicable VoiceHub dispositions.

The local and official Models pages rendered at 1,440 x 900, 1,024 x 768, and 390 x 844 with matching 804-, 658-, and 342-pixel articles. Light and dark were verified on both sites at all three sizes, and every render retained zero document overflow. The local page preserved its active Models path, four-link desktop TOC, edit/copy/adjacent navigation, three distinct public source targets, and four mobile tables inside 374-pixel scrolling wrappers. The mobile drawer retained API, Main Classes, Models, and Full API reference in both palettes.

Local pointer copy replaced a sentinel with the complete 5,106-character article. Its Model outputs transition retained #model-outputs, placed the heading at y = 64, and left exactly one matching TOC link active. Official pointer copy replaced its sentinel with the complete 49,539-character Markdown source. Its PreTrainedModel transition retained #transformers.PreTrainedModel, focused the requested link, and placed the target at y = 36. The first collapsed generated signature exposed all 14 hidden parameters, and the light and dark official mobile menus preserved their active Models item. Native keyboard copy activation and complete sequential focus remain unverified and are not counted as passed. No source, stylesheet, generator, registry, runtime, optimization, or package change was needed.

The focused Models API contract passed before and after the evidence edit with 50 subtests. The complete documentation contract reported 45 passes and 1,379 subtests, while the base API, speech-core, and registry slice reported 49 passes and 216 subtests. All 68 model pages and 59 model notebooks remained generator-current, release alignment found five benchmark files and 68 documented providers, and the strict eleven-language documentation build passed. The selected file-integrity, whitespace, credential, and case-conflict pre-commit hooks passed; code and Markdown hooks had no matching files and are not counted as executed evidence. The public optimization suites, distribution probes, packaging checks, and complete Python suite were not rerun for this evidence-only change; their immediately preceding evidence remains applicable. Exact-commit remote CI, protected publisher configuration, publication approval, native keyboard activation, complete sequential focus, and the five native-kernel hardware gates remain open. The untracked uv.lock stayed unchanged.

The next bounded evidence slice closes the stale Goal-level universal optimization lifecycle audit. The live public registry remains six passes— codec-kernels, compile, custom-kernels, diffusion-cache, diffusion-sampling, and flash-attention-4—across 68 registered models, for 408 model/pass pairs. The audit derives both dimensions from the public registries; it contains no provider or pass allowlist in runtime code.

Every pair executed through apply_optimization_plan() on a CPU-safe normalized TTS, ASR, or VAD semantic runtime. The contract checks validation before mutation, the resolved architecture context, an explicit non-skipped outcome, reason text for universal fallbacks, manifest reporting, strict JSON serialization, output type and value preservation, checkpoint-state preservation, deterministic restoration, result cleanup, and post-restore semantic/state equality. All 68 registered classes separately retain the five public lifecycle methods from BaseSpeechModel rather than overriding them with provider-specific behavior.

The executed inventory reported 68 reasoned eager-fallback outcomes for the CPU compile path and 340 reasoned not-applicable outcomes for the five architecture-specific passes. No pair reported skipped, and no fallback omitted its reason. These universal fallbacks are lifecycle support rather than acceleration claims. The separate pass-specific suites exercise real configured or compiled paths, state-key safety, semantic behavior, failure rollback, manifest data, and restoration.

The narrow 408-pair contract reported one pass and 408 subtests; the shared 68-model lifecycle-owner contract reported one pass and 68 subtests. The full native optimization lifecycle file reported 29 passes and 482 subtests with no skips. The registry and all public pass-specific CPU suites reported 305 passes and 1,049 subtests. Five opt-in checks were skipped and remain unpassed: three require VOICEHUB_TEST_TRITON_KERNELS=1 on a Triton CUDA host, and two require VOICEHUB_TEST_CUDA_EXTENSIONS=1 on a CUDA-toolkit host. One platform-specific PyTorch decomposition warning and seven weight-normalization deprecation warnings remain warnings rather than passes or failures.

The complete documentation contract reported 45 passes and 1,379 subtests. All 68 model pages and 59 model notebooks remained generator-current, release alignment found five benchmark files and 68 documented providers, and the strict eleven-language documentation build passed. The selected file-integrity, whitespace, credential, and case-conflict pre-commit hooks passed; code hooks had no matching files and are not counted as executed evidence. No model, runtime, optimization, registry, generator, package, or test code changed, so distribution probes, packaging checks, and the complete Python suite were not rerun; their immediately preceding evidence remains applicable. Exact-commit remote CI, protected publisher configuration, publication approval, native keyboard activation, complete sequential focus, and the five native-kernel hardware gates remain open. The untracked uv.lock stayed unchanged.

The next bounded packaging-evidence slice replaces a stale single source- distribution size with two fresh current-worktree builds. The focused distribution and release contracts reported 13 passes. The complete scripts/check_distribution.py probe built an isolated wheel and source distribution, installed the wheel, source distribution, and editable checkout into separate dependency-free virtual environments, and reported version 0.3.0, 68 models, 81 provenance manifests, 193 compliance files, all required representative package data, zero runtime dependency violations, and no eager PyTorch import in every installation mode. That build produced a 57,189,200- byte wheel and a 55,439,867-byte source distribution.

A second isolated build passed scripts/check_release.py --dist-dir with the same source, documentation, five-file benchmark, 34 TTS, 23 ASR, 11 VAD, and 68-provider alignment. It produced the same 57,189,200-byte wheel and a 55,438,790-byte source distribution. The source-distribution byte difference is expected from gzip build timestamps and is not presented as reproducible artifact identity; the tagged workflow must retain the exact published pair and record its hashes. Both fresh artifact pairs remained below PyPI's 100 MB per-file limit. Neither build used --with-dependencies, so dependency resolution and runtime execution are not claimed by this slice; the supported- platform CI and complete CPU-safe suite remain the applicable runtime evidence.

No source, packaging configuration, runtime, model, test, or generated file changed in this slice. Existing repository-local dist/ artifacts were not overwritten. Exact-commit CI for the uncommitted worktree, the protected PyPI environment and trusted publisher, tag and publication approval, five native- kernel hardware gates, seven inaccessible asset/checkpoint/oracle paths, native documentation keyboard activation, and complete sequential focus remain open. The untracked uv.lock stayed unchanged.

The next bounded public-lifecycle slice corrects AutoProcessor.from_pretrained() option routing. Previously, configuration loader values such as revision and local_files_only were omitted from AutoConfig.from_pretrained() and then leaked into the processor constructor or restored processor_config.json. Explicit model_type construction had the same leak and ignored configuration overrides. The factory now accepts the same explicit config_kwargs separation as VoiceHub's auto-model factories, validates the mapping before loading, rejects ambiguous config plus config_kwargs, and reserves ordinary keyword arguments for the processor.

The first focused regression executed before the implementation and reported three failures: remote configuration options were omitted, explicit-model-type configuration overrides were ignored, and local artifact restoration retained the leaked config_kwargs. After correction, all three regressions passed; the complete focused file reported 10 passes. The broader auto, base API, speech-core, task-registry, pipeline, and registry slice reported 90 passes and 298 subtests. The two registry-derived optimization lifecycle contracts reported two passes and 476 subtests. A fresh subprocess constructed the registered Kokoro processor without importing PyTorch, and the public root inventory remained 261 exports with no eager PyTorch import.

The complete Python 3.12.12 suite reported 2,471 passes, 15 explicit skips, 3,729 subtests, and 35 warnings in 112.60 seconds. Those skips remain unpassed. The documentation contract reported 45 passes and 1,379 subtests; the combined documentation, release, and canonical-guidance slice reported 58 passes and 1,390 subtests; all 68 model pages remained generator-current; and the strict eleven-language build passed. The distribution and release source contracts reported 23 passes and 76 subtests, and release alignment found five benchmark records plus all 68 documented providers. The fresh wheel, source-distribution, and editable probe reported version 0.3.0, 68 models, 81 provenance manifests, 193 compliance files, all required representative package data, zero runtime dependency violations, and no eager PyTorch import. The wheel measured 57,189,217 bytes and the source distribution 55,440,131 bytes.

Every applicable selected pre-commit hook passed. Markdown formatting found no matching files and is not counted. The official Transformers main reference remained b3a36037d3feb22e3f0174b3dd4248fcc0f0f722, its navigation SHA-256 remained f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a, and the documentation endpoint returned HTTP 200. The refreshed inventories remain 68 models (34 TTS, 23 ASR, and 11 VAD), 102 aliases, 68 provider pages, six public optimization passes, 408 model/pass pairs, five benchmark records, and no missing, orphaned, or lowercase-first model pages. Exact-commit remote CI for the uncommitted worktree, the protected PyPI environment and publisher, tag and publication approval, native documentation keyboard activation, complete sequential focus, five native-kernel hardware gates, and seven inaccessible asset/checkpoint/oracle paths remain open. The untracked uv.lock stayed unchanged.

The following bounded public-lifecycle slice completes the processor artifact half of AutoProcessor.from_pretrained(). The pinned Transformers AutoProcessor source delegates the original local directory or Hub identifier to the selected processor class after resolving its configuration. VoiceHub previously delegated only an existing local directory containing processor_config.json; Hub identifiers, direct processor configuration files, and missing-optional-artifact fallback never reached the processor loader.

VoiceHub now resolves the registered processor class without constructing a temporary instance, delegates every source form to that class's from_pretrained() method, and reuses subfolder, cache_dir, revision, token, and local_files_only from configuration loading when the processor call does not override them. The base processor consumes those loader-only values before either its direct-file or resolved-file branch, so credentials and cache settings cannot become serialized processor state. A missing optional base processor_config.json remains an explicit constructor fallback.

The first focused regression reported three failures before implementation: the remote processor loader was never called, a missing remote processor artifact never exercised fallback, and a direct processor configuration file was ignored. An expanded direct-file regression then failed because its token reached a temporary processor constructor. Neither failed run is counted as a pass. The corrected focused file reported 13 passes. The broader auto, base API, speech-core, task-registry, registry, and Hub-transport slice reported 105 passes and 295 subtests. The two registry-derived optimization lifecycle contracts reported two passes and 476 subtests.

The complete Python 3.12.12 suite reported 2,474 passes, 15 explicit skips, 3,729 subtests, and 35 warnings in 112.17 seconds; the skips remain unpassed. The combined documentation, release, canonical-guidance, distribution-source, and packaging-metadata slice reported 72 passes and 1,466 subtests. All 68 model pages remained generator-current, release alignment found five benchmark records and all 68 documented providers, and the strict eleven-language build passed. A fresh wheel, source-distribution, and editable probe reported version 0.3.0, 68 models, 81 provenance manifests, 193 compliance files, all required representative package data, zero runtime dependency violations, and no eager PyTorch import. The wheel measured 57,189,342 bytes and the source distribution 55,440,749 bytes.

Every applicable selected pre-commit hook passed; Markdown formatting found no matching files and is not counted. Transformers main remained b3a36037d3feb22e3f0174b3dd4248fcc0f0f722. The pinned processing_auto.py SHA-256 is c1e82fbd511b4fb3838cae570a29716d54903c0658c4e37de10e17fba8583cad, the pinned processing_utils.py SHA-256 is 77386cb128481c3bc5bd330f6a6687d69822e9eacb8072d2f3c3fc930f426483, and the navigation SHA-256 remains f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a. The refreshed inventories remain 68 models (34 TTS, 23 ASR, and 11 VAD), 102 aliases, 68 provider pages, six public optimization passes, 408 model/pass pairs, five benchmark records, and no missing, orphaned, or lowercase-first model pages. Exact-commit remote CI for the uncommitted worktree, the protected PyPI environment and publisher, tag and publication approval, native documentation keyboard activation, complete sequential focus, five native-kernel hardware gates, and seven inaccessible asset/checkpoint/oracle paths remain open. The untracked uv.lock stayed unchanged.

The next bounded lazy-discovery slice removes model-wrapper imports from registry-wide processor selection. The first fresh-process inventory constructed and restored all 68 processors successfully, but 13 registrations imported PyTorch merely to discover their processor class. That inventory exited nonzero and is not counted as a pass. ModelSpec now records a validated lazy processor import target, supplies the shared TTS or audio processor by task when no override is declared, and preserves a custom importable processor_class through the recommended extension registrar. AutoProcessor resolves that metadata directly instead of importing a model wrapper.

The focused processor-metadata contracts reported three passes and 68 subtests. The broader task-registry, registry, auto-configuration, base API, and speech-core slice reported 88 passes and 358 subtests. The two registry-derived optimization lifecycle contracts reported two passes and 476 subtests. A post-change fresh-process inventory constructed and offline-loaded the task-appropriate processor for all 68 registrations, exercised the text or audio input envelope, and imported none of PyTorch, Transformers, or the named optional speech backends. The 68 generated pages and navigation remained current. Documentation, release, distribution-source, and packaging contracts reported 68 passes and 1,455 subtests, and the strict eleven-language build passed. A local 1,280-by-720 browser render showed the new registry metadata on the light API page and the eight-step contribution guide in dark mode with no document overflow. The in-app surface offered no viewport-resize control, so the 1,440-, 1,024-, and 390-pixel comparisons were not rerun and no new responsive-parity claim is made.

The complete Python 3.12.12 suite reported 2,476 passes, 15 explicit skips, 3,797 subtests, and 35 warnings in 118.53 seconds; the skips remain unpassed. The fresh wheel, source-distribution, and editable probe reported version 0.3.0, 68 models, 81 provenance manifests, 193 compliance files, all required representative package data, zero runtime dependency violations, and no eager PyTorch import. The wheel measured 57,189,672 bytes and the source distribution 55,442,882 bytes. Every applicable selected pre-commit hook passed; Markdown formatting found no matching files and is not counted.

A staged fresh-process probe intentionally combined AutoConfig.for_model() with AutoProcessor.from_config() and still exited nonzero because seven configuration facades import PyTorch: llasa, f5tts, parlertts, zonos, zonos2, csm, and asr_seamless_m4t_v2. The processor-only phase of that probe imported no additional heavy backend, so those seven failures are a separate configuration-facade gap rather than processor discovery evidence. They remain unpassed. Exact-commit remote CI for the uncommitted worktree, the protected PyPI environment and publisher, tag and publication approval, native documentation keyboard activation, complete sequential focus, five native-kernel hardware gates, and seven inaccessible asset/checkpoint/oracle paths also remain open. The untracked uv.lock stayed unchanged.

The next bounded lazy-configuration slice closes that seven-model gap. The first registry-wide regression and per-model fresh-process inventory both exited nonzero because AutoConfig.for_model() imported PyTorch for llasa, f5tts, parlertts, zonos, zonos2, csm, and asr_seamless_m4t_v2; neither failing check is counted as a pass. Four model packages exported runtime objects eagerly, five public config classes lived in Torch-backed inference modules, and the SeamlessM4T-v2 config reached its language table through a Torch-backed tokenizer.

The five config classes now live in their dependency-light configuration modules, the affected package exports resolve lazily, and the SeamlessM4T-v2 language table lives in a framework-free metadata module shared with its tokenizer. Runtime wrappers import those same config class objects, so model construction, serialization, checkpoint loading, and public compatibility paths retain one class identity. This follows the pinned Transformers lazy configuration mapping rather than adding a provider branch to AutoConfig.

The corrected registry-wide config regression reported one pass after its initial failure. A separate post-change process for each of all 68 registered models reported zero PyTorch, Transformers, or named optional speech-backend imports. The affected LLaSA, F5-TTS, Parler-TTS, Zonos, ZONOS2, CSM, and SeamlessM4T-v2 suites reported 168 passes, 15 subtests, and eight warnings. The broader task-registry, registry, auto-configuration, base API, and speech-core slice reported 89 passes and 358 subtests. Native dependency and shared provider-independence policies plus the two registry-derived optimization lifecycle contracts reported 25 passes and 476 subtests.

The complete Python 3.12.12 suite reported 2,477 passes, 15 explicit skips, 3,797 subtests, and 35 warnings in 109.21 seconds; the skips remain unpassed. All 68 model pages and navigation entries remained generator-current. The documentation, release, distribution-source, and packaging contracts reported 68 passes and 1,455 subtests, and the strict eleven-language build passed. A fresh wheel, source-distribution, and editable probe reported version 0.3.0, 68 models, 81 provenance manifests, 193 compliance files, all required representative package data, zero runtime dependency violations, and no eager PyTorch import. The wheel measured 57,192,256 bytes and the source distribution 55,444,113 bytes.

Every applicable selected pre-commit hook passed; Markdown formatting found no matching files and is not counted. The official Transformers main revision remained b3a36037d3feb22e3f0174b3dd4248fcc0f0f722; its configuration_auto.py SHA-256 was 5eaeb741896b05e852fcca8d316b329e42fa3788e2017c560d2e95002b84f0e9, and the official Auto Classes endpoint returned HTTP 200. Exact-commit remote CI for the uncommitted worktree, the protected PyPI environment and publisher, tag and publication approval, native documentation keyboard activation, complete sequential focus, five native-kernel hardware gates, and seven inaccessible asset/checkpoint/oracle paths remain open. The untracked uv.lock stayed unchanged.

The next bounded documentation-shell slice closes the shared page-copy keyboard gap. Before correction, a rendered Enter press left its clipboard sentinel unchanged. The focused source contract then failed after being strengthened to require an explicit, single keyboard copy path. The first focused command used an unavailable bare python executable and did not run; the corrected virtual-environment invocation produced the expected pre-change failure. Neither result is counted as a pass.

The shared action now handles Enter and Space directly, prevents the browser's second native activation, and invokes the existing copy lifecycle exactly once. It retains the selection fallback, aria-busy and live success state, focus restoration, and pointer path without adding a page- or provider-specific branch. The focused contract passed after correction.

At the available 1280 x 720 browser viewport, Enter and Space each replaced a fresh sentinel with the complete 5,252-character Quickstart article. Both activations returned aria-busy to false, reported Copied, and left the button focused with its focus-visible class. Enter repeated the same result in the dark palette with zero horizontal overflow. Pointer activation still copied all 5,252 characters and removed the keyboard-only focus class. The official Quicktour at Transformers commit b3a36037d3feb22e3f0174b3dd4248fcc0f0f722 exposed its corresponding action as a button, but Enter left its 18-character sentinel unchanged, so that probe is recorded as failed rather than passed. Its pinned source SHA-256 remains ecfb99781204bcaea1ca63bcb4ad9ef70c99812e5f965a49b29de23cedd25bd7.

The in-app browser did not expose viewport resizing. This JavaScript-only correction changes no markup or stylesheet, so the immediately preceding exact 1440 x 900, 1024 x 768, and 390 x 844 geometry remains applicable; no fresh exact-viewport keyboard claim is made. The complete documentation contract reported 45 passes and 1,379 subtests. All 68 model pages and navigation entries remained generator-current, release alignment found five benchmark records and all 68 documented providers, and the strict eleven-language build passed. Every applicable selected pre-commit hook for the changed script and contract test passed; Markdown formatting had no matching files and is not counted.

Registry runtime, optimization, package, and complete-suite checks were not rerun because the bounded change is confined to the shared documentation action and its source/evidence contracts. Their immediately preceding evidence remains applicable: 68 models, six public optimization passes, 408 model/pass pairs, five benchmark records, a 2,477-pass complete suite, and the fresh 57,192,256-byte wheel plus 55,444,113-byte source distribution. Exact-commit remote CI for the uncommitted worktree, native Enter activation of focused right-TOC links, complete sequential focus, the protected PyPI environment and publisher, tag and publication approval, five native-kernel hardware gates, and seven inaccessible asset/checkpoint/oracle paths remain open. The untracked uv.lock stayed unchanged.

The next bounded documentation-shell slice closes the shared right-table-of- contents Enter gap. Before correction, Enter left the local Parameters link focused but retained an empty hash, no active TOC link, and a Parameters heading more than 2,700 pixels below the viewport. Pointer activation already preserved #parameters, one matching active link, and the y = 63.8 heading offset. The official Pipeline link likewise stayed focused without navigating under the same Enter probe, so neither pre-change keyboard result is a pass.

The strengthened source regression failed before the handler existed and passed after correction. The shared secondary rail now handles only unmodified Enter, prevents a second native activation, delegates to its existing click path, and restores focus without scrolling. Modified Ctrl+Enter remains unintercepted. No page, provider, heading, or route allowlist was added.

At the available 1280 x 720 desktop viewport, Enter on Parameters in the dark palette retained #parameters, settled the heading at y = 63.8, left exactly that link active and focused, and rendered its solid two-pixel outline with a two-pixel offset. Enter on Large models repeated the result in the light palette at y = 63.9. Pointer activation of Tasks still produced #tasks, one matching active link, and y = 64.0. All checked states retained zero horizontal overflow. The first 900-millisecond post-change probe captured an intermediate y = 128.8 state and the preceding active link before the existing settling timers completed; it is not counted as final evidence. The later settled probes above are the passing state.

The in-app browser again exposed no viewport resizing. This shared JavaScript change alters no markup or CSS, and the right rail is already verified as hidden at 1024 and 390 pixels, so the preceding exact responsive geometry remains applicable; no fresh 1440-pixel Enter claim is made. The official Transformers main revision remained b3a36037d3feb22e3f0174b3dd4248fcc0f0f722, its navigation SHA-256 remained f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a, the Pipeline source SHA-256 was 80a95da29bb1bc960d7570065322a267c47052b132b7a5f8130b77e809d3914f, and the rendered documentation endpoint returned HTTP 200.

The focused contract passed; the complete documentation contract reported 45 passes and 1,379 subtests. All 68 model pages and navigation entries remained generator-current, release alignment found five benchmark records and all 68 documented providers, and the strict eleven-language build passed. Every applicable selected pre-commit hook for the changed script and contract test passed; Markdown formatting had no matching files and is not counted.

Registry runtime, optimization, package, and complete-suite checks were not rerun because this bounded change is confined to one shared documentation interaction and its source/evidence contracts. Their immediately preceding evidence remains applicable: 68 models, 102 aliases, six public optimization passes, 408 model/pass pairs, five benchmark records, a 2,477-pass complete suite, and the fresh 57,192,256-byte wheel plus 55,444,113-byte source distribution. Complete sequential focus is now the remaining shared documentation keyboard gate. Exact-commit remote CI for the uncommitted worktree, the protected PyPI environment and publisher, tag and publication approval, five native-kernel hardware gates, and seven inaccessible asset/checkpoint/oracle paths remain open. The untracked uv.lock stayed unchanged.

The next bounded documentation-shell slice removes inactive left-navigation descendants from the desktop focus inventory. Before correction, the unchecked Base classes branch alone exposed 88 focusable descendants, including all 68 model pages, before the current Pipeline article. The strengthened source contract first failed because the desktop shell had no collapse rule. Moving the existing grid, opacity, and visibility treatment to desktop made that source contract pass, but the rendered inventory still exposed every inactive descendant because descendant visibility rules overrode the parent. The contract was strengthened again to require actual display removal and failed against that intermediate implementation. A later accessibility strengthening also failed before section labels became operable controls. None of those failed or intermediate results is counted as a pass.

Unchecked desktop navigation branches now use display: none; checked and indeterminate branches use display: block. Section labels are pointer- operable controls, expose aria-controls, synchronize aria-expanded, and handle unmodified Enter and Space through one shared initializer. The focused regression passed after correction. At the available 1280 x 720 desktop viewport, both VoiceHub palettes exposed 63 visible focusable elements, including the version summary, and zero focusable descendants inside inactive branches. All eight root controls remained visible; the checked Inference and Pipeline API branches retained Pipeline, Speech recognition, and VAD, while unchecked Serving descendants remained hidden. Pointer activation expanded Base classes, synchronized aria-expanded, revealed exactly Models, Preprocessors, and Architecture, and left nested Models collapsed. The local article began at x = 318 and measured 644 pixels, matching the official Pipeline page at this viewport, and every checked state retained zero horizontal overflow.

Native Tab could not be replayed: the in-app browser, DOM keyboard surface, and locator surface did not advance focus. Locator-driven Enter and Space on a branch label also failed because the browser reported that its focused input target no longer matched the resolved locator. The source-tested handler is therefore not rendered keyboard evidence; complete sequential focus and rendered branch-key activation remain pending. The in-app surface exposed no viewport resizing, so no fresh exact 1440 x 900, 1024 x 768, or 390 x 844 claim is made. The complete documentation contract reported 45 passes and 1,379 subtests. Every applicable selected pre-commit hook for the stylesheet, shared script, contract test, and evidence passed. All 68 model pages remained generator-current, release alignment found five benchmark records and all 68 documented providers, and the strict eleven-language build passed.

Registry runtime, optimization, package, and complete-suite checks were not rerun because this bounded change is confined to the shared documentation shell and its source/evidence contracts. Their immediately preceding evidence remains applicable: 68 models, 102 aliases, six public optimization passes, 408 model/pass pairs, five benchmark records, a 2,477-pass complete suite, and the fresh 57,192,256-byte wheel plus 55,444,113-byte source distribution. Exact-commit remote CI for the uncommitted worktree, native sequential focus, rendered branch-key activation, the protected PyPI environment and publisher, tag and publication approval, five native-kernel hardware gates, and seven inaccessible asset/checkpoint/oracle paths remain open. The untracked uv.lock stayed unchanged.

The next bounded documentation-validation slice turns the representative-route navigation matrix into a tagged-candidate gate. Native Tab replay was attempted first, but the browser remained focused on BODY; complete sequential focus therefore remains unavailable and unpassed. A rendered audit then checked Home, Installation, Quickstart, Pipeline, Auto Classes, SpeechT5, Trainer, Optimization overview, Add a model, and Models API at the available 1280 x 720 viewport. In both VoiceHub palettes, every route exposed exactly one expected active primary-navigation link, its exact checked ancestors, all eight root controls, zero focusable descendants inside inactive branches, and zero horizontal overflow.

The initial browser audit failed because its evaluation used an unavailable HTMLElement constructor; the corrected node-type check produced the route matrix. The focused source contract then failed before the post-build checker and workflow steps existed. The checker's first site execution failed because it treated Material's active-page __toc control as a product branch and expected a branch tabindex. The corrected implementation scopes branch checks to __nav_, retaining every product-navigation toggle while excluding the separate page-TOC control. The first selected pre-commit run also failed after YAPF modified the new script and docformatter exited nonzero. None of these failed or intermediate runs is counted as passing evidence.

The dependency-free checker parses the ten generated English pages and requires their expected H1, exact eight-root navigation order, sole active anchor, exact checked ancestor sequence, focusable branch labels, and panel aria-expanded state synchronized with each toggle. Documentation CI and the tagged release workflow now run it immediately after the strict build. The corrected focused contract passed, and the post-build validator reported ten representative routes and all eight roots. The complete documentation contract reported 46 passes and 1,379 subtests, the strict eleven-language build plus the post-build validator passed, the release-workflow contract reported nine passes, and every applicable selected hook passed after formatting. Markdown formatting had no matching files and is not counted.

All 68 model pages remained generator-current, release alignment found five benchmark records and all 68 documented providers, and the refreshed inventories remain 34 TTS, 23 ASR, 11 VAD, 102 aliases, six public optimization passes, 408 model/pass pairs, and eight contribution steps. The official Transformers main revision remained b3a36037d3feb22e3f0174b3dd4248fcc0f0f722, its navigation SHA-256 remained f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a, the current Installation source SHA-256 was d050e3e0e1c89d543c71c25367a455a9b49ea89c92f8c2376bb58294a4a4cf3b, and the rendered endpoint returned HTTP 200.

Runtime, optimization-lifecycle, distribution, and complete-suite checks were not rerun because this slice changes only documentation validation, its source contract, and the two workflows that consume the already-built site. Their immediately preceding evidence remains applicable: a 2,477-pass complete suite and fresh 57,192,256-byte wheel plus 55,444,113-byte source distribution. The validator proves rendered static structure and ARIA state; it does not prove pixel geometry or native keyboard order. Exact-commit remote CI for the current uncommitted worktree, native sequential focus, rendered branch-key activation, fresh exact 1440/1024/390 checks, protected publisher configuration, tag and publication approval, five native-kernel hardware gates, and seven inaccessible asset/checkpoint/oracle paths remain open. The untracked uv.lock stayed unchanged.

The next bounded documentation-validation slice adds executable responsive geometry and palette evidence for the same ten representative routes. The checker pins Playwright 1.62.0 and Chromium 151.0.7922.34 revision 1234, then renders each route in both VoiceHub palettes at 1,440 x 900, 1,024 x 768, and 390 x 844. The resulting 60 cases require exact article, header, primary-rail, secondary-rail, and closed mobile-drawer geometry; the expected colors, H1, active link, checked ancestors, and eight navigation roots; zero horizontal overflow; and zero inactive desktop/tablet focusables. Documentation CI and the tagged release workflow install the pinned Chromium runtime and execute the checker after the strict build and static DOM validator.

The pre-implementation focused contract failed because the visual checker did not exist. Its first post-implementation run failed on an over-specific test fragment, the first selected hook run exited nonzero after YAPF formatted the new script, and a scratch forced palette-checkbox action failed because the hidden input was outside the viewport. The corrected checker uses the theme control's change-event path. None of these failed or intermediate results is counted as passing evidence.

The final focused contract and every applicable selected hook passed. The strict eleven-language build completed; the structural validator reported ten routes and eight roots; and the responsive validator passed all 60 cases. The complete documentation contract reported 47 passes and 1,379 subtests. The related documentation, release, distribution, and packaging slice reported 70 passes and 1,455 subtests. The complete Python 3.12.12 suite reported 2,479 passes, 15 skipped, 3,797 subtests, and 35 warnings in 112.83 seconds. The fresh wheel, source-distribution, and editable probe reported 68 models, 81 provenance manifests, 193 compliance files, required package data, zero runtime dependency violations, and no eager PyTorch import; the wheel measured 57,192,274 bytes and the source distribution 55,444,958 bytes.

All 68 model pages and 59 notebooks remained generator-current, release alignment found five benchmark records and all 68 documented providers, and the refreshed inventories remain 34 TTS, 23 ASR, 11 VAD, 102 aliases, six public optimization passes, 408 model/pass pairs, five benchmark records, eight contribution steps, and ten representative routes. The official Transformers revision remained b3a36037d3feb22e3f0174b3dd4248fcc0f0f722, its navigation SHA-256 remained f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a, and the official Pipeline endpoint returned HTTP 200.

Computed geometry and palette validation is not screenshot pixel comparison or native keyboard evidence. At that point native sequential Tab and rendered Enter/Space branch activation remained open; the following slice closes those two executable gates. Exact-commit remote CI, protected publisher configuration, tag and publication approval, five native-kernel hardware gates, and seven inaccessible asset/checkpoint/oracle paths remain open. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

The next bounded documentation-accessibility slice converts the two remaining shared keyboard gaps into a tagged-candidate gate. The first native Chromium probe reached a zero-height search result region and invisible palette radio, skipped all eight root branch labels because their generated tabindex was empty, and left Base classes collapsed after Enter. Those product failures are not counted as passing evidence.

The shared correction removes the search result viewport and palette radios from sequential focus, normalizes branch labels to tabindex="0", adds their two-pixel visible focus outline, and replaces click delegation with one capture- phase Enter/Space toggle plus the existing bubbling change lifecycle. The Playwright checker now traverses the full Pipeline page from document start in both palettes at desktop and mobile widths. Four cycles contain 228 stable native Tab stops, enter no invisible element or inactive branch, visit the exact header, root navigation, expanded Inference descendants, right table of contents, article, previous and next, and footer order, reach BODY, and return to the skip link. A second complete replay contained 230 stops because Material's scroll-dependent Back to top control became focusable in two cycles; both conditional stops were visible and passed the same checks.

Separate default-palette Enter and slate-palette Space cases reach Base classes through nine native Tabs. Each key opens and closes the branch while retaining focus and the visible outline, synchronizing checked state, aria-expanded, and block/none panel display, preserving the sole active Pipeline route, and retaining zero overflow. An inspected 1,440 x 900 slate screenshot shows the focused control inside the unchanged three-column shell. The official Transformers Pipeline page returned HTTP 200, and its first 25 native Tab stops at the same viewport were all visible.

The responsive continuation makes the mobile menu label keyboard-operable and synchronizes its accessible name, controls, and expanded state. Closed mobile navigation is inert and the closed search input leaves sequential focus. Enter and Space open the drawer in separate palettes, wait until the first navigation control is fully on canvas, and retain zero overflow; Escape closes it, restores inert state, and returns focus to the trigger. The opened 390 x 844 screenshot was inspected with the Pipeline drawer visible.

The pre-implementation focused regression failed against the old markup and control contract. A first corrected source run then over-counted one Jinja loop declaration, the first selected hook run exited nonzero after pyupgrade rewrote the JavaScript evaluation string, and two rendered runs started after the theme control because the palette setup intentionally restored focus there. The final checker reloads the persisted palette before beginning from document start. None of those failed or intermediate runs is counted as a pass.

The mobile audit then failed on the old off-canvas drawer and search sequence, and its focused source regression failed before the inert and trigger lifecycle existed. Early rendered checks applied viewport containment to long article stops that the automation did not scroll and sampled the opening transition before the drawer reached x = 0. These harness failures are not passes; the final check scopes viewport containment to fixed shell controls and waits for the transition to finish.

The two focused source regressions, all applicable selected hooks, strict eleven-language build, ten-route DOM check, 60 responsive cases, eight keyboard cases, and 228 stable focus stops passed; a repeat also validated two conditional Back to top stops. The complete documentation and release- contract slice reported 57 passes and 1,379 subtests. The complete Python 3.12.12 suite reported 2,480 passes, 15 skipped, 3,797 subtests, and 35 warnings in 113.79 seconds. All 68 model pages and 59 notebooks remained generator- current, release alignment found five benchmark records and all 68 documented providers, and the refreshed inventories remain 34 TTS, 23 ASR, 11 VAD, 102 aliases, six public optimization passes, 408 model/pass pairs, five benchmark records, eight contribution steps, and ten representative routes.

This slice changes no package runtime or metadata, so distribution probes were not rerun; the immediately preceding wheel, source-distribution, and editable evidence remains applicable. At that point screenshot pixel comparison, exact- commit remote CI, protected publisher configuration, tag and publication approval, five native-kernel hardware gates, and seven inaccessible asset/checkpoint/oracle paths remain open. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Screenshot-derived visual regression gate

The next bounded documentation slice replaces the local screenshot-pixel gap with a reproducible regression gate for all ten representative routes. The official reference remains Transformers commit b3a36037d3feb22e3f0174b3dd4248fcc0f0f722, with navigation SHA-256 f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a. The official Pipeline endpoint returned HTTP 200 when the reference was refreshed.

Each of the existing 60 route, viewport, and palette cases now captures an in-memory PNG after fonts settle and animations are disabled. Pillow 12.3.0 verifies the exact 1,440 x 900, 1,024 x 768, or 390 x 844 raster dimensions, applies a two-pixel Gaussian blur, and derives an aspect-preserving 64-pixel- wide difference hash plus mean RGB values. The checked-in schema-one JSON manifest contains exactly 60 signatures generated by Playwright 1.62.0 and Chromium 151.0.7922.34. Default validation rejects missing, orphaned, malformed, dimension-mismatched, more than 8%-different perceptual hashes, or mean channel deltas above 6.0. Baseline generation is an explicit CLI mode that prints JSON for review; neither documentation nor tagged-release CI can update it.

The pre-implementation focused regression failed because the baseline artifact did not exist. The first selected hook run then let YAPF format the Python source and exited nonzero; its formatted rerun passed. An initial baseline- application attempt failed before changing the tree because the generated patch lacked its final newline. These failed tooling runs are not passing evidence.

The corrected source contract passed. The rebuilt site then passed all 60 DOM- geometry cases, all 60 screenshot comparisons, eight keyboard cases, and four complete focus cycles. Two deliberate negative controls prove the gate is sensitive: replacing the first expected perceptual hash with zeroes failed at 566 changed bits, or 22.109%, against the 8% limit; replacing its expected mean RGB with zeroes failed at channel deltas of 229.142, 231.456, and 239.484 against the 6.0 limit. Neither intentional negative control is a product pass.

The complete documentation and release-contract slice reported 58 passes and 1,379 subtests. The complete Python 3.12.12 suite reported 2,481 passes, 15 skipped, 3,797 subtests, and 35 warnings in 110.41 seconds. The fresh distribution probe passed wheel, source-distribution, and editable installs, including dependency and lazy-import checks, with a 57,192,282-byte wheel and 55,446,461-byte source distribution. Release alignment still finds five benchmark records and all 68 documented providers; all 68 model pages and 59 notebooks remain generator-current.

Fresh official Pipeline screenshots returned HTTP 200 for light desktop, tablet, and mobile plus dark desktop and tablet. The first dark-mobile request was rate-limited with HTTP 429 and is not a pass; a later isolated retry returned HTTP 200 and its screenshot was inspected. The local desktop light/dark and mobile light/dark screenshots were visually inspected. The new manifest protects VoiceHub against local raster regressions; it does not claim raw-pixel identity with upstream's intentionally different Hugging Face branding, community panel, prose, or VoiceHub's approved palette. Exact- commit remote CI, protected publisher configuration, tag and publication approval, five native- kernel hardware gates, and seven inaccessible asset/checkpoint/oracle paths remain open. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Representative-route native focus gate

The next bounded documentation-accessibility slice expands the native keyboard contract from the Pipeline sample to every representative page. The official reference remains Transformers commit b3a36037d3feb22e3f0174b3dd4248fcc0f0f722, and the current navigation file retains SHA-256 f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a. The official Pipeline endpoint returned HTTP 200 during the refresh.

A pre-edit Chromium probe established three real viewport prefixes: desktop starts with skip, logo, search, version, language, theme, and source; tablet omits the compacted logo; mobile starts with skip, drawer, search trigger, and version. The same probe found 239 focus stops on the desktop model index, so the previous fixed 200-step ceiling could not represent a complete cycle. The checker now derives a safe bound from each rendered DOM and executes a full cycle inside all 60 route, viewport, and palette cases. Every stop must be rendered, every inactive navigation branch must remain unreachable, fixed prefix controls must stay on canvas, branch, drawer, and content-tab controls must expose the two-pixel focus indicator, each cycle must reach BODY, and a forward Tab from that boundary must return to Skip to content.

The first full render resumed after the theme control because a plain blur did not reset Chromium's sequential-navigation cursor. A document-start focus reset corrected that harness boundary. The next run reached a zero-sized Installation content-tab radio whose visible label had no focus indicator. The shared tab initializer now maps focused radios to their rendered labels, and the checker validates the label as the visible focus proxy. A later Models API tablet cycle reached BODY but Chromium resumed at a nearby scrolled article link. The shared document boundary now routes an unmodified forward Tab from body focus to the skip link. These failed runs are regression evidence and are not counted as passes. An initial focused test invocation also named the wrong test class and collected no tests; its corrected pre-implementation run failed as intended.

After correction, the two focused source contracts and every applicable selected pre-commit hook passed. Markdown formatting had no matching selected file and is not counted. The strict eleven-language build and ten-route DOM validator passed. The Playwright 1.62.0 and Chromium 151.0.7922.34 gate passed all 60 geometry cases, all 60 screenshot comparisons, all 60 complete focus cycles, and four separate Enter/Space activation cases. The latest run visited 4,075 focus stops; this is an observed count rather than a fixed expectation because Material may conditionally expose Back to top.

The documentation and release-contract slice reported 59 passes and 1,379 subtests. The registry and public-optimization slice reported 231 passes, 1,203 subtests, and four recorded warnings. The complete Python 3.12.12 suite reported 2,482 passes, 15 skipped, 3,797 subtests, and 35 warnings in 110.60 seconds. The fresh distribution probe passed wheel, source-distribution, and editable installs with 68 models, 81 provenance manifests, 193 compliance files, no eager Torch import, and zero dependency violations. Its wheel was 57,192,282 bytes and its source distribution was 55,446,859 bytes.

Exact-commit remote CI for the uncommitted worktree, protected publisher configuration, tags, publication approval, five native-kernel hardware gates, and seven inaccessible asset, checkpoint, or oracle paths remain open. The untracked uv.lock remained unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Rendered semantic-accessibility gate

The next bounded documentation slice closes the missing automated rendered accessibility contract without claiming a complete WCAG or manual audit. The official structural reference remains Transformers commit b3a36037d3feb22e3f0174b3dd4248fcc0f0f722; the navigation fingerprint remains f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a. The checker pins axe-playwright-python 0.1.8 and reports its bundled Axe Core 4.12.1 engine beside the existing Playwright matrix.

The pre-edit 60-case inventory found five shared failure classes: invalid button roles on labels, insufficient code-token and footer contrast, non-unique code-action landmarks, links distinguished only by color, and closed search results that still exposed an unfocusable scroll region. Drawer, search, theme, and runtime navigation toggles now use native buttons; closed search output is both inert and hidden from the accessibility tree; code-action landmarks receive stable per-page names; paragraph links remain underlined; and the affected light and dark color tokens exceed the automated contrast threshold. The first corrected matrix then found a Quickstart table that was scrollable only by pointer at mobile width. All rendered table wrappers now have a named keyboard stop and a visible two-pixel focus indicator.

The focused source contract failed before those changes and passed afterward. Two early full-browser attempts timed out before a product assertion while waiting for the first local networkidle state, and the first completed Axe run failed on the Quickstart table; none is counted as a pass. A deliberate unnamed-button injection triggered Axe's button-name rule, demonstrating that the new gate fails observably. The final current-code run passed all 60 Axe cases, all 60 geometry cases, all 60 screenshot comparisons, all 60 complete focus cycles, and four separate Enter/Space activations. Its 4,213 observed focus stops are not a fixed expectation because Material conditionally exposes Back to top.

The documentation source suite reported 51 passes and 1,379 subtests. The registry and all public-optimization suites reported 231 passes, 1,203 subtests, and four warnings. The combined documentation, release, guidance, distribution-source, and packaging slice reported 78 passes and 1,466 subtests. The full Python 3.12.12 suite reported 2,483 passes, 15 skips, 3,797 subtests, and 35 warnings in 113.36 seconds. The fresh distribution probe passed wheel, source-distribution, and editable installs with 68 models, 81 provenance manifests, 193 compliance files, no eager Torch import, and zero dependency violations; its wheel was 57,192,297 bytes and its source distribution was 55,446,839 bytes. The first selected pre-commit run reformatted the Python source and returned nonzero; the final applicable selected hooks passed. Markdown formatting had no matching file in that pre-evidence run and is not counted there.

Exact-commit remote CI, protected publisher configuration, tags, publication approval, five native-kernel hardware gates, and seven inaccessible asset, checkpoint, or oracle paths remain open. The untracked uv.lock remained unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Expanded-state semantic-accessibility gate

The next bounded documentation-accessibility slice extends the automated Axe contract from settled page loads to settled shared-shell interaction states. The official reference remains Transformers commit b3a36037d3feb22e3f0174b3dd4248fcc0f0f722; the current _toctree.yml fingerprint remains f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a, and the official Pipeline endpoint returned HTTP 200 during the refresh.

The new 30-case inventory contains six search-open, six search-results, six search-empty, six version-open, four desktop/tablet branch-open, and two mobile drawer-open states across both VoiceHub palettes and every applicable viewport. Before editing, 12 cases failed. Dynamically generated search-result code regions were scrollable but not keyboard-focusable; the version disclosure mixed list and menu roles; nested navigation landmarks repeated their names; and the opened mobile drawer exposed an insufficiently contrasting nested title. Search-result code regions now receive stable names, keyboard stops, and visible focus through a mutation observer. The version selector uses native details, summary, list, and link semantics. Runtime navigation panels receive unique labels, and the mobile nested title uses the shared readable foreground token.

The checker also waits for the search dialog's opacity transition to finish before auditing it. This is necessary because an intermediate mobile result probe measured the search container at opacity 0.666036; its apparent result count contrast was 4.24:1 even after the final foreground token was correct. The settled state exceeds the automated threshold. The initial inventory, a malformed branch-state JavaScript predicate, a palette focus-restoration timeout, missing shell python, a stale pytest node ID, the 3.2:1, 4.2:1, and 4.24:1 contrast probes, and the YAPF/docformatter rewrite conflict are all recorded failures or harness errors, not passing evidence. Ordinary concatenated JavaScript literals made the two Python formatters converge, and the final selected hook run passed every applicable hook.

The exact current-code Playwright 1.62.0 and Chromium 151.0.7922.34 run passed 60 base Axe cases, 30 expanded-state Axe cases, 60 geometry checks, 60 screenshot comparisons, 60 complete focus cycles with 4,229 observed focus stops, and four independent Enter/Space activation cases. The source contract reported two focused passes. The documentation suite reported 52 passes and 1,379 subtests; registry and public-optimization suites reported 231 passes, 1,203 subtests, and four warnings; and the focused guidance, release, distribution-source, and pipeline slice reported 24 passes and 19 subtests. The complete Python 3.12.12 suite reported 2,484 passes, 15 skips, 3,797 subtests, and 35 warnings in 110.14 seconds.

Fresh wheel, source-distribution, and editable probes passed with 68 models, 81 provenance manifests, 193 compliance files, all required package data, no eager Torch import, and zero dependency violations. The artifacts measured 57,192,297 and 55,446,883 bytes. The refreshed inventories remain 34 TTS, 23 ASR, 11 VAD, 102 aliases, 68 current model pages, 59 current notebooks, six public optimization passes, 408 model/pass pairs, five benchmark records, eight top-level navigation roots, ten representative routes, and eight contribution steps.

This evidence covers automated detectable rules in the enumerated settled states; it is not a complete WCAG or manual accessibility audit. Exact-current- worktree remote CI, protected publisher configuration, tags, publication approval, five native-kernel hardware gates, and seven inaccessible asset, checkpoint, or oracle paths remain open. The untracked uv.lock remained unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Current Installation route parity gate

The next bounded documentation slice maps the current official Transformers Installation page to VoiceHub's speech-specific installation and offline workflow. The refreshed reference is Transformers commit b3a36037d3feb22e3f0174b3dd4248fcc0f0f722; its _toctree.yml SHA-256 is f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a, and the official Installation endpoint returned HTTP 200.

The official rendered page currently uses Installation, Virtual environment, Python, Source install, Editable install, conda, Set up, Cache directory, and Offline mode in that order, with no content-tab component. Its article measured 804 pixels at x = 318 on desktop, 658 pixels at x = 318 on tablet, and 342 pixels at x = 24 on mobile, with zero document overflow. VoiceHub's previous source instead used numbered sections and an OS tab set. The pre-edit focused source contract failed on that stale hierarchy and is not counted as a pass.

The rewritten VoiceHub page now matches the current heading and TOC hierarchy while retaining original speech-library instructions: uv-first environment management, published, source, and editable installs, an honest conda-owned environment boundary, the exact shared Hub cache precedence, and global or call-level offline operation. It does not invent a conda-forge VoiceHub package, report checkpoint-specific offline inference as verified, or imply that the development contract is already available from the older PyPI release. Six rendered route cases enforce the exact headings, TOC, absence of stale tabs, required cache/offline markers, shared geometry, both palettes, three viewports, Axe checks, complete native focus, and reviewed screenshot signatures.

The first focused rendered run exposed a real light-theme syntax-token failure: Pygments variable names rendered at 4.486:1 against the code background, below the 4.5:1 Axe threshold. The shared light-theme variable token is now #686a72, which measures 4.948:1 against #f4f5f8. That failed run is not passing evidence. The first selected hook run also let YAPF format the Python checker and returned nonzero; the formatted rerun passed every applicable hook.

The final strict eleven-language build and ten-route DOM validator passed. Playwright 1.62.0 with Chromium 151.0.7922.34 and Axe Core 4.12.1 passed 60 base accessibility, geometry, screenshot, and complete-focus cases, six Installation structure cases, 30 expanded-state accessibility cases, and four independent Enter/Space activation cases. The matrix traversed 4,250 visible focus stops. The documentation suite reported 53 passes and 1,380 subtests; the registry and public-optimization slice reported 231 passes, 1,203 subtests, and four warnings; and the focused guidance, release, distribution- source, and pipeline slice reported 24 passes and 19 subtests. The complete Python 3.12.12 suite reported 2,485 passes, 15 skips, 3,798 subtests, and 35 warnings in 110.86 seconds.

Package code and metadata did not change, so wheel, source-distribution, and editable probes were not rebuilt for this documentation-only slice. The immediately preceding successful artifacts remain 57,192,297 and 55,446,883 bytes and are not new evidence here. The refreshed inventories remain 34 TTS, 23 ASR, 11 VAD, 102 aliases, 68 current model pages, 59 current notebooks, six public optimization passes, 408 model/pass pairs, five benchmark records, eight top-level navigation roots, ten representative routes, and eight contribution steps.

Exact-current-worktree remote CI, protected publisher configuration, tags, publication approval, five native-kernel hardware gates, and seven inaccessible asset, checkpoint, or oracle paths remain open. The untracked uv.lock remained unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Current Quickstart route parity gate

The bounded release slice mapped VoiceHub's Quickstart route to the current official Transformers Quickstart at commit b3a36037d3feb22e3f0174b3dd4248fcc0f0f722. The upstream _toctree.yml retained SHA-256 f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a, and the rendered endpoint returned HTTP 200. Its current hierarchy is Quickstart, Set up, Agent skills, Pretrained models, Pipeline, Trainer, and Next steps; its articles measured 804, 658, and 342 pixels at the matched desktop, tablet, and mobile viewports with zero overflow.

VoiceHub's stale Discover models and Inference hierarchy had no current tab or callout contract. The pre-edit source regression failed on that mismatch. The corrected route keeps VoiceHub instructions original and speech-specific while mapping the current page structure: uv/pip setup, source-checkout agent skills, shared pretrained contracts, TTS/ASR/VAD pipelines, and an explicit Trainer support boundary. Source and rendered tests require the exact hierarchy, three tab sets with seven options, two tips, one model table, at least twelve code blocks, required content markers, and parsed Python examples.

Rendered negative evidence remained explicit. The iteration found and fixed an off-canvas keyboard target, tab state persisted between viewports, an unnamed keyboard-inaccessible code scroller, eight pixels of mobile overflow, an unnamed scrollable option strip, and transient low-contrast Back to top text. The final shared behavior gives overflowing code and option strips named keyboard stops, keeps mobile tabs inside the document, resets tab state for deterministic baselines, and removes the opacity transition that produced the low-contrast intermediate state. None of the failed runs is counted as a pass.

The final strict eleven-language build and ten-route DOM validator passed. Playwright 1.62.0 with Chromium 151.0.7922.34 and Axe Core 4.12.1 passed 60 base accessibility, geometry, screenshot, and complete-focus cases, six Quickstart structure cases, six post-activation Quickstart tab cases, 30 expanded-state accessibility cases, and four independent branch/drawer activations. The matrix traversed 4,297 visible focus stops. The documentation suite reported 53 passes and 1,380 subtests; the registry and public-optimization slice reported 231 passes, 1,203 subtests, and four warnings; and the focused guidance, release, distribution-source, and pipeline slice reported 24 passes and 19 subtests. The complete Python 3.12.12 suite reported 2,485 passes, 15 skips, 3,798 subtests, and 35 warnings in 107.12 seconds.

The first combined selected-file pre-commit runs returned docformatter exit 3. The failure was reproducible only after the preceding hooks because docformatter 1.7.8 misclassified the multiline JavaScript callback used to reset Quickstart tabs and proposed damaging whitespace inside the string. A serial-execution experiment did not fix the observable command and was discarded. Replacing only that callback with equivalent concatenated strings made docformatter's direct check, its hook, and the full selected-file sequence exit zero with identical before/after hashes. The failed runs are not passes. Package code and metadata did not change during the Quickstart slice. A subsequent exact-current-worktree distribution iteration rebuilt and validated the wheel, source distribution, and editable install; its fresh evidence is recorded in the next section.

The refreshed inventories remain 34 TTS, 23 ASR, 11 VAD, 102 aliases, 68 current model pages, 59 current notebooks, six public optimization passes, 408 model/pass pairs, five benchmark records, eight top-level navigation roots, ten representative routes, and eight contribution steps. Exact-current- worktree remote CI, protected publisher configuration, tags, publication approval, five native-kernel hardware gates, and seven inaccessible asset, checkpoint, or oracle paths remain open. The untracked uv.lock remained unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree distribution evidence

The focused distribution and release contracts reported 13 passes. The source alignment check found version 0.3.0, 34 TTS, 23 ASR, 11 VAD, all 68 documented providers, and five benchmark records. The full temporary scripts/check_distribution.py run then built the current worktree and probed wheel, source-distribution, and editable installations in three isolated environments without installing runtime dependencies.

All three installation modes reported version 0.3.0, 68 models, 193 compliance files, every required representative data file, zero runtime-dependency violations, and no eager PyTorch import. The build contained 81 pinned and licensed provenance manifests. The validated wheel was 57,192,297 bytes and the validated source distribution was 55,447,546 bytes, both below the 100 MB publication limit. The temporary environments and artifacts were removed after the probe; nothing was copied into the repository or published.

A second temporary build passed scripts/check_release.py --dist-dir and recorded exact-run fingerprints:

Artifact	Bytes	SHA-256
voicehub-0.3.0-py3-none-any.whl	57,192,297	83b32ea12255f7829d8e152069a32f85f5eb786d53a31bb01d458676426b2a74
voicehub-0.3.0.tar.gz	55,448,121	eb7284f989b2c3f8365959c1f790c3071c7e43f58853f1852e86cc5fb2c1fbe1

The two successful sdist builds differed by 575 bytes, so this report does not claim byte-for-byte reproducible archives; the fingerprints identify only the second exact build. Tagged-workflow artifacts and their authoritative hashes remain a separate release gate. Exact-current-worktree remote CI, protected publisher configuration, tags, publication approval, five native-kernel hardware gates, and seven inaccessible asset, checkpoint, or oracle paths also remain open. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree supported-Python evidence

The next bounded release iteration exercised the complete current worktree on the two supported interpreters whose full-suite records predated the latest documentation and packaging contracts. Python 3.10.19 and Python 3.11.15 each created a new temporary virtual environment. uv 0.11.21 then used its CPU PyTorch backend to install the current source directly with uv pip install; the install target was -e ".[test]". Project synchronization was not used, and the repository lock file was neither read nor modified.

Both environments first passed the 13 focused release and distribution contracts. The corrected Hub transport regression separately reported 19 passes and five subtests on each interpreter. The complete default-offline suite, with all opt-in asset variables left unset, then reported identical coverage:

Interpreter	Passed	Skipped	Subtests	Warnings	Duration
Python 3.10.19	2,485	15	3,798	35	108.21 seconds
Python 3.11.15	2,485	15	3,798	35	98.60 seconds

Those runs did not count any of the 15 skipped paths as passed. The next default-runtime iteration executes the three import gates separately. An initial command incorrectly forced VOICEHUB_OFFLINE=1, which prevented 12 mocked Hub transport tests from reaching their patched URL opener. Those two executions failed with 12 failures, 2,473 passes, 15 skips, 3,798 subtests, and 35 warnings in 213.83 seconds on Python 3.10 and 200.08 seconds on Python 3.11. They are recorded as failed command evidence and are not counted as supported- version passes.

The successful Python 3.10 and 3.11 results join the fresh Python 3.12.12 current-worktree result above, so the exact current source has now passed its complete CPU-safe suite on every supported interpreter on macOS. This does not substitute for exact-current-worktree Linux or Windows CI. At that point, tagged-workflow artifact hashes, publisher configuration, publication approval, and the explicit hardware and asset skip boundaries remained open; the next section refreshes their current status. The temporary environments were removed after verification, and the untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree default-runtime evidence

The next bounded release iteration refreshed the official parity reference before selecting a gate. Transformers main remained b3a36037d3feb22e3f0174b3dd4248fcc0f0f722, its documentation toctree retained SHA-256 f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a, and the official documentation endpoint remained available. No parity mapping changed in this release-evidence iteration.

Python 3.12.12 created a fresh temporary virtual environment. uv 0.11.21 used its CPU PyTorch backend to install the exact current source with .[test] and without project synchronization. With VOICEHUB_FULL_RUNTIME_TEST=1, the focused default-runtime file reported five passes and 138 registry subtests in 16.58 seconds. The same isolated environment then completed the entire suite with 2,488 passes, 12 skips, 3,936 subtests, and 35 warnings in 157.75 seconds. The three default-runtime tests that the ordinary suite skips all executed and passed.

The remaining 12 skips retain separate boundaries:

• Three Triton and two compiled CUDA-extension tests require unavailable CUDA hardware or a CUDA toolkit and remain unpassed.
• ESPNet, SenseVoice, SpeechBrain, TEN-VAD, and QuartzNet are opt-in in the default suite but retain their separately executed successful evidence.
• The WeNet checkpoint and tokenizer remain inaccessible and unpassed.

The ignored local dist/ directory was inspected without modification. Its 54,572,319-byte wheel has SHA-256 d37fee52843a2a8ac654894509199aaccc5493830b6d19b6f3ccaddf252352f4, and its 52,863,267-byte source distribution has SHA-256 d0224a0d39c58ca7a0398fad356a4e8c40577131f250c58b459ed2d3902c1b42. Those files do not match the exact current-worktree fingerprints recorded above, were preserved as user-owned local artifacts, and are not release- candidate evidence.

This closes the current-worktree default-runtime gate locally; execution of the same uncommitted source in remote Linux CI remains pending. Exact-current- worktree Linux and Windows matrix coverage, tagged-workflow artifact hashes, protected publisher configuration, tags, publication approval, five native- kernel hardware gates, and two inaccessible WeNet asset paths remain open. The temporary environment was removed after verification, and the untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree Pipeline interaction evidence

The next bounded documentation iteration refreshed the official task-guide pair before selecting the remaining local shell gap. Transformers main remained b3a36037d3feb22e3f0174b3dd4248fcc0f0f722, and _toctree.yml retained SHA-256 f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a. The mapped routes remain /docs/transformers/main/en/pipeline_tutorial and /voicehub/guides/inference/.

The focused source test first failed on the absence of a rendered Pipeline contract. The initial exact render then found that all six Pipeline code blocks had zero copy buttons despite the configured content.code.copy feature; that failure is not passing evidence. The shared page-action runtime now creates native, labelled code-copy buttons with Clipboard API and selection-fallback paths, explicit busy/success states, and retained keyboard focus. The first implementation copied syntax-highlighted visual line breaks rather than exact source text and failed its clipboard comparison. That failure is also excluded; the corrected action copies the code element's exact text content.

The final strict eleven-language build passed. Playwright 1.62.0 with Chromium 151.0.7922.34 and Axe Core 4.12.1 then passed 60 representative route/viewport/palette cases, including six exact Pipeline structure/content cases and six Pipeline keyboard-copy interaction cases. The copy cases use Enter and Space across desktop, tablet, mobile, light, and dark states; they verify exact clipboard text, visible success state, focus retention, zero overflow, and a second zero-violation Axe result. The full matrix recorded 60 base Axe checks, 60 reviewed screenshot signatures, 60 complete focus cycles with 4,572 stops, 30 expanded-state Axe checks, four branch/drawer activation cases, six Quickstart tab cases, and six Pipeline copy cases. The ten-route DOM validator also passed.

The complete documentation suite reported 53 passes and 1,380 subtests. A risk-proportional registry, universal-optimization, distribution-compliance, and release-readiness slice reported 97 passes and 452 subtests. Refreshed inventories remain 68 models (34 TTS, 23 ASR, 11 VAD), 102 aliases, zero invalid display names, 68 current provider pages, 59 current model notebooks, six public optimization passes, and 408 model/pass pairs.

The first combined final-file sequence exposed a YAPF line split that broke one source-level CI fragment assertion; that run reported one failure, 52 passes, and 1,380 subtests and is not passing evidence. Computing the aggregate keyboard-case count before constructing the result mapping restored stable formatting. The complete selected-file pre-commit sequence and all 53 documentation tests then passed together.

This closes the local Pipeline structure and code-copy interaction gate. The complete Python suite and physical package builds were not rerun for this documentation-runtime slice; their exact preceding current-worktree evidence remains separate. Exact-current-worktree remote CI, exact tagged artifact hashes, protected publisher configuration, tags, publication approval, five native-kernel hardware gates, and two inaccessible WeNet asset paths remain open. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree Auto Classes interaction evidence

The next bounded documentation iteration refreshed the official Auto Classes pair before selecting the remaining local shell gap. Transformers main remained b3a36037d3feb22e3f0174b3dd4248fcc0f0f722, _toctree.yml retained SHA-256 f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a, the official route returned HTTP 200, and the pinned 301-line Auto Classes source has SHA-256 557f5836c0722fef6a484c46805dfab0eb69a387b028a914b132350edf09f167. The mapped routes remain /docs/transformers/main/en/model_doc/auto and /voicehub/models/providers/.

The focused source contract intentionally failed before implementation with one failure and 68 passing subtests because no model-index-specific rendered contract existed. That run is not passing evidence. The first selected-file pre-commit run later failed when docformatter rewrote a multiline JavaScript predicate; it is also excluded. The predicate now uses formatter-safe concatenated strings, and the complete selected-file sequence passes.

The final strict eleven-language build passed. Playwright 1.62.0 with Chromium 151.0.7922.34 and Axe Core 4.12.1 passed all 60 representative route/viewport/palette cases, including six exact Auto Classes structure and content cases. The contract protects nine headings, eight right-TOC entries, four registry-derived tables with 3/34/23/11 rows, 68 unique uppercase-first provider links, three code blocks and copy actions, and the required configuration, processing, task-model, registry-discovery, and lazy-loading markers.

Six additional cases used Enter in the default palette and Space in slate at desktop, tablet, and mobile widths to activate page copy. They verified exact clipboard text, visible success and idle state, focus retention, zero overflow, and a second zero-violation Axe result. The complete matrix recorded 60 base Axe checks, 60 reviewed screenshot signatures, 60 complete focus cycles with 4,620 stops, 30 expanded-state Axe checks, four branch/drawer activation cases, six Quickstart tab cases, six Pipeline copy cases, and six Auto Classes page-copy cases, for 82 keyboard cases total. The ten-route DOM validator also passed.

The complete documentation suite reported 53 passes and 1,380 subtests. A risk-proportional registry, speech-task, universal-optimization, distribution-compliance, and release-readiness slice reported 97 passes and 452 subtests. All 68 generated provider pages and 59 generated model notebooks remain current, and the five-record release-alignment check passed.

This closes the local Auto Classes structure and page-copy interaction gate. The complete Python suite and physical package builds were not rerun for this documentation-only slice; their exact preceding current-worktree evidence remains separate. Exact-current-worktree remote CI, exact tagged artifact hashes, protected publisher configuration, tags, publication approval, five native-kernel hardware gates, and two inaccessible WeNet asset paths remain open. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree SpeechT5 interaction evidence

The next bounded documentation iteration refreshed the official SpeechT5 model-detail pair. Transformers main remained b3a36037d3feb22e3f0174b3dd4248fcc0f0f722, _toctree.yml retained SHA-256 f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a, the route returned HTTP 200, and the pinned 87-line source retained SHA-256 71bba8a2921cf637383fb8d6f2fd66df9cd95deb59118b9f49e1362485c27eb5. The mapped routes remain /docs/transformers/main/en/model_doc/speecht5 and /voicehub/models/providers/speecht5/.

The first focused command named a nonexistent unittest class, collected no tests, and exited 4; it is not passing evidence. The corrected focused test then intentionally failed because no SpeechT5-specific rendered contract existed. The first full render later failed because the new table inventory expected ten checkpoint rows while the source correctly renders nine. Neither failure is counted as a pass. Correcting the expectation required no generated page change.

The final strict eleven-language build passed. Playwright 1.62.0 with Chromium 151.0.7922.34 and Axe Core 4.12.1 passed all 60 representative route/viewport/palette cases, including six exact SpeechT5 structure and content cases. The contract protects 14 headings, 13 right-TOC entries, eight generated tables with 6/3/4/2/6/1/9/8 rows, seven code blocks and copy actions, two local facade-source links, and required auto-model, processor, normalized-output, checkpoint, real-evidence, and fail-closed optimization markers.

Six additional cases used Enter in the default palette and Space in slate at desktop, tablet, and mobile widths to activate page copy. They verified exact clipboard text, visible success and idle state, focus retention, zero overflow, and a second zero-violation Axe result. The complete matrix recorded 60 base Axe checks, 60 reviewed screenshot signatures, 60 complete focus cycles with 4,637 stops, 30 expanded-state Axe checks, four branch/drawer activation cases, six Quickstart tab cases, six Pipeline copy cases, six Auto Classes page-copy cases, and six SpeechT5 page-copy cases, for 88 keyboard cases total. The ten-route DOM validator also passed.

The complete documentation suite reported 53 passes and 1,380 subtests. A risk-proportional registry, speech-task, universal-optimization, distribution-compliance, and release-readiness slice reported 97 passes and 452 subtests. All 68 generated provider pages and 59 generated model notebooks remain current, and the five-record release-alignment check passed.

This closes the local SpeechT5 model-detail structure and page-copy interaction gate. The complete Python suite and physical package builds were not rerun for this documentation-only slice; their exact preceding current-worktree evidence remains separate. Exact-current-worktree remote CI, exact tagged artifact hashes, protected publisher configuration, tags, publication approval, five native-kernel hardware gates, and two inaccessible WeNet asset paths remain open. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree Trainer interaction evidence

The next bounded documentation iteration refreshed the official Trainer overview pair. Transformers main remained b3a36037d3feb22e3f0174b3dd4248fcc0f0f722, _toctree.yml retained SHA-256 f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a and the Trainer overview then Fine-tuning order, the route returned HTTP 200, and the pinned 30-line source retained SHA-256 e7c5368c1223c2b195321468fcd1ac64c5cda52a1e44d4adf854a6473b6c9ee0. The mapped routes remain /docs/transformers/main/en/trainer and /voicehub/guides/trainer/.

The focused source contract intentionally failed before implementation because no Trainer-specific rendered contract existed. That run is not passing evidence. The new contract protects the exact two-heading and one-entry TOC inventory, zero tables and code blocks, one page-copy action, four exact next-step destinations, the edit and Fine-tuning footer targets, and the Trainer, TrainingArguments, model-owned objective, exact-resume, and fail-closed speech-training markers.

The final strict eleven-language build passed. Playwright 1.62.0 with Chromium 151.0.7922.34 and Axe Core 4.12.1 passed all 60 representative route/viewport/palette cases, including six exact Trainer structure and content cases. Six additional cases used Enter in the default palette and Space in slate at desktop, tablet, and mobile widths to activate page copy. They verified exact clipboard text, visible success and idle state, focus retention, zero overflow, and a second zero-violation Axe result. The complete matrix recorded 60 base Axe checks, 60 reviewed screenshot signatures, 60 complete focus cycles with 4,771 stops, 30 expanded-state Axe checks, four branch/drawer activation cases, six Quickstart tab cases, six Pipeline copy cases, six Auto Classes page-copy cases, six SpeechT5 page-copy cases, and six Trainer page-copy cases, for 94 keyboard cases total. The ten-route DOM validator also passed.

The complete documentation suite reported 53 passes and 1,380 subtests. A risk-proportional registry, speech-task, universal-optimization, distribution-compliance, and release-readiness slice reported 97 passes and 452 subtests. All 68 generated provider pages and 59 generated model notebooks remain current, and the five-record release-alignment check passed.

This closes the local Trainer structure and page-copy interaction gate. The complete Python suite and physical package builds were not rerun for this documentation-only slice; their exact preceding current-worktree evidence remains separate. Exact-current-worktree remote CI, exact tagged artifact hashes, protected publisher configuration, tags, publication approval, five native-kernel hardware gates, and two inaccessible WeNet asset paths remain open. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree Optimization interaction evidence

The next bounded documentation iteration refreshed the official Optimization overview pair. Transformers main remained b3a36037d3feb22e3f0174b3dd4248fcc0f0f722, _toctree.yml retained SHA-256 f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a and the optimization_overview title Overview, the route returned HTTP 200, and the pinned 178-line source retained SHA-256 19622667a7299f258f5c9a72940c9f26492619636f35d9bd592701c02745b620. The mapped routes remain /docs/transformers/main/en/optimization_overview and /voicehub/guides/optimization-overview/.

The focused source contract intentionally failed before implementation with one failure and six passing subtests because no Optimization-specific rendered contract existed. That run is not passing evidence. The new contract protects the exact eight-heading and seven-entry TOC inventory, one six-row technique table, one code block and copy action, one page-copy action, all six registered optimization pass names, five exact workflow destinations, the edit and previous/next footer targets, and the application, validation, manifest, restoration, unsupported-quantization, parallelism, and continuous-batching boundaries.

The final strict eleven-language build passed. Playwright 1.62.0 with Chromium 151.0.7922.34 and Axe Core 4.12.1 passed all 60 representative route/viewport/palette cases, including six exact Optimization structure and content cases. Six additional cases used Enter in the default palette and Space in slate at desktop, tablet, and mobile widths to activate page copy. They verified exact clipboard text, visible success and idle state, focus retention, zero overflow, and a second zero-violation Axe result. The complete matrix recorded 60 base Axe checks, 60 reviewed screenshot signatures, 60 complete focus cycles with 4,603 stops, 30 expanded-state Axe checks, four branch/drawer activation cases, six Quickstart tab cases, six Pipeline copy cases, six Auto Classes page-copy cases, six SpeechT5 page-copy cases, six Trainer page-copy cases, and six Optimization page-copy cases, for 100 keyboard cases total. The ten-route DOM validator also passed.

The complete documentation suite reported 53 passes and 1,380 subtests. A risk-proportional registry, speech-task, universal-optimization, distribution-compliance, and release-readiness slice reported 97 passes and 452 subtests. All 68 generated provider pages and 59 generated model notebooks remain current, and the five-record release-alignment check passed. Refreshed inventories remain 68 models (34 TTS, 23 ASR, and 11 VAD), 102 aliases, zero invalid display names, six public optimization passes, 408 model/pass pairs, eight top-level navigation roots, ten representative routes, and eight contribution steps.

This closes the local Optimization overview structure and page-copy interaction gate. The complete Python suite and physical package builds were not rerun for this documentation-only slice; their exact preceding current-worktree evidence remains separate. Exact-current-worktree remote CI, exact tagged artifact hashes, protected publisher configuration, tags, publication approval, five native-kernel hardware gates, and two inaccessible WeNet asset paths remain open. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree Contribution interaction evidence

The next bounded documentation iteration refreshed the official modular- contribution pair. Transformers main remained b3a36037d3feb22e3f0174b3dd4248fcc0f0f722, _toctree.yml retained SHA-256 f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a and the modular then legacy contribution order, the route returned HTTP 200, and the pinned 500-line source retained SHA-256 a52af1ee1754e6e658535f071a727fac31a057a3a848c28c18873a5ff371dd96. The mapped routes remain /docs/transformers/main/en/modular_transformers and /voicehub/project/adding-a-model/.

The focused source contract intentionally failed before implementation with one failure and 18 passing subtests because no Contribution-specific rendered contract existed. That run is not passing evidence. Two read-only pre-edit HTML inventory probes also failed because Beautiful Soup is not installed and the first dependency-free parser selector assumed the content root was a div; neither is verification evidence. The corrected built-in parser established the exact local inventory without changing the site.

The new contract protects the exact 10-heading and nine-entry TOC inventory, the eight ordered process labels, three tables with 8/3/7 rows, 13 code blocks and copy actions, one page-copy action, two final contribution-guide targets, the edit and adjacent footer destinations, and the scaffold, registry, task bases, training, optimization, generation, distribution, and unverified- evidence boundaries.

The final strict eleven-language build passed. Playwright 1.62.0 with Chromium 151.0.7922.34 and Axe Core 4.12.1 passed all 60 representative route/viewport/palette cases, including six exact Contribution structure and content cases. Six additional cases used Enter in the default palette and Space in slate at desktop, tablet, and mobile widths to activate page copy. They verified exact clipboard text, visible success and idle state, focus retention, zero overflow, and a second zero-violation Axe result. The complete matrix recorded 60 base Axe checks, 60 reviewed screenshot signatures, 60 complete focus cycles with 4,572 stops, 30 expanded-state Axe checks, four branch/drawer activation cases, six Quickstart tab cases, six Pipeline copy cases, six Auto Classes page-copy cases, six SpeechT5 page-copy cases, six Trainer page-copy cases, six Optimization page-copy cases, and six Contribution page-copy cases, for 106 keyboard cases total. The ten-route DOM validator also passed.

The complete documentation suite reported 53 passes and 1,380 subtests. The model-scaffold contract reported 20 passes and 35 subtests. A risk-proportional registry, speech-task, universal-optimization, distribution-compliance, and release-readiness slice reported 97 passes and 452 subtests. All 68 generated provider pages and 59 generated model notebooks remain current, and the five-record release-alignment check passed. Refreshed inventories remain 68 models (34 TTS, 23 ASR, and 11 VAD), 102 aliases, zero invalid display names, six public optimization passes, 408 model/pass pairs, eight top-level navigation roots, ten representative routes, and eight contribution steps.

This closes the local Contribution structure and page-copy interaction gate. The complete Python suite and physical package builds were not rerun for this documentation-only slice; their exact preceding current-worktree evidence remains separate. Exact-current-worktree remote CI, exact tagged artifact hashes, protected publisher configuration, tags, publication approval, five native-kernel hardware gates, and two inaccessible WeNet asset paths remain open. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree Models API interaction evidence

The next bounded documentation iteration refreshed the official Models API pair. Transformers main remained b3a36037d3feb22e3f0174b3dd4248fcc0f0f722, _toctree.yml retained SHA-256 f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a and the Models entry between Logging and Text Generation, the route returned HTTP 200, and the pinned 44-line source retained SHA-256 a4899c758b5d621075b2e2f39f0aa79671010c88b08d1508f7af5731375c2871. The mapped routes remain /docs/transformers/main/en/main_classes/model and /voicehub/reference/models/.

The focused source contract intentionally failed before implementation with one failure and 26 passing subtests because no Models-API-specific rendered contract existed. That run is not passing evidence. The first full render then failed because the initial lifecycle-link selector included code-line number anchors. That run is also excluded. The corrected selector ignores links inside pre and protects only the declared user-facing content links.

The new contract protects the exact five-heading and four-entry TOC inventory, four tables with 6/3/3/5 rows, two code blocks and copy actions, one page-copy action, four exact facade-source targets, three exact internal lifecycle targets, the edit and Model audit/Full API footer destinations, and the configuration, factory, lazy-loading, task-base, normalized-output, training- output, portable-state, and unsupported-sharing boundaries.

The final strict eleven-language build passed. Playwright 1.62.0 with Chromium 151.0.7922.34 and Axe Core 4.12.1 passed all 60 representative route/viewport/palette cases, including six exact Models API structure and content cases. Six additional cases used Enter in the default palette and Space in slate at desktop, tablet, and mobile widths to activate page copy. They verified exact clipboard text, visible success and idle state, focus retention, zero overflow, and a second zero-violation Axe result. The complete matrix recorded 60 base Axe checks, 60 reviewed screenshot signatures, 60 complete focus cycles with 4,571 stops, 30 expanded-state Axe checks, four branch/drawer activation cases, six Quickstart tab cases, six Pipeline copy cases, six Auto Classes page-copy cases, six SpeechT5 page-copy cases, six Trainer page-copy cases, six Optimization page-copy cases, six Contribution page-copy cases, and six Models API page-copy cases, for 112 keyboard cases total. The ten-route DOM validator also passed.

The complete documentation suite reported 53 passes and 1,380 subtests. The speech-core, AutoConfig, and registry slice reported 54 passes and 216 subtests. A risk-proportional registry, speech-task, universal-optimization, distribution-compliance, and release-readiness slice reported 97 passes and 452 subtests. All 68 generated provider pages and 59 generated model notebooks remain current, and the five-record release-alignment check passed. Refreshed inventories remain 68 models (34 TTS, 23 ASR, and 11 VAD), 102 aliases, zero invalid display names, six public optimization passes, 408 model/pass pairs, eight top-level navigation roots, ten representative routes, and eight contribution steps.

This closes the local Models API structure and page-copy interaction gate. The complete Python suite and physical package builds were not rerun for this documentation-only slice; their exact preceding current-worktree evidence remains separate. Exact-current-worktree remote CI, exact tagged artifact hashes, protected publisher configuration, tags, publication approval, five native-kernel hardware gates, and two inaccessible WeNet asset paths remain open. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree Home interaction evidence

The next bounded documentation iteration refreshed the official Home pair. Transformers main remained b3a36037d3feb22e3f0174b3dd4248fcc0f0f722; the 65-line index.md source has SHA-256 e24e0eddadccef8b15e3f25c55acb0df0f65de21a4b5bf709d119d22d0eb555b, and _toctree.yml retained SHA-256 f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a with Transformers, Installation, and Quickstart at the start of Get started. The official route returned HTTP 200. The mapped routes remain /docs/transformers/main/en/index and /voicehub/.

The focused source contract intentionally failed before implementation because the local Home exposed only What is VoiceHub? instead of the current Features, Design, and Learn hierarchy. An intermediate raw-source assertion also failed on an existing line-wrapped bold registry label. Neither run is passing evidence. Normalizing the Markdown source corrected the assertion without weakening rendered checks. The first rendered run then passed the ten-route DOM validator but failed the stale desktop-light Home screenshot baseline by 289 bits (11.289%); that run is also excluded.

The replacement candidate differed materially above threshold only for the six Home viewport/palette signatures. Only those six reviewed entries were accepted, and no non-Home baseline was replaced. Rendered review also moved the long wrapping H1 to the compact product name VoiceHub, while retaining the original one-lifecycle message in the VoiceHub-specific tagline.

The final contract protects the VoiceHub, Features, Design, and Learn heading order; three TOC entries; Pipeline, Trainer, and speech-generation links; one tip; two design principles; 13 resource cards; four exact badge targets; six images with two decorative and four named status images; zero tables and code blocks; one page-copy action; the edit target; no previous destination; Installation as next; exact live registry counts; and the lazy-checkpoint, training-extra, and third-party-license boundaries.

The strict eleven-language build passed. Playwright 1.62.0 with Chromium 151.0.7922.34 and Axe Core 4.12.1 passed all 60 representative route/viewport/palette cases, including six exact Home structure/content cases and six Home keyboard-copy interaction cases. The copy cases use Enter and Space across desktop, tablet, mobile, light, and dark states; they verify exact clipboard text, visible success and idle state, focus retention, zero overflow, and a second zero-violation Axe result. The complete matrix recorded 60 base Axe checks, 60 reviewed screenshot signatures, 60 complete focus cycles with 4,713 stops, 30 expanded-state Axe checks, four branch/drawer activation cases, six Home page-copy cases, six Quickstart tab cases, six Pipeline copy cases, and six page-copy cases each for Auto Classes, SpeechT5, Trainer, Optimization, Contribution, and Models API, for 118 keyboard cases total. The ten-route DOM validator also passed.

A successful fresh official render confirmed the exact Transformers, Features, Design, and Learn hierarchy, an 804-pixel desktop article, and zero overflow. Subsequent six-state remote retries returned an unhydrated shell and were interrupted; those retries are not counted as passed. The preceding official six-state Home shell evidence remains tied to the same upstream commit, while the current source, HTTP, and local six-state evidence support this bounded slice.

The complete documentation suite reported 54 passes and 1,408 subtests. A risk-proportional registry, speech-task, universal-optimization, packaging- metadata, distribution-compliance, and release-readiness slice reported 97 passes and 522 subtests. All 68 generated provider pages and 59 generated model notebooks remain current, and the five-record release-alignment check passed. Refreshed inventories remain 68 models (34 TTS, 23 ASR, and 11 VAD), 102 aliases, zero invalid display names, six public optimization passes, 408 model/pass pairs, eight top-level navigation roots, ten representative routes, and eight contribution steps.

This closes the local Home structure and page-copy interaction gate. The complete Python suite and physical package builds were not rerun for this documentation-only slice; their exact preceding current-worktree evidence remains separate. Exact-current-worktree remote CI, exact tagged artifact hashes, protected publisher configuration, tags, publication approval, five native-kernel hardware gates, and two inaccessible WeNet asset paths remain open. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree Installation interaction evidence

The next bounded documentation iteration refreshed the official Installation pair at Transformers commit b3a36037d3feb22e3f0174b3dd4248fcc0f0f722. Its 164-line docs/source/en/installation.md source has SHA-256 d050e3e0e1c89d543c71c25367a455a9b49ea89c92f8c2376bb58294a4a4cf3b, and the 1,576-line _toctree.yml retained SHA-256 f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a with Transformers, Installation, and Quickstart at the start of Get started. The official route returned HTTP 200. Its rendered hierarchy remains Installation, Virtual environment, Python, Source install, Editable install, conda, Set up, Cache directory, and Offline mode, with code-copy controls, an edit action, and Transformers/Quickstart footer actions. The mapped local route remains /voicehub/getting-started/installation/.

The observable contract was fixed before implementation: six viewport/palette states must enforce the exact nine headings and eight-entry TOC, zero stale tab sets, 15 code blocks and code-copy buttons, one page-copy button, four exact external package-manager destinations, three exact internal workflow destinations, edit and previous/next targets, and the installation, cache, offline, and evidence-safety markers. Six Enter/Space cases must copy the first command, while six more must copy the readable page. Both action families must verify exact clipboard content, visible success and idle states, focus retention, zero overflow, and a post-action zero-violation Axe result.

The focused source contract intentionally failed before implementation because the checker had no Installation destination constants, copy helpers, or result counters. The first rendered run also failed because an assertion searched for the narrower phrase without importing PyTorch rather than the existing without downloading a checkpoint or importing PyTorch sentence. The exact phrase correction passed; neither failed run is counted. The Installation source itself required no change.

The final strict eleven-language build and ten-route DOM validator passed. Playwright 1.62.0 with Chromium 151.0.7922.34 and Axe Core 4.12.1 passed all 60 representative route/viewport/palette cases, including six exact Installation structure/destination cases, six Installation first-code copy cases, and six Installation page-copy cases. The matrix recorded 60 base Axe checks, 60 reviewed screenshot signatures, 60 complete focus cycles with 4,613 observed stops, 30 expanded-state Axe checks, four branch/drawer activation cases, and 130 keyboard cases. Installation used Enter in the default palette and Space in slate at desktop, tablet, and mobile widths; both actions returned from the success state to idle without losing focus or introducing overflow.

The complete documentation suite reported 54 passes and 1,408 subtests. A risk-proportional registry, speech-task, universal-optimization, packaging- metadata, distribution-compliance, and release-readiness slice reported 97 passes and 522 subtests. Selected pre-commit hooks passed in one invocation. All 68 generated provider pages and 59 generated model notebooks remain current. The five-record release-alignment check passed, and refreshed inventories remain 68 models (34 TTS, 23 ASR, and 11 VAD), 102 aliases, zero invalid display names, six public optimization passes, 408 model/pass pairs, eight top-level navigation roots, ten representative routes, and eight contribution steps.

This closes the current local Installation structure, destination, code-copy, and page-copy gate. The complete Python suite and physical package builds were not rerun for this documentation-checker-only slice; their exact preceding current-worktree evidence remains separate. Exact-current-worktree remote CI, exact tagged artifact hashes, protected publisher configuration, tags, publication approval, five native-kernel hardware gates, and two inaccessible WeNet asset paths remain open. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree Quickstart interaction evidence

The next bounded documentation iteration refreshed the official Quickstart pair at Transformers commit b3a36037d3feb22e3f0174b3dd4248fcc0f0f722. Its 312-line docs/source/en/quicktour.md source has SHA-256 ecfb99781204bcaea1ca63bcb4ad9ef70c99812e5f965a49b29de23cedd25bd7, and the 1,576-line _toctree.yml retained SHA-256 f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a with Transformers, Installation, and Quickstart at the start of Get started. The official route returned HTTP 200. Its rendered hierarchy remains Quickstart, Set up, Agent skills, Pretrained models, Pipeline, Trainer, and Next steps, with content tabs, code-copy controls, page actions, and previous/next navigation. The mapped local route remains /voicehub/getting-started/quickstart/. All three declared GitHub skill destinations also returned HTTP 200 in the refreshed link probe.

The observable contract was fixed before implementation: six viewport/palette states must enforce the exact seven headings and six-entry TOC, three tab sets with seven options, two tips, one three-row table, 12 code blocks and code-copy buttons, one page-copy button, three external skill destinations, nine internal workflow-link occurrences, and exact edit and previous/next targets. The six existing tab cases must still activate the final option in every set. Six Enter/Space cases must then copy the readable active-tab state with exact clipboard content, visible success and idle states, focus retention, zero overflow, and a post-action zero-violation Axe result.

The focused source contract intentionally failed before implementation because the checker had no Quickstart destination constants, page-copy helper, or result counter. The first rendered run also failed because Material's seven generated tab-control hashes entered the initial content-link inventory. The corrected inventory excludes those component controls while their dedicated tab-activation contract remains intact. Neither failed run is counted. The Quickstart source and screenshot baselines required no change.

The final strict eleven-language build and ten-route DOM validator passed. Playwright 1.62.0 with Chromium 151.0.7922.34 and Axe Core 4.12.1 passed all 60 representative route/viewport/palette cases, including six exact Quickstart structure/destination cases, six Quickstart tab cases, and six Quickstart page-copy cases. The matrix recorded 60 base Axe checks, 60 reviewed screenshot signatures, 60 complete focus cycles with 4,644 observed stops, 30 expanded- state Axe checks, four branch/drawer activation cases, and 136 keyboard cases. Quickstart page copy used Enter in the default palette and Space in slate at desktop, tablet, and mobile widths and returned from success to idle without losing focus or introducing overflow.

The complete documentation suite reported 54 passes and 1,408 subtests. A risk-proportional registry, speech-task, universal-optimization, packaging- metadata, distribution-compliance, and release-readiness slice reported 97 passes and 522 subtests. Selected pre-commit hooks passed in one invocation. All 68 generated provider pages and 59 generated model notebooks remain current. The five-record release-alignment check passed, and refreshed inventories remain 68 models (34 TTS, 23 ASR, and 11 VAD), 102 aliases, zero invalid display names, six public optimization passes, 408 model/pass pairs, eight top-level navigation roots, ten representative routes, and eight contribution steps.

This closes the current local Quickstart structure, destination, tab, and page-copy gate. The complete Python suite and physical package builds were not rerun for this documentation-checker-only slice; their exact preceding current- worktree evidence remains separate. Exact-current-worktree remote CI, exact tagged artifact hashes, protected publisher configuration, tags, publication approval, five native-kernel hardware gates, and two inaccessible WeNet asset paths remain open. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree representative TOC activation evidence

The next bounded documentation iteration refreshed the shared table-of-contents reference at Transformers commit b3a36037d3feb22e3f0174b3dd4248fcc0f0f722. Its 65-line docs/source/en/index.md source has SHA-256 e24e0eddadccef8b15e3f25c55acb0df0f65de21a4b5bf709d119d22d0eb555b, and the 1,576-line _toctree.yml retained SHA-256 f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a. The official Home route returned HTTP 200. No new remote rendered interaction was executed, so prior official shell evidence remains separate.

The observable contract was fixed before implementation. At the 1,440-pixel desktop viewport, all ten representative routes must activate their final TOC target by pointer and by unmodified Enter in both palettes. Each of the 40 cases must require the current hash to identify the CSS :target, exactly one matching active link after smooth scrolling settles, visible heading alignment beneath the header, zero overflow, and a post-action zero-violation Axe result. The 20 Enter cases must retain focus on the activated link.

The focused source contract intentionally failed before implementation because the checker had no representative-route activation matrix, helper, or result counters. The first full rendered run then failed on the intentional Home observer alignment at 63.703 pixels beneath the 65-pixel header. A focused rendered probe also failed by inspecting Installation before smooth scrolling and Material's observer settled. The corrected helper applies the existing two-pixel observer tolerance, waits for the exact hash and active link, and then requires another 600 milliseconds of stable state. Those failed runs are not counted as passing evidence.

The final strict eleven-language build and ten-route DOM validator passed. Playwright 1.62.0 with Chromium 151.0.7922.34 and Axe Core 4.12.1 passed all 60 base route/viewport/palette cases and 60 reviewed screenshot signatures. It completed 60 focus cycles with 4,607 stops, 30 expanded-state Axe cases, four branch/drawer activation cases, 20 pointer TOC cases, and 20 Enter TOC cases. Every TOC interaction received a second Axe pass, and the total keyboard inventory increased from 136 to 156 cases.

The complete documentation suite reported 55 passes and 1,408 subtests. A risk-proportional registry, speech-task, universal-optimization, packaging- metadata, distribution-compliance, and release-readiness slice reported 97 passes and 522 subtests. Selected pre-commit hooks passed. All 68 generated provider pages and 59 generated model notebooks remain current. The five-record release-alignment check passed, and refreshed inventories remain 68 models (34 TTS, 23 ASR, and 11 VAD), 102 aliases, zero invalid display names, six public optimization passes, 408 model/pass pairs, eight top-level navigation roots, ten representative routes, and eight contribution steps.

This closes representative-route TOC activation in the local rendered matrix. Non-representative routes and future templates remain outside that matrix. The complete Python suite and physical package builds were not rerun for this documentation-checker-only slice; their exact preceding current-worktree evidence remains separate. Exact-current-worktree remote CI, exact tagged artifact hashes, protected publisher configuration, tags, publication approval, five native-kernel hardware gates, and two inaccessible WeNet asset paths remain open. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree representative search activation evidence

The next bounded documentation iteration refreshed the shared header-search reference at Transformers commit b3a36037d3feb22e3f0174b3dd4248fcc0f0f722. Its 65-line docs/source/en/index.md source has SHA-256 e24e0eddadccef8b15e3f25c55acb0df0f65de21a4b5bf709d119d22d0eb555b, and the 1,576-line _toctree.yml retained SHA-256 f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a. The official Home route returned HTTP 200. No new remote rendered search interaction was executed, so prior official shell evidence remains separate.

The observable contract was fixed before implementation. All ten representative routes must activate and close search at desktop, tablet, and mobile widths in both palettes. Forty desktop/tablet cases use Ctrl+K, and 20 mobile cases use the visible pointer trigger. Each opened state must expose the checked toggle, expanded trigger, open body state, focused and visible input, active input/result tab stops, non-hidden and non-inert output, zero overflow, and a zero-violation Axe result. Escape must close the dialog, restore the hidden/inert/tab-order state, preserve zero overflow, and return focus to the visible inline input on desktop/tablet or the search trigger on mobile.

The focused source contract intentionally failed before implementation because the checker lacked the route/viewport/palette search matrix, helper, counters, and breakpoint-aware close-focus behavior. The implementation corrected the hidden desktop trigger as Escape's universal focus destination by selecting the visible inline input outside the mobile breakpoint. A six-case Home breakpoint probe passed. The first complete documentation-suite run after the rendered matrix then failed one legacy static assertion that still required trigger.focus(); the updated assertion now protects the breakpoint-aware helper. Neither failed run is passing evidence.

The final strict eleven-language build and ten-route DOM validator passed. Playwright 1.62.0 with Chromium 151.0.7922.34 and Axe Core 4.12.1 passed 60 base route/viewport/palette cases, 60 reviewed screenshot signatures, 60 complete focus cycles with 4,654 stops, 30 expanded-state Axe cases, four branch/drawer activation cases, 60 search activation/closure cases, and 40 TOC activation cases. Every opened search state received another Axe pass. The complete keyboard inventory increased from 156 to 196 cases.

The complete documentation suite reported 56 passes and 1,408 subtests. A risk-proportional registry, speech-task, universal-optimization, packaging- metadata, distribution-compliance, and release-readiness slice reported 97 passes and 522 subtests. All 68 generated provider pages and 59 generated model notebooks remain current. The five-record release-alignment check passed, and refreshed inventories remain 68 models (34 TTS, 23 ASR, and 11 VAD), 102 aliases, zero invalid display names, six public optimization passes, 408 model/pass pairs, eight top-level navigation roots, ten representative routes, and eight contribution steps.

This closes search activation and Escape restoration in the representative local route/viewport/palette matrix. Version, language, theme, and source route-specific activation remain outside this slice. The complete Python suite and physical package builds were not rerun; their exact preceding current-worktree evidence remains separate. Exact-current-worktree remote CI, exact tagged artifact hashes, protected publisher configuration, tags, publication approval, five native-kernel hardware gates, and two inaccessible WeNet asset paths remain open. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree representative version-control activation evidence

The next bounded documentation iteration refreshed the shared header-version reference at Transformers commit af0993dda925a8cac0a590f6e43a239933cc6d5b. Its 65-line docs/source/en/index.md source retains SHA-256 e24e0eddadccef8b15e3f25c55acb0df0f65de21a4b5bf709d119d22d0eb555b, and the 1,576-line _toctree.yml retains SHA-256 f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a. The official Home route returned HTTP 200. No new remote rendered version-menu interaction was executed, so prior official shell evidence remains separate.

The observable contract was fixed before implementation. All ten representative routes must activate and close the version control at desktop, tablet, and mobile widths in both palettes. The 30 default-palette cases use unmodified Enter and the 30 slate cases use the pointer. Each opened state must expose the exact three labels and destinations, one current item, the open and expanded state, summary focus, a visible menu inside the viewport, zero overflow, and a zero-violation Axe result. Escape must close and hide the menu, restore the collapsed ARIA state, preserve zero overflow, and return focus to the version summary.

The focused source contract intentionally failed before implementation because the checker had no version activation matrix, helper, or counters. The first rendered probe then exposed a real geometry failure: the desktop and tablet menu was right-aligned to a 64-pixel left-rail control and extended outside the viewport. The minimal responsive CSS correction left-aligns the LTR menu and right-aligns the RTL menu inside that rail while preserving mobile placement. A fresh six-case Home breakpoint/palette probe passed. Neither failed run is counted as passing evidence.

The strict eleven-language build and ten-route DOM validator passed. Playwright 1.62.0 with Chromium 151.0.7922.34 and Axe Core 4.12.1 passed 60 base route/viewport/palette cases, 60 reviewed screenshot signatures, 60 complete focus cycles with 4,704 stops, 30 expanded-state Axe cases, four branch/drawer activation cases, 60 version activation/closure cases, 60 search activation/closure cases, and 40 TOC activation cases. Every opened version state received another Axe pass. The complete keyboard inventory increased from 196 to 226 cases.

The complete documentation suite reported 57 passes and 1,408 subtests. A risk-proportional registry, speech-task, universal-optimization, packaging- metadata, distribution-compliance, and release-readiness slice reported 97 passes and 522 subtests. All 68 generated provider pages and 59 generated model notebooks remain current. The five-record release-alignment check passed, and refreshed inventories remain 68 models (34 TTS, 23 ASR, and 11 VAD), 102 aliases, zero invalid display names, six public optimization passes, 408 model/pass pairs, eight top-level navigation roots, ten representative routes, and eight contribution steps.

This closes version activation and Escape restoration across the representative local route/viewport/palette matrix. Language, theme, and source route-specific activation remain outside this slice. The complete Python suite and physical package builds were not rerun; their exact preceding current-worktree evidence remains separate. Exact-current-worktree remote CI, exact tagged artifact hashes, protected publisher configuration, tags, publication approval, five native-kernel hardware gates, and two inaccessible WeNet asset paths remain open. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree representative language-control activation evidence

The next bounded documentation iteration refreshed the shared header-language reference at Transformers commit af0993dda925a8cac0a590f6e43a239933cc6d5b. Its 65-line docs/source/en/index.md source retains SHA-256 e24e0eddadccef8b15e3f25c55acb0df0f65de21a4b5bf709d119d22d0eb555b, and the 1,576-line _toctree.yml retains SHA-256 f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a. The official Home route returned HTTP 200. No new remote rendered language-switch interaction was executed, so prior official shell evidence remains separate.

The observable contract was fixed before implementation. The language control is intentionally hidden at the shared mobile breakpoint, so all ten representative routes must switch locale at desktop and tablet widths in both palettes. Each of the 40 cases must expose one visible, focusable select with the exact ordered EN, TR, ES, FR, DE, PT, ZH, JA, KO, RU, and AR inventory and production-base destination for the current route. The 20 default-palette cases use unmodified ArrowDown to select Turkish; the 20 slate cases use pointer activation and semantic native-option selection to select Arabic. The target must retain the palette, route, selected locale, LTR or RTL direction, zero overflow, and a zero-violation Axe result.

The focused source contract intentionally failed before implementation because the checker had no language activation matrix, helper, or counters. Rendered probes then exposed that the local server did not mount /voicehub/, headless Chromium did not surface native picker key events, and a locale navigation reset slate to default. The test server now serves the production base, the native select explicitly commits unmodified ArrowUp and ArrowDown changes, and a one-navigation sessionStorage transfer restores the chosen palette before removing its temporary key. Follow-up failures corrected exact native-option selection, body-owned direction, and nested-route mounting. The first full matrix failed on that nested mount before the corrected four-case Installation probe passed. None of the failed or timed-out runs is counted as passing evidence.

The strict eleven-language build and ten-route DOM validator passed. Playwright 1.62.0 with Chromium 151.0.7922.34 and Axe Core 4.12.1 passed 60 base route/viewport/palette cases, 60 reviewed screenshot signatures, 60 complete focus cycles with 4,752 stops, 30 expanded-state Axe cases, four branch/drawer activation cases, 40 language switches, 60 version activation/closure cases, 60 search activation/closure cases, and 40 TOC activation cases. Every localized destination received another Axe pass. The complete keyboard inventory increased from 226 to 246 cases.

The complete documentation suite reported 58 passes and 1,408 subtests. A risk-proportional registry, speech-task, universal-optimization, packaging- metadata, distribution-compliance, and release-readiness slice reported 97 passes and 522 subtests. All 68 generated provider pages and 59 generated model notebooks remain current. The five-record release-alignment check passed, and refreshed inventories remain 68 models (34 TTS, 23 ASR, and 11 VAD), 102 aliases, zero invalid display names, six public optimization passes, 408 model/pass pairs, eight top-level navigation roots, ten representative routes, and eight contribution steps. The selected pre-commit sequence exited zero for every applicable hook; its Markdown hook reported no matching files and is not counted as a pass.

This closes language switching across the visible representative local desktop/tablet matrix. The mobile-hidden state remains covered by all 20 mobile base cases. Theme and source route-specific activation remain outside this slice. The complete Python suite and physical package builds were not rerun; their exact preceding current-worktree evidence remains separate. Exact-current-worktree remote CI, exact tagged artifact hashes, protected publisher configuration, tags, publication approval, five native-kernel hardware gates, and two inaccessible WeNet asset paths remain open. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree representative theme-control activation evidence

The next bounded documentation iteration refreshed the shared theme reference at Transformers commit b317ff31cd2491c2d4fc05d25fa06f35c527bcf6. Its 65-line docs/source/en/index.md source retains SHA-256 e24e0eddadccef8b15e3f25c55acb0df0f65de21a4b5bf709d119d22d0eb555b, and the 1,576-line _toctree.yml retains SHA-256 f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a. The official Home route returned HTTP 200. Later repeated rendered probes were rate-limited with HTTP 429 and are excluded rather than reported as passing comparison evidence.

The observable contract was fixed before implementation. Because the theme control is intentionally hidden at the shared mobile breakpoint, all ten representative routes must switch palette at desktop and tablet widths from both starting palettes. Twenty default-palette cases use Enter to switch to slate; 20 slate cases use pointer activation to switch to default. Each case requires the exact two toggle labels and target inputs, one visible native tab stop at the reference-sized 34 by 24-pixel geometry, unchanged route and English locale, focus on the newly visible toggle, exact target colors, zero overflow, and a zero-violation Axe result after activation.

The focused source contract intentionally failed before implementation because the checker had no theme activation matrix, helper, or counters. The first four-case rendered probe rejected a one-pixel y-coordinate error, and its corrected rerun passed. The first full matrix then exposed a two-animation- frame focus race between programmatic palette setup and native Tab traversal. The setup helper now waits for those exact frames. The previously failing Models tablet/slate route then passed all 239 focus stops, and a fresh complete matrix passed. None of the failed or interrupted commands is counted as passing evidence. The production theme template, JavaScript, and CSS required no change.

The strict eleven-language build and ten-route DOM validator passed. Playwright 1.62.0 with Chromium 151.0.7922.34 and Axe Core 4.12.1 passed 60 base route/viewport/palette cases, 60 reviewed screenshot signatures, 60 complete focus cycles with 4,709 stops, 30 expanded-state Axe cases, four branch/drawer activation cases, 40 theme switches, 40 language switches, 60 version activation/closure cases, 60 search activation/closure cases, and 40 TOC activation cases. Every switched theme received another Axe pass. The complete keyboard inventory increased from 246 to 266 cases.

The complete documentation suite reported 59 passes and 1,408 subtests. A risk-proportional registry, speech-task, universal-optimization, packaging- metadata, distribution-compliance, and release-readiness slice reported 97 passes and 522 subtests. All 68 generated provider pages and 59 generated model notebooks remain current. The five-record release-alignment check passed, and refreshed inventories remain 68 models (34 TTS, 23 ASR, and 11 VAD), 102 aliases, zero invalid display names, six public optimization passes, 408 model/pass pairs, eight top-level navigation roots, ten representative routes, and eight contribution steps. The selected pre-commit sequence exited zero for every applicable hook; its Markdown hook reported no matching files and is not counted as a pass.

This closes palette switching across the visible representative local desktop/tablet matrix. The mobile-hidden state remains covered by all 20 mobile base cases. Source-link route-specific activation remains outside this slice. The complete Python suite and physical package builds were not rerun; their exact preceding current-worktree evidence remains separate. Exact-current- worktree remote CI, exact tagged artifact hashes, protected publisher configuration, tags, publication approval, five native-kernel hardware gates, and two inaccessible WeNet asset paths remain open. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree representative source-link activation evidence

The next bounded documentation iteration refreshed the shared source-control reference at Transformers commit b317ff31cd2491c2d4fc05d25fa06f35c527bcf6. Its 65-line docs/source/en/index.md source retains SHA-256 e24e0eddadccef8b15e3f25c55acb0df0f65de21a4b5bf709d119d22d0eb555b, and the 1,576-line _toctree.yml retains SHA-256 f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a. The official Home route returned HTTP 200.

The observable contract was fixed before implementation. Because the source link is intentionally hidden at the shared mobile breakpoint, all ten representative routes must activate it at desktop and tablet widths in both palettes. Twenty default-palette cases use Enter and 20 slate cases use a pointer. Each requires exactly one native tab stop named Open VoiceHub source repository, the exact https://github.com/kadirnar/voicehub target, 55 by 16-pixel geometry at x = 198 and y = 155, a two-pixel focus outline, route and palette stability before activation, zero overflow, and a zero-violation Axe result. The checker then requires exact browser navigation while intercepting the declared external target with a deterministic fixture; GitHub availability and GitHub-page accessibility are not reported as local passes.

The focused contract first failed because no source activation matrix, helper, or counters existed. A later focused assertion rejected a quote-sensitive checker fragment and was corrected before the contract passed. The four-case Home render then passed both palettes and visible viewports. The first two full matrix attempts exposed a real sequential-focus race: closing Material search after native Tab could asynchronously restore focus to the search input and skip the source link. Neither run is counted. The production header control now cancels only that delayed restoration when Tab leaves the desktop input; Escape closure still restores focus. A 20-cycle focused probe and a fresh full matrix passed after the strict rebuild.

The strict eleven-language build and ten-route DOM validator passed. Playwright 1.62.0 with Chromium 151.0.7922.34 and Axe Core 4.12.1 passed 60 base route/viewport/palette cases, 60 reviewed screenshot signatures, 60 complete focus cycles with 4,621 stops, 30 expanded-state Axe cases, four branch/drawer activation cases, 40 source navigations, 40 theme switches, 40 language switches, 60 version activation/closure cases, 60 search activation/closure cases, and 40 TOC activation cases. Every focused source link received its own local-page Axe pass. The complete keyboard inventory is now 286 cases.

The complete documentation suite reported 60 passes and 1,408 subtests. A risk-proportional registry, speech-task, universal-optimization, packaging- metadata, distribution-compliance, and release-readiness slice reported 97 passes and 522 subtests. All 68 generated provider pages and 59 generated model notebooks remain current. The five-record release-alignment check passed, and refreshed inventories remain 68 models (34 TTS, 23 ASR, and 11 VAD), 102 aliases, zero invalid display names, six public optimization passes, 408 model/pass pairs, eight top-level navigation roots, ten representative routes, and eight contribution steps.

This closes source-link activation across the visible representative local desktop/tablet matrix; all 20 mobile base cases retain the intentional hidden state. Non-representative routes and future shared controls remain outside the matrix. The complete Python suite and physical package builds were not rerun; their exact preceding current-worktree evidence remains separate. Exact- current-worktree remote CI, exact tagged artifact hashes, protected publisher configuration, tags, publication approval, five native-kernel hardware gates, and two inaccessible WeNet asset paths remain open. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree root left-navigation activation evidence

The next bounded documentation-shell iteration refreshed the official navigation reference at Transformers commit b317ff31cd2491c2d4fc05d25fa06f35c527bcf6. Its 65-line docs/source/en/index.md source retains SHA-256 e24e0eddadccef8b15e3f25c55acb0df0f65de21a4b5bf709d119d22d0eb555b, and the 1,576-line _toctree.yml retains SHA-256 f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a. The official rendered Home route returned HTTP 200 and presents the same eight ordered top-level disclosures as the VoiceHub rail.

The observable contract was fixed before implementation. On the Pipeline route, every root branch must activate and restore at desktop and tablet widths in both palettes. Sixteen default-palette cases use Enter and 16 slate cases use a pointer. Each case requires the exact eight-label order, native checked state and disclosure ARIA, correctly labeled panel visibility, the active Pipeline link and unchanged route, palette stability, visible reference rail geometry, zero horizontal overflow, and an Axe pass whenever the branch is expanded. Keyboard activation also requires the two-pixel visible focus ring; pointer activation retains DOM focus without incorrectly displaying that keyboard-only ring. The mobile drawer remains covered separately.

The focused contract first failed because no root activation matrix, helper, or counters existed. Three rendered probes then rejected, respectively, a desktop coordinate applied to tablet, an expectation that ignored the final API control's legitimate eight-pixel focus scroll, and an incorrect demand for a keyboard-only focus ring after pointer input. Those expectations were fixed without changing production markup or styling, and none of the failed probes is counted. The complete 32-case rerun passed with Axe Core 4.12.1.

The strict eleven-language build and ten-route DOM validator passed. Playwright 1.62.0 with Chromium 151.0.7922.34 and Axe Core 4.12.1 passed 60 base route/viewport/palette cases, 60 reviewed screenshot signatures, 60 complete focus cycles with 4,737 stops, 30 expanded-state Axe cases, 32 root-branch activation/restoration cases, two mobile drawer activation cases, 40 source navigations, 40 theme switches, 40 language switches, 60 version activation/closure cases, 60 search activation/closure cases, and 40 TOC activation cases. The complete keyboard inventory is now 300 cases.

The complete documentation suite reported 61 passes and 1,408 subtests. A risk-proportional registry, speech-task, universal-optimization, packaging- metadata, distribution-compliance, and release-readiness slice reported 97 passes and 522 subtests. All 68 generated provider pages and 59 generated model notebooks remain current. The five-record release-alignment check passed, and refreshed inventories remain 68 models (34 TTS, 23 ASR, and 11 VAD), 102 aliases, zero invalid display names, six public optimization passes, 408 model/pass pairs, eight top-level navigation roots, ten representative routes, and eight contribution steps. The selected pre-commit sequence exited zero for every applicable hook; its Markdown hook reported no matching files and is not counted as a pass.

This closes root-branch activation on the representative Pipeline route across the visible desktop/tablet matrix. Nested branch activation and sticky behavior outside that recorded route remain pending. The complete Python suite and physical package builds were not rerun; their preceding exact-current-worktree evidence remains separate. Exact-current-worktree remote CI, exact tagged artifact hashes, protected publisher configuration, tags, publication approval, five native-kernel hardware gates, and two inaccessible WeNet asset paths remain open. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree API model-navigation hierarchy evidence

The next bounded documentation iteration refreshed Transformers main at commit b317ff31cd2491c2d4fc05d25fa06f35c527bcf6. The 65-line docs/source/en/index.md source retains SHA-256 e24e0eddadccef8b15e3f25c55acb0df0f65de21a4b5bf709d119d22d0eb555b, and the 1,576-line _toctree.yml retains SHA-256 f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a. That source places Auto Classes under API → Main Classes and model guides, including SpeechT5, under API → Models. The official Home and Auto Classes routes returned HTTP 200. The SpeechT5 shell rendered its selected model link, but its content pane was rate-limited with a visible 429. That current content capture is unavailable and is not counted as a page-content pass.

The shared generator writes Auto Classes exactly once under API → Main Classes and all 68 unique model paths under Base classes → Models, grouped as 34 TTS, 23 ASR, and 11 VAD guides. Base classes retains lifecycle and support material, contribution, preprocessors, and architecture. Auto Classes expands API → Main Classes; SpeechT5 expands Base classes → Models → Text to speech. Model routes remain stable, and the generator is idempotent.

The first complete visual run rejected a stale Quickstart footer expectation after the correct navigation move; that failed run is not counted. The strict eleven-language build and ten-route DOM validator then passed. Reviewed local light and dark captures covered Auto Classes and SpeechT5 at desktop, tablet, and mobile widths. The complete Playwright 1.62.0, Chromium 151.0.7922.34, and Axe Core 4.12.1 matrix passed 60 base cases, all 60 existing screenshot signatures, 60 complete focus cycles with 4,745 stops, 36 nested-branch activation/restoration cases, and 348 keyboard cases. The final model-route slice revalidated exact Base classes ancestry, screenshots, Axe, and sticky navigation behavior after formatting and the final strict build.

The complete documentation suite reported 62 passes and 1,476 subtests. The risk-proportional registry, speech-task, universal-optimization, packaging- metadata, distribution-compliance, and release-readiness slice reported 97 passes and 522 subtests. All 68 model pages and 59 model notebooks remain current. The five-record release-alignment check passed, and refreshed inventories remain 68 models, 102 aliases, zero invalid display names, six public optimization passes, 408 model/pass pairs, eight top-level navigation roots, ten representative routes, and eight contribution steps. The selected pre-commit sequence passed every applicable hook on its second invocation; the first invocation changed formatting and is not counted as a pass, while the Markdown hook found no matching files in either run.

This closes the model-navigation placement gap. The current official SpeechT5 content capture remains rate-limited and unpassed. Nested branch activation and sticky behavior outside the existing representative matrix remain pending. The complete Python suite and physical package builds were not rerun for this navigation-only slice; their preceding exact-current-worktree evidence remains separate. Exact-current-worktree remote CI, exact tagged artifact hashes, protected publisher configuration, tags, publication approval, five native- kernel hardware gates, and two inaccessible WeNet asset paths remain open. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree nested navigation and sticky-rail evidence

The next bounded shell iteration refreshed Transformers main at commit 339c18c08bf0c143b8307c255004506e358984f2. The 65-line docs/source/en/index.md source remains SHA-256 e24e0eddadccef8b15e3f25c55acb0df0f65de21a4b5bf709d119d22d0eb555b, and the 1,576-line _toctree.yml remains SHA-256 f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a. The official SpeechT5 route returned HTTP 200 and rendered current content at desktop, tablet, and mobile widths without the earlier 429 response. The upstream rail still selects SpeechT5 inside the API model inventory; VoiceHub retains original speech-specific content and its approved color palette.

The observable contract was fixed before implementation on VoiceHub's deepest representative route. Its nine visible nested disclosure paths are Base classes → Models, Base classes → Models → Catalogs and support, Base classes → Models → Text to speech, Base classes → Models → Text to speech → SpeechT5, Base classes → Models → Automatic speech recognition, Base classes → Models → Voice activity detection, Base classes → Models → Contribute, Base classes → Preprocessors, and Base classes → Architecture. Each must activate and return to its exact initial checked, aria-expanded, controlled-panel, focus, route, active- link, palette, and overflow state at desktop and tablet widths in both palettes. Eighteen default-palette cases use Enter and 18 slate cases use a pointer. Every target also scrolls the document to 320 pixels and requires the 270-pixel sticky rail to move from top 65 with viewport-minus-65 height to top 0 with full viewport height while the shell offset becomes 65 pixels. Axe runs in all 24 sticky target states.

The focused source test first failed because no nested matrix, helper, or counters existed. The first rendered case then rejected a checker query that included the active Usage table-of-contents link after scrolling; that run is not a pass. The corrected query excludes embedded secondary navigation. The first complete nested matrix reached the active SpeechT5 disclosure and then failed Axe with four landmark-unique violations because its expanded page table of contents duplicated unnamed or identically named subsection landmarks. That run is also failed evidence. The shared navigation initializer now gives every nested page-section landmark a deterministic SpeechT5-scoped accessible name before the panel opens. The corrected 24-case focused render passed with Axe Core 4.12.1.

The strict eleven-language build and ten-route DOM validator passed. The full Playwright 1.62.0, Chromium 151.0.7922.34, and Axe Core 4.12.1 matrix passed 60 base cases, all 60 existing screenshot signatures, 60 complete native focus cycles with 4,628 stops, 30 expanded-state Axe cases, 32 root-branch cases, 24 nested-branch/sticky cases, two mobile drawer cases, 40 source navigations, 40 theme switches, 40 language switches, 60 version cases, 60 search cases, and 40 TOC cases. The complete keyboard inventory is now 312 cases.

The complete documentation suite reported 63 passes and 1,476 subtests. The risk-proportional registry, speech-task, universal-optimization, packaging- metadata, distribution-compliance, and release-readiness slice reported 97 passes and 522 subtests. All 68 generated model pages and 59 notebooks remain current, and the five-record release-alignment check passed. Refreshed inventories remain 68 models (34 TTS, 23 ASR, and 11 VAD), 102 aliases, zero invalid display names, six public optimization passes, 408 model/pass pairs, eight top-level navigation roots, ten representative routes, and eight contribution steps.

This closes nested disclosure activation and sticky-rail behavior on the deepest representative model route. Non-representative navigation structures, remaining page-action states, the object-by-object public API audit, and exact comparison with the current Modular Transformers contribution path remain open. The complete Python suite and physical package builds were not rerun for this documentation-only slice; their preceding exact-current-worktree evidence remains separate. Exact-current-worktree remote CI, exact tagged artifact hashes, protected publisher configuration, tags, publication approval, five native-kernel hardware gates, and two inaccessible WeNet asset paths also remain open. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree representative page-action evidence

The page-action iteration refreshed Transformers main at commit 179a24360e55f1daff1bc20f0d11f4e8f47a6f44. The official 65-line docs/source/en/index.md and 1,576-line _toctree.yml retain SHA-256 values e24e0eddadccef8b15e3f25c55acb0df0f65de21a4b5bf709d119d22d0eb555b and f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a. The current SpeechT5 page returned HTTP 200 and its rendered source contains the shared Copy page control and DocFooterNav adjacency for Speech2Text and UniSpeech. That is current structural reference evidence, not a claim that an external page or destination passed VoiceHub's local accessibility checks.

The observable local contract covers all ten representative routes at desktop, tablet, and mobile widths in both palettes. Thirty default-palette cases use Enter and 30 slate cases use a pointer. Each case requires exactly one native edit link, the route's exact optional previous link and exact next link, and one native Back to top button. Edit activation must navigate to the declared GitHub URL through a deterministic local intercept. Footer activation must perform real local browser navigation to every declared destination. Back to top must appear after the documented down-then-up scroll direction, expose a visible two-pixel focus outline, return to scroll position zero, preserve the route and palette, introduce no horizontal overflow, and pass Axe in its visible focused state.

The focused source contract first failed because no page-action inventory, helper, matrix, or counters existed. A bare python invocation was unavailable in the shell and is not test evidence; the repository virtual environment then produced the intended failing regression. The first rendered case timed out because Material reveals Back to top only after upward scroll. The corrected probe then exposed a real missing focus outline on that button. After the shared edit/footer/top focus rule was added, Axe caught a transient 2.05:1 contrast state during the declared 125-millisecond color/background transition. None of those runs is counted as passing. The final checker models the directional reveal and waits 150 milliseconds for the transition to settle before the accessibility audit.

The complete Playwright 1.62.0, Chromium 151.0.7922.34, and Axe Core 4.12.1 matrix passed 60 base cases, all 60 screenshot signatures, 60 complete focus cycles with 4,576 stops, 30 expanded-state Axe cases, 32 root disclosures, 24 nested sticky disclosures, two mobile drawer activations, and the existing source, theme, language, version, search, and TOC matrices. The new slice passed 60 page-action cases, 60 exact edit activations, 114 exact footer navigations, 60 Back to top activations, and 60 focused interaction-state Axe audits. Its 30 keyboard cases raise the aggregate keyboard inventory to 342.

The strict eleven-language build, ten-route DOM validator, 64 documentation tests with 1,476 subtests, and the 97-test release-risk slice with 522 subtests passed. All 68 generated model pages and 59 generated notebooks remain current, and the five-record release-alignment check passed. A fresh distribution check passed wheel, sdist, and editable installs with 68 models, 81 provenance manifests, 193 compliance files, required package data, zero runtime dependency violations, and no eager PyTorch import. It produced a 57,192,297-byte wheel and a 55,450,920-byte sdist. A separate local fingerprint build produced:

Artifact	Bytes	SHA-256
voicehub-0.3.0-py3-none-any.whl	57,192,297	d6fc30aa0d037d6624425e877fac6d5d356ad07bc901adccdc72b8c930304d98
voicehub-0.3.0.tar.gz	55,450,952	2a9becc4014c7e487eb1a60d1cd1009977dfd8df2f493397b8c7296d522326a7

The first selected pre-commit sequence is not a pass: docformatter exited 3 after YAPF collapsed three multiline JavaScript object callbacks. A standalone docformatter run restored them, proving a formatter oscillation rather than a source failure. Those callbacks now use equivalent concatenated strings. The complete selected sequence and git diff --check then exited zero; its Markdown hook had no matching files and is not counted as a pass.

This closes actual page-action behavior on the complete representative local matrix. Non-representative routes and future controls remain outside that matrix, and the deterministic edit intercept does not make any claim about GitHub's page content or availability. Exact-current-worktree remote Linux and Windows CI, exact tagged-workflow artifacts, protected publisher configuration, tags, publication approval, five native-kernel hardware gates, and two inaccessible WeNet asset paths remain open. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree AutoConfig subfolder evidence

This bounded public-lifecycle iteration used Transformers main commit 179a24360e55f1daff1bc20f0d11f4e8f47a6f44. Its current AutoConfig.from_pretrained() delegates configuration discovery to PreTrainedConfig.get_config_dict(), whose _get_config_dict() consumes subfolder and passes it to the cached-file resolver. VoiceHub's concrete configuration loader already followed that contract, but automatic model-type discovery omitted subfolder and searched only the repository root.

The observable regression uses a local artifact whose only config.json is nested/config.json. AutoConfig.from_pretrained(root, subfolder="nested") must discover and restore its canonical vits type, and AutoModelForTextToSpeech.from_pretrained() must preserve the same path through config_kwargs without loading the model runtime. Both tests first failed at the root lookup with FileNotFoundError; after the discovery resolver received the existing loader option, both passed. A bare python invocation failed with exit 127 because that shell alias is absent and is not counted as test evidence; the repository virtual environment produced the intended red and green runs.

The focused auto-configuration file passed 16 tests. The configuration, Hub, pipeline, inference, task-registry, and registry slice passed 141 tests and 364 subtests. The registry-related optimization slice passed 55 tests and 204 subtests. Documentation passed 64 tests and 1,476 subtests, the strict eleven-language build, the ten-route DOM validator, all 68 model-page and 59 notebook freshness checks, and the complete Playwright/Axe matrix. That matrix reported 60 base and screenshot cases, 60 focus cycles with 4,583 stops, 342 keyboard cases, and all existing copy, search, version, language, theme, source, TOC, root, nested, and page-action interaction inventories. The first DOM invocation used an unsupported --site-dir option, exited 2, and is not a pass; the positional invocation then passed.

The complete Python 3.12.12 suite passed 2,499 executed tests and 3,894 subtests with 35 warnings in 105.59 seconds. Its 15 skips are not passes:

• three complete-dependency import checks remain assigned to the default-runtime CI job;
• three Triton and two compiled CUDA-extension checks require their declared CUDA hosts;
• ESPNet, NeMo QuartzNet, SenseVoice, SpeechBrain, and TEN-VAD each require their declared opt-in release asset or checkpoint; and
• WeNet conversion and tokenizer validation require two currently unavailable asset paths.

Some opt-in assets have separate earlier successful evidence in the candidate matrix; this run does not relabel its skips as successes. The release, packaging-metadata, and distribution-compliance slice passed 23 tests and 76 subtests. A fresh physical distribution check passed wheel, sdist, and editable installs with 68 models, 81 provenance manifests, 193 compliance files, required package data, zero runtime dependency violations, and no eager PyTorch import. It produced a 57,192,308-byte wheel and a 55,451,111-byte sdist.

The selected-file pre-commit sequence passed every applicable source hook in one invocation. Its Markdown hook found no matching files and is not counted as a pass.

This closes automatic configuration discovery from local or Hub subfolders and its task-factory propagation. The broader object-by-object public API audit remains open. Exact-current-worktree Linux, Windows, Python 3.10, Python 3.11, default-runtime, and tagged-workflow evidence; protected publisher configuration; tags; publication approval; the five CUDA/Triton gates; and the listed opt-in asset paths remain pending. No protected action was taken. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree configuration-secret boundary evidence

This bounded data-safety iteration refreshed Transformers main at commit 179a24360e55f1daff1bc20f0d11f4e8f47a6f44; the 1,576-line navigation and 500-line Modular Transformers source retain SHA-256 values f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a and a52af1ee1754e6e658535f071a727fac31a057a3a848c28c18873a5ff371dd96. The upstream revision is reference context. VoiceHub's safe checkpoint and runtime-credential policy is the controlling speech-domain requirement for this slice.

The observable contract requires VoiceHubConfig construction to reject top-level and nested credential-shaped fields, rejects secrets attached after construction before dictionary or file serialization, and validates the final payload returned by a subclass before diff, JSON, representation, or checkpoint output. A concrete loader's top-level Hub token remains runtime-only and is omitted, while model fields such as pad_token_id remain serializable. A failed checkpoint write must not create config.json.

The first focused run proved the gap: two base tests and 35 of 68 registry subtests failed, while 33 configurations passed only because they carried their own local protection. The shared constructor and dictionary boundary then made all 68 subtests pass. The first broader base/registry run is not a pass: four TTS configs intentionally accepted runtime Hub tokens and attempted to remove them after the new base check, producing seven failures. The common boundary now omits only the top-level runtime token before validating all remaining state; no provider branch was added. A separate malicious-subclass regression first showed that final serialization and repr could still expose a secret added after the base call. Final-payload validation closed those paths. The focused result is five tests with 68 registry subtests passed.

The configuration, processing, inference, and provider security slice passed 225 tests and 348 subtests. The registry, universal-optimization, packaging, distribution-compliance, and release-policy slice passed 95 tests and 473 subtests. The complete Python 3.12.12 suite passed 2,504 executed tests and 3,962 subtests with 35 warnings in 106.84 seconds. Its 15 skips remain unpassed: three complete-dependency default-runtime CI checks, three Triton and two CUDA-extension checks, five opt-in ESPNet/NeMo/SenseVoice/SpeechBrain/ TEN-VAD asset or oracle checks, and two inaccessible WeNet paths.

Documentation passed 64 tests and 1,476 subtests, all 68 model-page and 59 model-notebook freshness checks, the strict eleven-language build, and the ten-route DOM validator. The complete Playwright 1.62.0, Chromium 151.0.7922.34, and Axe Core 4.12.1 matrix passed all 60 route/viewport/palette cases, 60 screenshots, and 60 focus cycles with 4,610 visible focus steps and 342 keyboard cases. Its complete interaction inventory passed search, version, language, theme, source, table-of-contents, disclosure, page-action, copy, and accessibility paths. A fresh physical distribution check passed wheel, sdist, and editable installs with 68 models, 81 provenance manifests, 193 compliance files, required package data, zero runtime dependency violations, and no eager PyTorch import. It produced a 57,192,400-byte wheel and a 55,451,994-byte sdist.

The first selected-file pre-commit invocation is not a pass because YAPF rewrote the new code. The focused contract passed after that rewrite, and the second invocation passed every applicable source hook. The Markdown hook found no matching files and is not counted as a pass.

This closes the shared configuration credential boundary for the complete registry. The broader object-by-object public API audit remains open. Exact-current-worktree Linux, Windows, Python 3.10, Python 3.11, default-runtime, tagged-workflow, and publisher evidence; publication approval; five CUDA/Triton gates; and the seven opt-in asset paths remain pending. No protected action was taken. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree generation-configuration secret boundary evidence

This bounded data-safety iteration refreshed Transformers main at commit d79d0c1ed3e8fc3f4742cc34aee46d92abd0ea44. Its 1,576-line navigation and 500-line Modular Transformers source retain SHA-256 values f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a and a52af1ee1754e6e658535f071a727fac31a057a3a848c28c18873a5ff371dd96. The upstream revision is structural context; VoiceHub's safe speech-checkpoint boundary controls this slice.

The observable contract requires standalone TTSGenerationConfig construction and untrusted checkpoint loading to reject top-level or nested credential-shaped fields. Secrets attached later must fail before dictionary, representation, or file serialization, and an unsafe subclass payload must be revalidated by representation and checkpoint output. A failed write must not create generation_config.json. The Hub loader token remains runtime-only, while legitimate model settings such as pad_token_id, eos_token_id, and max_new_tokens remain serializable. Embedded generation defaults must enter the same configuration constructor boundary for every registered model.

A direct pre-implementation probe confirmed that standalone generation configuration serialized a nested api_key. The first focused contract is not a pass: four tests failed on embedded construction, standalone construction, post-construction serialization, and final subclass output, while the existing runtime-only Hub token test passed. A separate pre-fix registry inventory found 18 configurations that still accepted the embedded secret at construction; 39 rejected it through the existing shared boundary and 11 through local credential-specific checks. The common explicit-field check now rejects all 68 at construction without a provider branch.

After correction, the focused contract passed seven tests and 68 registry subtests. The broader generation, configuration, auto-loading, pipeline, task-registry, and pretrained TTS slice passed 135 tests and 314 subtests. The dedicated registry and public-optimization slice passed 169 tests and 486 subtests with four recorded framework warnings. The complete Python 3.12.12 suite passed 2,510 executed tests and 4,030 subtests with 35 warnings in 106.64 seconds. Its 15 skips remain unpassed: three complete-dependency default-runtime checks, three Triton and two CUDA-extension checks, five opt-in ESPNet/NeMo/SenseVoice/SpeechBrain/TEN-VAD asset or oracle checks, and two inaccessible WeNet paths.

Documentation and scaffold validation passed 93 tests and 1,511 subtests; all 68 model pages, 68 model navigation entries, 59 model notebooks, and five benchmark records remained current. The strict eleven-language build and the ten-route DOM validator passed. The unchanged documentation shell retains the immediately preceding complete Playwright/Axe evidence; the visual matrix was not rerun for this API-safety and prose-only slice and is not presented as a new pass.

A fresh physical distribution check passed wheel, sdist, and editable installs with 68 models, 81 provenance manifests, 193 compliance files, required package data, zero runtime dependency violations, and no eager PyTorch import. It produced a 57,192,472-byte wheel and a 55,453,207-byte sdist. The first selected-file pre-commit invocation is not a pass because YAPF rewrote the new test formatting. The formatted focused contract passed, and the second invocation passed every applicable source hook. The Markdown hook found no matching files and is not counted as a pass.

This closes credential handling for both embedded and standalone generation configuration. The broader object-by-object public API audit remains open. Exact-current-worktree Linux, Windows, Python 3.10, Python 3.11, default-runtime, tagged-workflow, and publisher evidence; publication approval; five CUDA/Triton gates; and the seven opt-in asset paths remain pending. No protected action was taken. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree native-manifest secret boundary evidence

This bounded data-safety iteration refreshed Transformers main at commit d79d0c1ed3e8fc3f4742cc34aee46d92abd0ea44. Its 1,576-line navigation and 500-line Modular Transformers source retain SHA-256 values f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a and a52af1ee1754e6e658535f071a727fac31a057a3a848c28c18873a5ff371dd96. The upstream revision is structural context; VoiceHub's safe portable-artifact boundary controls this slice.

The observable contract requires VoiceHubManifest.metadata construction and untrusted manifest loading to reject top-level or nested credential-shaped fields. Because the frozen manifest contains a mutable metadata mapping, to_dict() and save() must revalidate post-construction changes. Save must also validate the final payload returned by a subclass before creating the artifact directory, manifest, or temporary file. Safe descriptive fields such as token_count must continue to round-trip.

The first focused contract is not a pass: all four new tests failed because construction, untrusted loading, post-construction dictionary output, and a malicious subclass save accepted secrets. The shared secret detector was moved to the shallow serialization layer while remaining available from the existing configuration module for compatibility. Manifest construction and dictionary output now validate metadata, and save validates the complete final payload before any filesystem mutation. The corrected focused contract passed four tests with 12 deselected. The checkpoint, configuration, inference, and task slice passed 132 tests and 294 subtests. The registry, public-optimization, checkpoint, distribution-compliance, and release-policy slice passed 269 tests and 1,339 subtests with four framework warnings.

Documentation passed 64 tests and 1,476 subtests; all 68 model pages and 59 model notebooks remained current. The strict eleven-language build and ten-route DOM validator passed. The unchanged documentation shell retains the immediately preceding complete Playwright/Axe evidence; the visual matrix was not rerun for this API-safety and prose-only slice and is not presented as a new pass. An initial notebook command referenced a nonexistent scripts/check_model_notebooks.py, exited 2, and is not a pass; the repository's actual generate_model_notebooks.py --check command then passed.

The complete Python 3.12.12 suite passed 2,514 executed tests and 4,030 subtests with 35 warnings in 105.83 seconds. Its 15 skips remain unpassed: three complete-dependency default-runtime checks, three Triton and two CUDA-extension checks, five opt-in ESPNet/NeMo/SenseVoice/SpeechBrain/TEN-VAD asset or oracle checks, and two inaccessible WeNet paths. A fresh physical distribution check passed wheel, sdist, and editable installs with 68 models, 81 provenance manifests, 193 compliance files, required package data, zero runtime dependency violations, and no eager PyTorch import. It produced a 57,192,624-byte wheel and a 55,452,570-byte sdist. The selected-file pre-commit invocation passed every applicable source hook; its Markdown hook found no matching files and is not counted as a pass.

This closes credential handling for native manifest metadata and keeps the credential detector in one serialization contract. The broader object-by-object public API audit remains open. Exact-current-worktree Linux, Windows, Python 3.10, Python 3.11, default-runtime, tagged-workflow, and publisher evidence; publication approval; five CUDA/Triton gates; and the seven opt-in asset paths remain pending. No protected action was taken. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree ASR/VAD inference-configuration secret boundary evidence

This bounded data-safety iteration refreshed Transformers main at commit d79d0c1ed3e8fc3f4742cc34aee46d92abd0ea44. Its 1,576-line navigation and 500-line Modular Transformers source retain SHA-256 values f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a and a52af1ee1754e6e658535f071a727fac31a057a3a848c28c18873a5ff371dd96. The upstream revision is structural context; VoiceHub's safe public task- configuration boundary controls this slice.

The observable contract requires ASR and VAD inference configuration construction and untrusted checkpoint loading to reject top-level or nested credential-shaped fields. Mutable state must be rechecked, and representation and checkpoint output must validate the final payload returned by a subclass. A failed write must not create transcription_config.json, vad_config.json, or their artifact directory. The Hub loader token remains runtime-only, while legitimate task fields such as max_new_tokens continue to round-trip.

The pre-implementation focused run is not a pass: two tests failed because a malicious subclass payload reached representation and checkpoint output; the untrusted-load test passed with two task subtests. Final-payload validation was added to the shared SpeechInferenceConfig representation and save boundaries without a task or provider branch. The corrected focused contract passed five tests and two subtests. The speech-core, base-API, inference-contract, and task- registry slice passed 120 tests and 319 subtests. The registry, native-artifact, universal-optimization, documentation, scaffold, release-policy, and distribution-compliance slice passed 173 tests and 1,934 subtests.

The complete Python 3.12.12 suite passed 2,517 executed tests and 4,032 subtests with 35 warnings in 120.34 seconds. Its 15 skips remain unpassed: three complete-dependency default-runtime checks, three Triton and two CUDA- extension checks, five opt-in ESPNet/NeMo/SenseVoice/SpeechBrain/TEN-VAD asset or oracle checks, and two inaccessible WeNet paths.

All 68 model pages and 59 model notebooks remained current. The strict eleven- language build and ten-route DOM validator passed. The unchanged documentation shell retains the immediately preceding complete Playwright/Axe evidence; the visual matrix was not rerun for this API-safety and prose-only slice and is not presented as a new pass. A fresh physical distribution check passed wheel, sdist, and editable installs with 68 models, 81 provenance manifests, 193 compliance files, required package data, zero runtime dependency violations, and no eager PyTorch import. It produced a 57,192,647-byte wheel and a 55,453,587-byte sdist. The selected-file pre-commit invocation passed every applicable source hook; its Markdown hook found no matching files and is not counted as a pass.

This closes final-payload credential handling for public ASR and VAD inference configuration. The broader object-by-object public API audit remains open. Exact-current-worktree Linux, Windows, Python 3.10, Python 3.11, default-runtime, tagged-workflow, and publisher evidence; publication approval; five CUDA/Triton gates; and the seven opt-in asset paths remain pending. No protected action was taken. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree processor secret boundary evidence

This bounded data-safety iteration refreshed Transformers main at commit d09f53a801f45ad73ec3510e17972024234bc0fd. The rendered Home and Modular Transformers routes returned current content. The 1,576-line navigation and 500-line Modular Transformers source retain SHA-256 values f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a and a52af1ee1754e6e658535f071a727fac31a057a3a848c28c18873a5ff371dd96. The upstream revision is structural context; VoiceHub's safe processor- artifact boundary controls this slice. An initial read-only refresh command was rejected before execution because it included a disallowed cleanup form; it is not evidence. The corrected command used a temporary directory without that cleanup and produced the revision and fingerprints above.

The observable contract requires VoiceHubProcessor and AudioProcessor construction and untrusted artifact loading to reject top-level or nested credentials. Dictionary and checkpoint serialization must recheck mutable state, and save must validate the final mapping returned by a subclass before creating processor_config.json or its artifact directory. A Hub loader token remains runtime-only, while safe construction fields such as normalization continue to round-trip.

The pre-implementation focused run is not a pass: four tests plus four task subtests passed, while the malicious-subclass save test failed because its final payload reached the file writer. The shared processor save boundary now validates that complete mapping before filesystem mutation, without a task or provider branch. The corrected focused contract passed five tests and four subtests. The first selected pre-commit invocation is not a pass because YAPF rewrote the new test; the formatted focused contract and the second invocation then passed every applicable source hook. Its Markdown hook found no matching files and is not counted as a pass.

The base API, auto-loading, speech-core, task-registry, inference, pipeline, and registry slice passed 165 tests and 524 subtests. The universal- optimization, documentation, scaffold, release-policy, and distribution- compliance slice passed 156 tests and 2,104 subtests. The complete Python 3.12.12 suite passed 2,522 executed tests and 4,036 subtests with 35 warnings in 122.47 seconds. Its 15 skips remain unpassed: three complete-dependency default-runtime checks, three Triton and two CUDA-extension checks, five opt-in ESPNet/NeMo/SenseVoice/SpeechBrain/TEN-VAD asset or oracle checks, and two inaccessible WeNet paths.

All 68 model pages and 59 model notebooks remained current. The strict eleven- language build and ten-route DOM validator passed. The unchanged documentation shell retains the immediately preceding complete Playwright/Axe evidence; the visual matrix was not rerun for this API-safety and prose-only slice and is not presented as a new pass. A fresh physical distribution check passed wheel, sdist, and editable installs with 68 models, 81 provenance manifests, 193 compliance files, required package data, zero runtime dependency violations, and no eager PyTorch import. It produced a 57,192,663-byte wheel and a 55,453,948-byte sdist.

This closes final-payload credential handling for public processor artifacts. The broader object-by-object public API audit remains open. Exact-current- worktree Linux, Windows, Python 3.10, Python 3.11, default-runtime, tagged- workflow, and publisher evidence; publication approval; five CUDA/Triton gates; and the seven opt-in asset paths remain pending. No protected action was taken. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree TrainerState secret boundary evidence

This bounded public-API safety iteration refreshed Transformers main at commit d09f53a801f45ad73ec3510e17972024234bc0fd. Its 1,576-line navigation and 500-line Modular Transformers sources retain SHA-256 values f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a and a52af1ee1754e6e658535f071a727fac31a057a3a848c28c18873a5ff371dd96. The current 784-line upstream trainer_callback.py has SHA-256 a9703b60f3d585627054ed976eca3dbea69fddcfd4eb01b1cc19bff67bccf00a. That dataclass and JSON shape are structural context; VoiceHub's safe portable- checkpoint boundary controls this slice.

The observable contract requires nested credential-shaped values to fail at TrainerState construction and untrusted JSON loading. Because log_history is mutable, the final state write must revalidate post-construction changes and a subclass's complete dataclass payload before creating the output path. Trainer.log() must reject the same values before state mutation or callback dispatch. Safe metric names such as loss, step, and token_count must continue to round-trip.

The pre-implementation focused contract is not a pass: all five new tests failed because construction and loading accepted nested secrets, mutated and subclass payloads reached the file writer, and unsafe live logs reached state and callbacks. The shared serialization detector now validates state at construction and immediately before output; the live log path validates the complete normalized record before mutation. No task or provider branch was added. The corrected focused contract passed five tests with 19 deselected, and the complete trainer file passed 24 tests and 14 subtests. The training, checkpoint, callback, and integration slice passed 185 tests and 188 subtests. The registry, speech-task, universal-optimization, packaging, distribution- compliance, and release-policy slice passed 99 tests and 658 subtests.

The complete Python 3.12.12 suite passed 2,527 executed tests and 4,036 subtests with 35 warnings in 129.02 seconds. Its 15 skips remain unpassed: three complete-dependency default-runtime checks, three Triton and two CUDA- extension checks, five opt-in ESPNet/NeMo/SenseVoice/SpeechBrain/TEN-VAD asset or oracle checks, and two inaccessible WeNet paths.

Documentation passed 64 tests and 1,476 subtests. All 68 model pages and 59 model notebooks remained current, the five-record release-alignment check passed, and the strict eleven-language build completed in 38.60 seconds. The ten-route DOM validator retained all eight ordered navigation roots. The unchanged documentation shell retains the immediately preceding complete Playwright/Axe evidence; the visual matrix was not rerun for this API-safety and prose-only slice and is not presented as a new pass.

A fresh physical distribution check passed wheel, sdist, and editable installs with 68 models, 81 provenance manifests, 193 compliance files, required package data, zero runtime dependency violations, and no eager PyTorch import. It produced a 57,192,782-byte wheel and a 55,454,411-byte sdist. The first selected-file pre-commit invocation is not a pass because YAPF rewrote the new test formatting. The corrected focused contract and second invocation passed every applicable source hook. The Markdown hook found no matching files and is not counted as a pass.

Refreshed inventories remain 68 models (34 TTS, 23 ASR, and 11 VAD), 102 aliases, zero invalid display names, 68 provider pages with no missing or orphaned page, six public optimization passes, 408 model/pass pairs, eight top-level navigation roots, ten representative routes, eight contribution steps, and 261 public exports. The public root import still avoids PyTorch.

This closes credential handling for public trainer logs and state artifacts. The broader object-by-object public API audit remains open. Exact-current- worktree Linux, Windows, Python 3.10, Python 3.11, default-runtime, tagged- workflow, and publisher evidence; publication approval; five CUDA/Triton gates; and the seven opt-in or inaccessible asset paths remain pending. No protected action was taken. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree checkpoint-manifest secret boundary evidence

This bounded public-API safety iteration refreshed Transformers main at commit d09f53a801f45ad73ec3510e17972024234bc0fd. Its 1,576-line navigation and 500-line Modular Transformers sources retain SHA-256 values f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a and a52af1ee1754e6e658535f071a727fac31a057a3a848c28c18873a5ff371dd96. The pinned 4,460-line upstream trainer.py has SHA-256 6ea719d4f225c6fcbccfcd2137cdfb304426c93c4dd65983ae4186f6e8fb3d8d. That Trainer lifecycle is structural context; VoiceHub's safe exact-resume artifact boundary controls this slice.

The observable contract requires normalized dataset, collator, stateful- callback, optimizer, and scheduler fingerprints to reject nested credential- shaped fields without reading or reporting their values. The complete checkpoint manifest, including output added by a Trainer subclass, must be revalidated immediately before its atomic write. An untrusted manifest must fail before model or runtime restoration. A rejected save must leave neither a final checkpoint nor an incomplete checkpoint directory, while safe identity and metric fields such as dataset_id and token_count remain serializable.

The pre-implementation focused command is not a pass: two tests and five credential-bearing subtests failed because unsafe fingerprints, subclass manifest output, and an untrusted manifest were accepted. Its safe fingerprint assertion completed, but the failed command is excluded from passing evidence. The shared fingerprint normalizer, final manifest write boundary, and untrusted manifest read boundary now apply the same credential detector without a task or provider branch. The corrected focused contract passed three tests and five subtests with 33 deselected. The complete training-runtime file passed 36 tests and five subtests. The training, checkpoint, callback, and integration slice passed 294 tests and 205 subtests with four warnings.

The registry, public-API, optimization, documentation, scaffold, release- policy, and distribution-compliance slice passed 326 tests and 2,737 subtests with four warnings. The complete Python 3.12.12 suite passed 2,530 executed tests and 4,041 subtests with 35 warnings in 127.62 seconds. Its 15 skips remain unpassed: three complete-dependency default-runtime checks, three Triton and two CUDA-extension checks, five opt-in ESPNet/NeMo/SenseVoice/SpeechBrain/TEN- VAD asset or oracle checks, and two inaccessible WeNet paths.

All 68 model pages and 59 model notebooks remained current, and the five- record release-alignment check passed. The strict eleven-language build completed in 37.39 seconds. The ten-route DOM validator retained all eight ordered navigation roots. An initial DOM invocation used an unsupported --site-dir option and is not evidence; the corrected positional invocation passed. The unchanged visual shell retains its preceding Playwright/Axe evidence; no fresh visual-parity claim is made for this API-safety and prose- only slice.

A fresh physical distribution check passed wheel, sdist, and editable installs with 68 models, 81 provenance manifests, 193 compliance files, required package data, zero runtime dependency violations, and no eager PyTorch import. It produced a 57,192,833-byte wheel and a 55,455,064-byte sdist. The selected- file pre-commit invocation passed every applicable source hook. Its Markdown hook found no matching files and is not counted as a pass.

Refreshed inventories remain 68 models (34 TTS, 23 ASR, and 11 VAD), 102 aliases, zero invalid display names, 68 provider pages with no missing or orphaned page, six public optimization passes, 408 model/pass pairs, eight top-level navigation roots, ten representative routes, eight contribution steps, and 261 public exports. The public root import still avoids PyTorch.

This closes credential handling for exact-resume checkpoint manifests. The broader object-by-object public API audit remains open. Exact-current-worktree Linux, Windows, Python 3.10, Python 3.11, default-runtime, tagged-workflow, and publisher evidence; publication approval; five CUDA/Triton gates; and the seven opt-in or inaccessible asset paths remain pending. No protected action was taken. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree binary checkpoint-state secret boundary evidence

This bounded public-API safety iteration refreshed Transformers main at commit d09f53a801f45ad73ec3510e17972024234bc0fd. Its 1,576-line navigation, 500-line Modular Transformers guide, and 4,460-line Trainer source retain SHA-256 values f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a, a52af1ee1754e6e658535f071a727fac31a057a3a848c28c18873a5ff371dd96, and 6ea719d4f225c6fcbccfcd2137cdfb304426c93c4dd65983ae4186f6e8fb3d8d. The official Home and Modular Transformers routes returned HTTP 200. That Trainer lifecycle is structural context; VoiceHub's exact-resume binary-state boundary controls this slice.

The observable contract requires optimizer, scheduler, random-generator, gradient-scaler, callback, sampler, and strategy state mappings to reject nested credential-shaped fields at their final binary write boundary. A rejected write must leave neither a final nor an incomplete checkpoint, and the error must expose only the owning state and key path. Loaded binary state must be checked before any model or runtime state is restored. Safe metric fields such as token_count must still round-trip through exact resume. The post-deserialization check is explicitly not a safe-unpickling claim: Python pickle input must come from a trusted VoiceHub checkpoint with intact manifest integrity.

The first command used an unavailable python executable and ran no tests; it is not evidence. The corrected pre-implementation command is also not a pass: it reported three failures alongside two passing results because unsafe callback and optimizer state completed a checkpoint and loaded optimizer state was reached only after model restoration. The shared final-payload helper now validates every Trainer-owned binary state mapping immediately before torch.save. The load path deserializes and validates all non-model state, including TrainerState, before applying the model or any runtime state. No task, provider, optimizer, callback, or strategy allowlist was added.

The corrected focused contract passed three tests and two subtests with 36 deselected. The complete training-runtime file passed 39 tests and seven subtests. The training, checkpoint, callback, and integration slice passed 306 tests and 207 subtests with four warnings. The registry, public-API, optimization, scaffold, packaging, distribution-compliance, and release- policy slice passed 319 tests and 1,454 subtests with four warnings.

The first documentation run is not a pass: its new warning pushed the concise training guide to 253 lines and one contract failed. The warning was condensed without weakening it; the corrected documentation file passed 64 tests and 1,476 subtests with the guide restored to its 250-line limit. The complete Python 3.12.12 suite then passed 2,533 executed tests and 4,043 subtests with 35 warnings in 113.29 seconds. Its 15 skips remain unpassed: three complete- dependency default-runtime checks, three Triton and two CUDA-extension checks, five opt-in ESPNet/NeMo/SenseVoice/SpeechBrain/TEN-VAD asset or oracle checks, and two inaccessible WeNet paths.

All 68 model pages and 59 model notebooks remained current, and the five- record release-alignment check passed. The strict eleven-language build completed in 32.93 seconds. The ten-route DOM validator retained all eight ordered navigation roots. The unchanged documentation shell retains its preceding Playwright/Axe evidence; no fresh visual-parity claim is made for this runtime-safety and prose-only slice.

A fresh physical distribution check passed wheel, sdist, and editable installs with 68 models, 81 provenance manifests, 193 compliance files, required package data, zero runtime dependency violations, and no eager PyTorch import. It produced a 57,193,103-byte wheel and a 55,455,792-byte sdist. The selected- file pre-commit invocation passed every applicable source hook. Its Markdown hook found no matching files and is not counted as a pass.

Refreshed inventories remain 68 models (34 TTS, 23 ASR, and 11 VAD), 102 aliases, zero invalid display names, 68 provider pages with no missing or orphaned page, six public optimization passes, 408 model/pass pairs, eight top-level navigation roots, ten representative routes, eight contribution steps, and 261 public exports. The public root import still avoids PyTorch.

This closes credential handling for exact-resume binary state mappings. The broader object-by-object public API audit remains open. Exact-current-worktree Linux, Windows, Python 3.10, Python 3.11, default-runtime, tagged-workflow, and publisher evidence; publication approval; five CUDA/Triton gates; and the seven opt-in or inaccessible asset paths remain pending. No protected action was taken. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree TrainingArguments secret boundary evidence

This bounded public-API safety iteration refreshed Transformers main at commit d09f53a801f45ad73ec3510e17972024234bc0fd. Its 1,576-line navigation and 500-line Modular Transformers sources retain SHA-256 values f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a and a52af1ee1754e6e658535f071a727fac31a057a3a848c28c18873a5ff371dd96. The current 2,906-line upstream training_args.py has SHA-256 cb55368c4d4b80633a7d289a790c5e4ab9f7ad03b527a969be7b0b0c2c432313. The official Home and Modular Transformers routes returned HTTP 200. The upstream argument object and its dictionary/JSON lifecycle are the structural reference; VoiceHub's credential-safe portable artifact boundary controls this slice.

The observable contract requires nested credential-shaped subclass fields to fail during normal construction and before an untrusted JSON payload can construct a subclass. Inherited dictionary, JSON-string, and file writers must revalidate the complete current dataclass payload even when construction validation was bypassed. Trainer.save_model() must reject the same unsafe arguments before creating its destination or writing model state. Errors may report only the owning object and key path. Safe metadata such as dataset_id and token_count must continue to round-trip.

The pre-implementation focused command is not a pass: three tests failed while the safe round-trip case passed because unsafe construction, untrusted loading, inherited serialization, and trainer export all accepted credential fields. The shared arguments normalizer now checks its complete dataclass payload at construction and every inherited persistence boundary. The untrusted loader checks the payload before subclass construction, and trainer export performs the same preflight before importing the training backend or creating the destination. No reporting backend, task, provider, or subclass allowlist was added.

The corrected focused contract passed four tests and two subtests with 24 deselected. The complete training, checkpoint, W&B, adapter, and optimization slice passed 180 tests and 212 subtests. The registry, public-API, universal- optimization, scaffold, packaging-compliance, and release-policy slice passed 151 tests and 621 subtests.

The complete Python 3.12.12 suite passed 2,537 executed tests and 4,045 subtests with 35 warnings in 109.04 seconds. Its 15 skips remain unpassed: three complete-dependency default-runtime checks, three Triton and two CUDA- extension checks, five opt-in ESPNet/NeMo/SenseVoice/SpeechBrain/TEN-VAD asset or oracle checks, and two inaccessible WeNet paths.

Documentation passed 64 tests and 1,476 subtests. All 68 model pages and 59 model notebooks remained current, the five-record release-alignment check passed, and the strict eleven-language build completed in 35.53 seconds. The ten-route DOM validator retained all eight ordered navigation roots. The unchanged documentation shell retains its preceding Playwright/Axe evidence; no fresh visual-parity claim is made for this API-safety and prose-only slice.

A fresh physical distribution check passed wheel, sdist, and editable installs with 68 models, 81 provenance manifests, 193 compliance files, required package data, zero runtime dependency violations, and no eager PyTorch import. It produced a 57,193,213-byte wheel and a 55,456,837-byte sdist. The first selected-file pre-commit invocation is not a pass because YAPF rewrote the new test formatting. The corrected focused contract and second invocation passed every applicable source hook. The Markdown hook found no matching files and is not counted as a pass.

Refreshed inventories remain 68 models (34 TTS, 23 ASR, and 11 VAD), 102 aliases, zero invalid display names, 68 provider pages with no missing or orphaned page, six public optimization passes, 408 model/pass pairs, eight top-level navigation roots, ten representative routes, eight contribution steps, and 261 public exports. The public root import still avoids PyTorch.

This closes credential handling for portable TrainingArguments artifacts. The broader object-by-object public API audit remains open. Exact-current- worktree Linux, Windows, Python 3.10, Python 3.11, default-runtime, tagged- workflow, and publisher evidence; publication approval; five CUDA/Triton gates; and the seven opt-in or inaccessible asset paths remain pending. No protected action was taken. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree optimization-manifest secret boundary evidence

This bounded public-API safety iteration refreshed Transformers main at commit d09f53a801f45ad73ec3510e17972024234bc0fd. Its 65-line Home, 1,576-line navigation, 500-line Modular Transformers, 30-line Trainer, and 2,906-line TrainingArguments sources retain SHA-256 values e24e0eddadccef8b15e3f25c55acb0df0f65de21a4b5bf709d119d22d0eb555b, f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a, a52af1ee1754e6e658535f071a727fac31a057a3a848c28c18873a5ff371dd96, e7c5368c1223c2b195321468fcd1ac64c5cda52a1e44d4adf854a6473b6c9ee0, and cb55368c4d4b80633a7d289a790c5e4ab9f7ad03b527a969be7b0b0c2c432313. The official Home and Modular Transformers routes returned HTTP 200. Upstream is structural context; VoiceHub's safe optimization-artifact boundary controls this slice.

The observable contract requires credential-shaped pass configuration to fail before validation or model mutation. Credential-shaped application metadata must fail before publishing the transformed graph and trigger deterministic rollback. Because runtime status can evolve after application, the final manifest call must revalidate it. Errors may report the owning manifest and key path, but not a credential value. Descriptive fields such as token_count must continue to round-trip.

A direct pre-implementation probe confirmed the gap: both nested api_key configuration and later nested auth_token runtime status reached the public optimization manifest, and the pass applied once. The focused regression is not a pass: it failed one test with 29 deselected because unsafe configuration did not raise. The shared canonical strict-JSON tree now applies the repository credential detector before normalization. That one boundary covers pass declarations, resolver-owned declaration snapshots, application metadata, mutable runtime status, generic and TTS optimization manifests, and every built-in or extension pass without a provider, pass, or architecture allowlist.

The corrected focused contract passed three tests with 27 deselected. The native, universal TTS, compiler, accelerator, codec, diffusion-cache, diffusion-sampling, and TTS optimization slice passed 190 tests and 684 subtests with one warning. The remaining codec, diffusion, serving, native- codec, compile-target, report, VITS, Trainer, and training-runtime slice passed 204 tests and 218 subtests with eight warnings. The registry, public API, speech-task, checkpoint, distribution-compliance, and release-policy slice passed 147 tests and 508 subtests. An earlier registry command named a nonexistent package-metadata test file, collected no tests, and is not passing evidence.

The complete Python 3.12.12 suite passed 2,538 executed tests and 4,045 subtests with 35 warnings in 102.32 seconds. Its 15 skips remain unpassed: three complete-dependency default-runtime checks, three Triton and two CUDA- extension checks, five opt-in ESPNet/NeMo/SenseVoice/SpeechBrain/TEN-VAD asset or oracle checks, and two inaccessible WeNet paths.

Documentation passed 64 tests and 1,476 subtests. The strict eleven-language build completed in 32.60 seconds, and the ten-route DOM validator retained all eight ordered navigation roots. The unchanged representative shell retains its preceding Playwright/Axe evidence; no fresh visual-parity claim is made for this API-safety and reference-prose slice. A fresh physical distribution check passed wheel, sdist, and editable installs with 68 models, 81 provenance manifests, 193 compliance files, required package data, zero runtime dependency violations, and no eager PyTorch import.

All 68 generated model pages and 59 generated model notebooks remained current. The five-record release-alignment check passed. Refreshed inventories remain 68 models (34 TTS, 23 ASR, and 11 VAD), 102 aliases, zero invalid display names, 68 provider pages with no missing or orphaned page, six public optimization passes, 408 model/pass pairs, eight top-level navigation roots, ten representative routes, eight contribution steps, and 261 public exports. The public root import still avoids PyTorch.

The first selected-file pre-commit invocation is not a pass because YAPF rewrote the new test formatting. The corrected focused contract and second selected-file invocation passed every applicable source hook; the Markdown hook found no matching files and is not counted as a pass. This closes credential handling for public optimization manifests. The broader object-by- object public API audit remains open. Exact-current-worktree Linux, Windows, Python 3.10, Python 3.11, default-runtime, tagged-workflow, and publisher evidence; publication approval; five CUDA/Triton gates; and the seven opt-in or inaccessible asset paths remain pending. No protected action was taken. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree training-recipe manifest secret boundary evidence

This bounded public-artifact safety iteration refreshed Transformers main at commit d09f53a801f45ad73ec3510e17972024234bc0fd. Its 65-line Home, 1,576-line navigation, 500-line Modular Transformers, 30-line Trainer, and 2,906-line TrainingArguments sources retain SHA-256 values e24e0eddadccef8b15e3f25c55acb0df0f65de21a4b5bf709d119d22d0eb555b, f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a, a52af1ee1754e6e658535f071a727fac31a057a3a848c28c18873a5ff371dd96, e7c5368c1223c2b195321468fcd1ac64c5cda52a1e44d4adf854a6473b6c9ee0, and cb55368c4d4b80633a7d289a790c5e4ab9f7ad03b527a969be7b0b0c2c432313. The official Home and Modular Transformers routes returned HTTP 200. Upstream is structural context; VoiceHub's portable training-recipe boundary controls this slice.

The observable contract requires the exact mapping returned by a training- adapter subclass to reject nested credential-shaped fields before importing the training backend, reading model state, creating the artifact destination, or starting a native export. The final mapping must be checked again before training_recipe.json is written. Errors may report the owning artifact and field path, but not a credential value, while safe descriptive fields such as dataset_id and token_count must continue to round-trip.

A direct pre-implementation probe confirmed the gap: a nested api_key from a custom adapter was persisted in training_recipe.json, and the destination was created. The focused regression is not a pass: it failed one test with 39 deselected because no error was raised. Trainer.save_model() now validates a mapping snapshot before backend import or artifact mutation and reuses that same validated payload for output. It rechecks the payload after adding the bounded native-export path and immediately before the deterministic JSON write. No task, provider, adapter, or recipe allowlist was added.

The corrected focused contract passed one test with 39 deselected. The training-runtime, Trainer, adapter, checkpoint, and native-optimization slice passed 132 tests and 575 subtests. The documentation, registry, speech-task, universal-optimization, distribution-compliance, and release-readiness slice passed 153 tests and 2,058 subtests. The complete Python 3.12.12 suite passed 2,539 executed tests and 4,045 subtests with 35 warnings in 104.51 seconds. Its 15 skips remain unpassed: three complete-dependency default-runtime checks, three Triton and two CUDA-extension checks, five opt-in ESPNet/NeMo/SenseVoice/ SpeechBrain/TEN-VAD asset or oracle checks, and two inaccessible WeNet paths.

Documentation passed 64 tests and 1,476 subtests. All 68 generated model pages and 59 generated model notebooks remained current, and the five-record release alignment check passed. The strict eleven-language build completed in 32.86 seconds, and the ten-route DOM validator retained all eight ordered navigation roots. The unchanged representative shell retains its preceding Playwright/ Axe evidence; no fresh visual-parity claim is made for this persistence and reference-prose slice.

A fresh physical distribution check passed wheel, sdist, and editable installs with 68 models, 81 provenance manifests, 193 compliance files, required package data, zero runtime dependency violations, and no eager PyTorch import. It produced a 57,193,356-byte wheel and a 55,458,161-byte sdist. The first selected-file pre-commit invocation is not a pass because YAPF rewrote the new test formatting. The formatted focused contract and second invocation passed every applicable source hook; the Markdown hook found no matching files and is not counted as a pass.

Refreshed inventories remain 68 models (34 TTS, 23 ASR, and 11 VAD), 102 aliases, zero invalid display names, 68 provider pages with no missing or orphaned page, six public optimization passes, 408 model/pass pairs, eight top-level navigation roots, ten representative routes, eight contribution steps, and 261 public exports. The public root import still avoids PyTorch.

This closes credential handling for adapter-supplied training-recipe manifests. The broader object-by-object public API audit remains open. Exact- current-worktree Linux, Windows, Python 3.10, Python 3.11, default-runtime, tagged-workflow, and publisher evidence; publication approval; five CUDA/ Triton gates; and the seven opt-in or inaccessible asset paths remain pending. No protected action was taken. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree model-artifact state secret boundary evidence

This bounded public-artifact safety iteration refreshed Transformers main at commit d09f53a801f45ad73ec3510e17972024234bc0fd. Its 65-line Home, 1,576-line navigation, 500-line Modular Transformers, 30-line Trainer, and 2,906-line TrainingArguments sources retain SHA-256 values e24e0eddadccef8b15e3f25c55acb0df0f65de21a4b5bf709d119d22d0eb555b, f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a, a52af1ee1754e6e658535f071a727fac31a057a3a848c28c18873a5ff371dd96, e7c5368c1223c2b195321468fcd1ac64c5cda52a1e44d4adf854a6473b6c9ee0, and cb55368c4d4b80633a7d289a790c5e4ab9f7ad03b527a969be7b0b0c2c432313. The official Home and Modular Transformers routes returned current content. Upstream is structural context; VoiceHub's safe binary model-artifact boundary controls this slice.

The observable contract requires the exact state returned by a model or portable optimization result to reject nested credential-shaped fields before artifact destination creation or a model-owned save. Because a model-owned save can mutate a previously returned mapping, the exact state must be checked again immediately before model_state.pt is written. Errors may report the owning artifact and field path but not a credential value. Tensors and safe descriptive metadata such as dataset_id and token_count must continue to round-trip.

A direct pre-implementation probe confirmed the gap: a nested api_key was persisted in model_state.pt, and the destination was created. The focused regression is not a pass: it failed one test with 40 deselected because no error was raised. Trainer.save_model() now applies one shared model-artifact state validator before filesystem mutation and again at the final binary write boundary. This covers ordinary model state and the universal portable- optimization state path without a task, provider, model, or optimization-pass allowlist.

The corrected focused contract passed one test with 40 deselected. The complete training-runtime file passed 41 tests and seven subtests. The training, checkpoint, optimization, compilation, and reporting slice passed 241 tests and 673 subtests with one platform warning. The registry, public API, inference, documentation, scaffold, packaging, distribution-compliance, and release-policy slice passed 264 tests and 2,089 subtests with one platform warning. Selected pre-commit hooks passed in one invocation; the Markdown hook found no matching files and is not counted as a pass.

The complete Python 3.12.12 suite passed 2,540 executed tests and 4,045 subtests with 35 warnings in 108.73 seconds. Its 15 skips remain unpassed: three complete-dependency default-runtime checks, three Triton and two CUDA- extension checks, five opt-in ESPNet/NeMo/SenseVoice/SpeechBrain/TEN-VAD asset or oracle checks, and two inaccessible WeNet paths.

Documentation passed 64 tests and 1,476 subtests. All 68 generated model pages and 59 generated model notebooks remained current, and the five-record release alignment check passed. The strict eleven-language build completed in 35.02 seconds, and the ten-route DOM validator retained all eight ordered navigation roots. The unchanged representative shell retains its preceding Playwright/ Axe evidence; this persistence and reference-prose slice makes no new visual- parity claim.

A fresh physical distribution check passed wheel, sdist, and editable installs with 68 models, 81 provenance manifests, 193 compliance files, all required package data, zero runtime-dependency violations, and no eager PyTorch import. It produced a 57,193,419-byte wheel and a 55,458,392-byte sdist. Refreshed inventories remain 68 models (34 TTS, 23 ASR, and 11 VAD), 102 aliases, zero invalid display names, 68 provider pages with no missing or orphaned page, six public optimization passes, 408 model/pass pairs, eight top-level navigation roots, ten representative routes, and eight contribution steps.

This closes credential handling for Trainer-owned binary model artifacts. The broader object-by-object public API audit remains open. Exact-current-worktree Linux, Windows, Python 3.10, Python 3.11, default-runtime, tagged-workflow, and publisher evidence; publication approval; five CUDA/Triton gates; and the seven opt-in or inaccessible asset paths remain pending. No protected action was taken. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree model-state load boundary evidence

This bounded public-artifact safety iteration refreshed Transformers main at commit d09f53a801f45ad73ec3510e17972024234bc0fd. Its 65-line Home, 1,576-line navigation, 500-line Modular Transformers, 30-line Trainer, and 2,906-line TrainingArguments sources retain SHA-256 values e24e0eddadccef8b15e3f25c55acb0df0f65de21a4b5bf709d119d22d0eb555b, f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a, a52af1ee1754e6e658535f071a727fac31a057a3a848c28c18873a5ff371dd96, e7c5368c1223c2b195321468fcd1ac64c5cda52a1e44d4adf854a6473b6c9ee0, and cb55368c4d4b80633a7d289a790c5e4ab9f7ad03b527a969be7b0b0c2c432313. The official Home, Modular Transformers, and Trainer routes returned current content. Upstream is structural context; VoiceHub's credential-safe exact- resume boundary controls this slice.

The observable contract requires credential-shaped values in deserialized model_state.pt to fail before load_state_dict() can mutate the model. The error may report only the owning artifact and key path, while ordinary model state must still restore exactly once. This is a post-deserialization credential check, not a safe-unpickling claim; Python pickle input must still come from a trusted VoiceHub checkpoint with intact manifest integrity.

The pre-implementation focused regression is not a pass. It reached the tracking model's load_state_dict() and raised a PyTorch missing/unexpected-key error instead of the credential-policy error. The shared model-artifact state validator now runs immediately after deserialization and before model mutation, without a task, provider, model, or optimization-pass allowlist. The corrected focused regression passed, and the complete training-runtime file passed 42 tests and seven subtests.

The proportional training and checkpoint slice passed 199 tests and 201 subtests. The optimization, registry, public-API, and speech-task slice passed 242 tests and 1,239 subtests with one platform warning. The documentation, release-policy, distribution-compliance, scaffold, pipeline, and automatic- configuration slice passed 120 tests and 1,519 subtests. The complete Python 3.12.12 suite passed 2,541 executed tests and 4,045 subtests with 35 warnings in 108.46 seconds. Its 15 skips remain unpassed: three complete-dependency default-runtime checks, three Triton and two CUDA-extension checks, five opt-in ESPNet/NeMo/SenseVoice/SpeechBrain/TEN-VAD asset or oracle checks, and two inaccessible WeNet paths.

The release-alignment check passed with five benchmark records and all 68 documented providers. All 68 generated model pages and 59 generated model notebooks remained current. Refreshed inventories remain 68 models (34 TTS, 23 ASR, and 11 VAD), 102 aliases, zero missing or orphaned provider pages, six public optimization passes, 408 model/pass pairs, eight top-level navigation roots, and 261 public exports. The public root import still avoids PyTorch.

The strict eleven-language build completed in 32.84 seconds, and the ten-route DOM validator retained all eight ordered navigation roots. The unchanged representative shell retains its preceding Playwright/Axe evidence; this Trainer runtime and reference-prose slice makes no new visual-parity claim. A fresh isolated distribution check passed wheel, sdist, and editable installs with 68 models, 81 provenance manifests, 193 compliance files, all required package data, zero runtime-dependency violations, and no eager PyTorch import. It produced a 57,193,429-byte wheel and a 55,458,434-byte sdist.

The first selected-file pre-commit invocation is not a pass because YAPF rewrote the new test. The formatted focused contract and second invocation then passed every applicable source hook; the Markdown hook found no matching files and is not counted as a pass. A direct release check against the repository dist/ directory also failed because that directory contains its tracked .gitignore; it is not distribution evidence. The isolated distribution checker above used a clean temporary artifact directory and passed.

Committed HEAD 8ea5e941fcbcc0b93e5a4dd180b7a4c15c235930 has successful remote Continuous Integration, Documentation, and Package CI runs, including the supported Python and operating-system matrix, but those runs do not contain the current dirty worktree and are not exact-candidate evidence. The broader object-by-object public API audit, exact-current-worktree remote platform and tagged-workflow evidence, publisher configuration, publication approval, five CUDA/Triton gates, and seven opt-in or inaccessible asset paths remain open. No protected action was taken. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree public portable model-state load boundary evidence

This bounded public-lifecycle iteration refreshed Transformers main at commit d09f53a801f45ad73ec3510e17972024234bc0fd. Its 65-line Home, 1,576-line navigation, 500-line Modular Transformers, 30-line Trainer, and 2,906-line TrainingArguments sources retain SHA-256 values e24e0eddadccef8b15e3f25c55acb0df0f65de21a4b5bf709d119d22d0eb555b, f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a, a52af1ee1754e6e658535f071a727fac31a057a3a848c28c18873a5ff371dd96, e7c5368c1223c2b195321468fcd1ac64c5cda52a1e44d4adf854a6473b6c9ee0, and cb55368c4d4b80633a7d289a790c5e4ab9f7ad03b527a969be7b0b0c2c432313. The official Home, Modular Transformers, and Trainer routes returned current content. Upstream is structural context; VoiceHub's shared public artifact load boundary controls this slice.

The observable contract requires every public TTS, ASR, and VAD pretrained base to reject nested credential-shaped fields after deserializing portable model_state.pt and before either a training adapter or runtime can mutate a fresh model. Errors may expose the owning artifact and field path but not the credential value. A rejected state must remain pending for retry, while safe ordinary state must load exactly once and restore inference mode. This is a post-deserialization credential boundary, not a safe-unpickling claim; Python pickle artifacts must still come from a trusted source.

The pre-implementation focused regression is not a pass: both public base families accepted the credential mapping and failed two tests because no policy error was raised. An earlier shell invocation also failed before test collection because python was not on PATH; it is not evidence. One shared task-neutral validator now runs immediately after torch.load() and before the adapter/runtime branch. The final focused contract passed two tests. Its TTS case proves the adapter-marked path cannot call adapter setup(), and its audio case proves ordinary state cannot call runtime load_state_dict(). Both cases prove redacted errors, retry preservation, and one successful safe restore. The complete inference-lifecycle and speech-core files passed 76 tests and 41 subtests.

The proportional training and checkpoint slice passed 168 tests and 189 subtests. The optimization, registry, public-API, automatic-model, pipeline, and speech-task slice passed 333 tests and 1,275 subtests with four warnings. The documentation, release-policy, distribution-compliance, packaging, and model-scaffold slice passed 107 tests and 1,587 subtests. The complete Python 3.12.12 suite passed 2,543 executed tests and 4,045 subtests with 35 warnings in 107.12 seconds. Its 15 skips remain unpassed: three complete-dependency default-runtime checks, three Triton and two CUDA-extension checks, five opt-in ESPNet/NeMo/SenseVoice/SpeechBrain/TEN-VAD asset or oracle checks, and two inaccessible WeNet paths.

The release-alignment check passed with five benchmark records and all 68 documented providers. All 68 generated model pages and 59 generated model notebooks remained current. Refreshed inventories remain 68 models (34 TTS, 23 ASR, and 11 VAD), 102 aliases, zero missing or orphaned provider pages, six public optimization passes, 408 model/pass pairs, eight top-level navigation roots, and 261 public exports. The public root import still avoids PyTorch.

The strict eleven-language build completed in 34.21 seconds, and the ten-route DOM validator retained all eight ordered navigation roots. Its first command used an unsupported option and did not run; only the corrected positional invocation is counted. The unchanged representative shell retains its preceding Playwright/Axe evidence; this lifecycle and reference-prose slice makes no new visual-parity claim. A fresh isolated distribution check passed wheel, sdist, and editable installs with 68 models, 81 provenance manifests, 193 compliance files, all required package data, zero runtime-dependency violations, and no eager PyTorch import. It produced a 57,193,552-byte wheel and a 55,458,754-byte sdist. Every applicable selected-file pre-commit hook passed; the Markdown hook found no matching files and is not counted.

Committed HEAD 8ea5e941fcbcc0b93e5a4dd180b7a4c15c235930 has successful remote Continuous Integration, Documentation, and Package CI runs, including the supported Python and operating-system matrix, but those runs do not contain the current dirty worktree and are not exact-candidate evidence. The broader object-by-object public API audit, exact-current-worktree remote platform and tagged-workflow evidence, publisher configuration, publication approval, five CUDA/Triton gates, and seven opt-in or inaccessible asset paths remain open. No protected action was taken. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree speech dataset manifest safety evidence

This bounded public-artifact iteration refreshed Transformers main at commit d09f53a801f45ad73ec3510e17972024234bc0fd. Revision-pinned revalidation now shows its 65-line Home, 1,576-line navigation, 500-line Modular Transformers, 30-line Trainer, and 2,906-line TrainingArguments sources with SHA-256 values e24e0eddadccef8b15e3f25c55acb0df0f65de21a4b5bf709d119d22d0eb555b, f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a, a52af1ee1754e6e658535f071a727fac31a057a3a848c28c18873a5ff371dd96, e7c5368c1223c2b195321468fcd1ac64c5cda52a1e44d4adf854a6473b6c9ee0, and cb55368c4d4b80633a7d289a790c5e4ab9f7ad03b527a969be7b0b0c2c432313. The official Home, Modular Transformers, and Trainer routes returned current content. The earlier 54-line Trainer response recorded here was not reproducible against the pinned revision and is excluded rather than carried forward. Upstream is structural context; VoiceHub's portable speech-dataset boundary controls this slice.

The observable contract requires both public TTS and ASR manifest readers and JSON Lines writers to reject nested credential-shaped fields. Reads must fail before dataset construction. Writes must validate and serialize every record before creating a parent directory, creating a file, or truncating an existing destination. Errors may expose the task, record index, and field path but not a credential value. Ordinary descriptive metadata such as token_count must continue to round-trip.

The corrected four-test pre-implementation regression is not a pass: all four selected writer/reader contracts accepted credentials, with two additional writer subtest failures because no policy error was raised. The first command used an incorrect ASR test class and collected no tests; it is also excluded. Both task modules now invoke the shared credential detector after portable path normalization and after parsing untrusted JSON, JSON Lines, CSV, or TSV records. TTS export now precomputes every line before filesystem mutation, matching the existing ASR preflight structure without introducing a task- provider allowlist.

The final focused contract passed five tests. It proves credential rejection for both writers and both readers, redacted errors, preservation of an existing destination, absence of a rejected destination's parent directory, TTS preservation on an ordinary JSON serialization failure, and safe token_count round-trips. The complete TTS and ASR dataset files passed 79 tests and 116 subtests.

The proportional training and dataset slice passed 244 tests and 317 subtests. The registry, optimization, public-API, automatic-model, pipeline, inference, and speech-task slice passed 227 tests and 1,117 subtests. The documentation, release-policy, distribution-compliance, packaging, and model-scaffold slice passed 107 tests and 1,587 subtests. The complete Python 3.12.12 suite passed 2,548 executed tests and 4,045 subtests with 35 warnings in 110.21 seconds. Its 15 skips remain unpassed: three complete-dependency default-runtime checks, three Triton and two CUDA-extension checks, five opt-in ESPNet/NeMo/SenseVoice/SpeechBrain/TEN-VAD asset or oracle checks, and two inaccessible WeNet paths.

The release-alignment check passed with five benchmark records and all 68 documented providers. All 68 generated model pages and 59 generated model notebooks remained current. Refreshed inventories remain 68 models (34 TTS, 23 ASR, and 11 VAD), 102 aliases, zero invalid display names, 68 provider pages with no missing or orphaned page, six public optimization passes, 408 model/pass pairs, eight top-level navigation roots, and 261 public exports. The public root import still avoids PyTorch.

The strict eleven-language build completed in 35.42 seconds, and the ten-route DOM validator retained all eight ordered navigation roots. The unchanged representative shell retains its preceding Playwright/Axe evidence; this data contract and reference-prose slice makes no new visual-parity claim. A fresh isolated distribution check passed wheel, sdist, and editable installs with 68 models, 81 provenance manifests, 193 compliance files, all required package data, zero runtime-dependency violations, and no eager PyTorch import. It produced a 57,193,721-byte wheel and a 55,459,521-byte sdist. Every applicable selected-file pre-commit hook passed; the Markdown hook found no matching files and is not counted.

Committed HEAD 8ea5e941fcbcc0b93e5a4dd180b7a4c15c235930 retains successful remote Continuous Integration, Documentation, and Package CI runs, including the supported Python and operating-system matrix, but those runs do not contain the current dirty worktree and are not exact-candidate evidence. The broader object-by-object public API audit, exact-current-worktree remote platform and tagged-workflow evidence, publisher configuration, publication approval, five CUDA/Triton gates, and seven opt-in or inaccessible asset paths remain open. No protected action was taken. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree shared JSON artifact atomicity evidence

This bounded data-safety iteration refreshed Transformers main at commit d09f53a801f45ad73ec3510e17972024234bc0fd. Its revision-pinned 65-line Home, 1,576-line navigation, 500-line Modular Transformers, 30-line Trainer, and 2,906-line TrainingArguments sources have SHA-256 values e24e0eddadccef8b15e3f25c55acb0df0f65de21a4b5bf709d119d22d0eb555b, f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a, a52af1ee1754e6e658535f071a727fac31a057a3a848c28c18873a5ff371dd96, e7c5368c1223c2b195321468fcd1ac64c5cda52a1e44d4adf854a6473b6c9ee0, and cb55368c4d4b80633a7d289a790c5e4ab9f7ad03b527a969be7b0b0c2c432313. The official Home, Modular Transformers, navigation, and Trainer routes returned current content. Upstream is structural context; VoiceHub's shared artifact writer controls this slice.

The observable contract requires shared model and Trainer JSON writers to encode finite deterministic JSON before creating a parent directory or touching a destination. A successful write must flush a temporary sibling and atomically replace the destination. Serialization or replacement failure must preserve an existing artifact and remove temporary files. Safe metadata such as token_count must continue to round-trip with the existing trailing-newline format. This is a per-document guarantee rather than a claim that a backend's multi-file native export is transactional without its staging contract.

The pre-implementation focused command exited one: nine subcases demonstrated parent creation, partial truncation, permissive non-finite output, or absence of an atomic replacement boundary. Its three reported top-level passes and three other subtest passes do not make the failed run passing evidence. The shared writer now precomputes strict JSON, writes and fsyncs a temporary sibling, uses os.replace(), and removes the temporary path on every failure. Trainer's JSON helper delegates to the same implementation, so configuration, processor, native-export helper, Trainer state, training arguments, and checkpoint manifest call sites no longer maintain divergent document-write behavior.

The corrected focused contract passed three tests and 12 subtests. The first selected-file pre-commit invocation is excluded because YAPF reformatted the new code; after formatting, the focused contract passed again and every applicable source hook passed. The Markdown hook found no applicable files and is not counted. A first dependency-free Python 3.10/3.11 probe is also excluded because its harness checked for the literal characters \n rather than a newline. The corrected exact-slice probe passed on Python 3.10.19 and 3.11.15.

The public API, configuration, processor, pipeline, training, checkpoint, and inference slice passed 207 tests and 76 subtests. The registry, optimization, documentation-policy, packaging-policy, distribution-policy, and release slice passed 193 tests and 2,616 subtests. The complete Python 3.12.12 suite passed 2,551 executed tests and 4,057 subtests with 35 warnings in 107.63 seconds. Its 15 skips remain unpassed: three complete-dependency default-runtime checks, three Triton and two CUDA-extension checks, five opt-in ESPNet/NeMo/SenseVoice/SpeechBrain/TEN-VAD asset or oracle checks, and two inaccessible WeNet paths.

Documentation and final release-policy validation passed 87 tests and 1,552 subtests. The strict eleven-language build completed successfully, and the ten-route DOM validator retained all eight ordered navigation roots. The unchanged representative shell retains its preceding Playwright/Axe evidence; this data-safety and reference-prose slice makes no new visual-parity claim. A fresh isolated distribution check passed wheel, sdist, and editable installs with 68 models, 81 provenance manifests, 193 compliance files, all required package data, zero runtime-dependency violations, and no eager PyTorch import. It produced a 57,193,870-byte wheel and a 55,459,554-byte sdist.

The five-record release-alignment check passed. All 68 generated model pages and 59 generated model notebooks remained current. Refreshed inventories remain 68 models (34 TTS, 23 ASR, and 11 VAD), 102 aliases, zero invalid display names, 68 provider pages with no missing or orphaned page, six public optimization passes, 408 model/pass pairs, eight top-level navigation roots, eight contribution steps, and 261 public exports. The public root import still avoids PyTorch.

Committed HEAD 8ea5e941fcbcc0b93e5a4dd180b7a4c15c235930 retains successful remote Continuous Integration, Documentation, and Package CI runs, but those runs do not contain the current dirty worktree and are not exact-candidate evidence. The broader object-by-object public API audit, exact-current-worktree complete Python 3.10/3.11, Linux, Windows, default-runtime, tagged-workflow, and publisher evidence, publication approval, five CUDA/Triton gates, and seven opt-in or inaccessible asset paths remain open. No protected action was taken. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree shared JSON artifact read-safety evidence

This bounded data-safety iteration refreshed Transformers main at commit d09f53a801f45ad73ec3510e17972024234bc0fd. Its revision-pinned 65-line Home, 1,576-line navigation, 500-line Modular Transformers, 30-line Trainer, and 2,906-line TrainingArguments sources have SHA-256 values e24e0eddadccef8b15e3f25c55acb0df0f65de21a4b5bf709d119d22d0eb555b, f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a, a52af1ee1754e6e658535f071a727fac31a057a3a848c28c18873a5ff371dd96, e7c5368c1223c2b195321468fcd1ac64c5cda52a1e44d4adf854a6473b6c9ee0, and cb55368c4d4b80633a7d289a790c5e4ab9f7ad03b527a969be7b0b0c2c432313. The official Home, navigation, and Trainer routes returned current content. The Modular Transformers route returned HTTP 429 and is unavailable evidence, not a pass; its revision-pinned raw source was still fetched and matched the recorded fingerprint. Upstream is structural context, while VoiceHub's shared artifact loader controls this slice.

The observable contract requires duplicate object keys to fail at every nesting level and requires NaN, positive or negative Infinity, and exponent overflow to fail before configuration construction or automatic model dispatch. Errors must identify the source and duplicate key or numeric path without printing discarded values. Ordinary finite JSON, including nested token_count metadata, must keep round-tripping.

The pre-implementation focused command exited one with seven failures; it accepted both duplicate-key documents, all four non-finite forms, and an ambiguous duplicate model_type before AutoConfig dispatch. Its three passes are not transferred out of the failed run. The first post-implementation command is also excluded: rejection behavior worked, but three message-case assertions failed. After the diagnostic wording was corrected, the focused contract passed seven tests and 18 subtests. It covers top-level and nested duplicates, redacted discarded values, all non-finite forms, numeric-path reporting for exponent overflow, pre-dispatch AutoConfig rejection, and a safe nested round-trip.

Dependent public API, automatic-configuration, inference, pipeline, speech- core, and native-checkpoint contracts passed 137 tests and 53 subtests. The registry and optimization selection passed 234 tests and 1,339 subtests. The complete Python 3.12.12 suite passed 2,555 executed tests and 4,063 subtests with 35 warnings in 106.93 seconds. Its 15 skips remain unpassed: three complete-dependency default-runtime checks, three Triton and two CUDA-extension checks, five opt-in ESPNet/NeMo/SenseVoice/SpeechBrain/TEN-VAD asset or oracle checks, and two inaccessible WeNet paths. Focused dependency-free probes passed on Python 3.10.19 and 3.11.15; complete exact-current-worktree execution on those interpreters remains pending.

The five-record release-alignment check passed. All 68 generated model pages and 59 generated model notebooks remain current. Refreshed inventories remain 68 models (34 TTS, 23 ASR, and 11 VAD), 102 aliases, zero invalid display names, 68 provider pages with no missing or orphaned page, six public optimization passes, 408 model/pass pairs, eight top-level navigation roots, eight contribution steps, and 261 public exports. The public root import still avoids PyTorch.

Release, distribution-policy, and documentation-policy validation passed 77 tests and 1,476 subtests. A fresh isolated distribution check passed wheel, sdist, and editable installs with 68 models, 81 provenance manifests, 193 compliance files, all required package data, zero runtime-dependency violations, and no eager PyTorch import. It produced a 57,194,351-byte wheel and a 55,460,094-byte sdist. Every applicable selected source/test pre-commit hook passed before this evidence update.

After the evidence update, the focused reader plus release, distribution, and documentation-policy selection passed 84 tests and 1,494 subtests. The strict eleven-language build completed successfully, and the ten-route DOM validator retained all eight ordered navigation roots. The unchanged representative shell retains its preceding Playwright/Axe evidence; this data-safety and reference-prose slice makes no new visual-parity claim.

Committed HEAD 8ea5e941fcbcc0b93e5a4dd180b7a4c15c235930 retains successful remote Continuous Integration, Documentation, and Package CI runs, but those runs do not contain the current dirty worktree and are not exact-candidate evidence. The broader object-by-object public API audit, exact-current-worktree complete Python 3.10/3.11, Linux, Windows, default-runtime, tagged-workflow, publisher, and publication gates, five CUDA/Triton checks, and seven opt-in or inaccessible asset paths remain open. No protected action was taken. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree Trainer JSON read-safety evidence

This bounded exact-resume iteration refreshed Transformers main at commit d09f53a801f45ad73ec3510e17972024234bc0fd. Its revision-pinned 65-line Home, 1,576-line navigation, 500-line Modular Transformers, 30-line Trainer, and 2,906-line TrainingArguments sources have SHA-256 values e24e0eddadccef8b15e3f25c55acb0df0f65de21a4b5bf709d119d22d0eb555b, f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a, a52af1ee1754e6e658535f071a727fac31a057a3a848c28c18873a5ff371dd96, e7c5368c1223c2b195321468fcd1ac64c5cda52a1e44d4adf854a6473b6c9ee0, and cb55368c4d4b80633a7d289a790c5e4ab9f7ad03b527a969be7b0b0c2c432313. The official Home, navigation, Modular Transformers, Trainer, documentation specification, and TrainingArguments routes returned current content. Upstream supplies the structural training mental model; VoiceHub's speech-specific artifact safety controls this slice.

The observable contract requires TrainingArguments, TrainerState, checkpoint discovery, and exact-resume validation to reject duplicate object keys and NaN, positive or negative Infinity, or exponent overflow through the shared artifact reader. Object construction and model or runtime state restoration must not occur first. Invalid checkpoint candidates must be ignored rather than selected. Errors must identify the source and duplicate key or numeric path without printing discarded values, while valid saved objects and exact-resume behavior remain unchanged.

The pre-implementation focused command exited one with 11 failures. Both public object loaders accepted duplicates or constructed before rejecting overflow, discovery selected both invalid latest checkpoints, and checkpoint, optimization, and Trainer-state documents reached later validation rather than the parse boundary. Its three reported top-level passes are not transferred out of the failed run. TrainingArguments, TrainerState, checkpoint discovery, and all three exact-resume JSON reads now delegate to the same strict loader; no new abstraction, provider branch, or artifact format was introduced.

The corrected focused contract passed three tests and 11 subtests. Its first selected-file pre-commit run is excluded because YAPF reformatted the added test data; the formatted focused contract passed again, followed by every applicable selected source/test hook. The Markdown hook had no matching files and is not counted. The complete Trainer and training-runtime files passed 73 tests and 34 subtests. The proportional training, checkpoint, and optimization selection passed 285 tests and 692 subtests; the registry and lifecycle selection passed 139 tests and 499 subtests.

The complete Python 3.12.12 suite passed 2,558 executed tests and 4,074 subtests with 35 warnings in 105.99 seconds. Its 15 skips remain unpassed: three complete-dependency default-runtime checks, three Triton and two CUDA-extension checks, five opt-in ESPNet/NeMo/SenseVoice/SpeechBrain/TEN-VAD asset or oracle checks, and two inaccessible WeNet paths. Focused dependency-free probes passed on Python 3.10.19 and 3.11.15; complete exact- current-worktree execution on those interpreters remains pending.

The five-record release-alignment check passed. All 68 generated model pages and 59 generated model notebooks remain current. Refreshed inventories remain 68 models (34 TTS, 23 ASR, and 11 VAD), 102 aliases, zero invalid display names, 68 provider pages with no missing or orphaned page, 10 representative page pairs, six shared component/state groups, six public optimization passes, 408 model/pass pairs, eight top-level navigation roots, eight contribution steps, and 261 public exports. The public root import still avoids PyTorch.

Release, distribution-policy, and documentation-policy validation passed 77 tests and 1,476 subtests. A fresh isolated distribution check passed wheel, sdist, and editable installs with 68 models, 81 provenance manifests, 193 compliance files, all required package data, zero runtime-dependency violations, and no eager PyTorch import. It produced a 57,194,294-byte wheel and a 55,460,694-byte sdist.

After the evidence update, the focused Trainer readers plus release, distribution, and documentation-policy selection passed 80 tests and 1,487 subtests. The strict eleven-language build completed successfully, and the ten-route DOM validator retained all eight ordered navigation roots. The unchanged representative shell retains its preceding Playwright/Axe evidence; this data-safety and reference-prose slice makes no new visual-parity claim.

Committed HEAD 8ea5e941fcbcc0b93e5a4dd180b7a4c15c235930 retains successful remote Continuous Integration, Documentation, and Package CI runs, but those runs do not contain the current dirty worktree and are not exact-candidate evidence. The broader object-by-object public API audit, exact-current-worktree complete Python 3.10/3.11, Linux, Windows, default-runtime, tagged-workflow, publisher, and publication gates, five CUDA/Triton checks, and seven opt-in or inaccessible asset paths remain open. No protected action was taken. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree native checkpoint JSON read-safety evidence

This bounded data-safety iteration refreshed Transformers main at commit d09f53a801f45ad73ec3510e17972024234bc0fd. Its revision-pinned 65-line Home, 1,576-line navigation, 500-line Modular Transformers, 30-line Trainer, and 2,906-line TrainingArguments sources retain SHA-256 values e24e0eddadccef8b15e3f25c55acb0df0f65de21a4b5bf709d119d22d0eb555b, f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a, a52af1ee1754e6e658535f071a727fac31a057a3a848c28c18873a5ff371dd96, e7c5368c1223c2b195321468fcd1ac64c5cda52a1e44d4adf854a6473b6c9ee0, and cb55368c4d4b80633a7d289a790c5e4ab9f7ad03b527a969be7b0b0c2c432313. The official Home, navigation, documentation specification, Modular Transformers, and Trainer routes returned current content. Upstream supplies the structural artifact-loading mental model; VoiceHub's bounded native checkpoint parsers control this slice.

The observable contract requires VoiceHubManifest, bounded Safetensors headers, and sharded Safetensors indexes to reject duplicate object keys at every nesting level and reject NaN, positive or negative Infinity, and exponent overflow. Rejection must occur before manifest construction, tensor materialization, or shard lookup. Diagnostics must identify the source and duplicate key or numeric path without exposing a discarded value. Valid deterministic manifests, headers, indexes, and descriptive token_count metadata must retain their existing formats.

The pre-implementation focused command exited one with eight failing subcases; its three reported top-level passes are not transferred out of that failed run. The shared decoder now accepts an in-memory JSON document as well as a file and applies one duplicate-key and finite-number policy. All three native checkpoint readers delegate to it and translate failures into their existing CheckpointFormatError boundary. No provider, model, task, optimization-pass, or shard allowlist was introduced.

The corrected focused contract passed three tests and eight subtests. The complete native-checkpoint and shared-JSON files passed 26 tests and 26 subtests. The proportional checkpoint, configuration, public API, Trainer, training-runtime, and speech-core selection passed 165 tests and 89 subtests; the registry, speech-task, inference, and optimization selection passed 124 tests and 969 subtests. Every applicable selected source/test pre-commit hook passed. The Markdown hook had no selected Markdown file and is not counted. Dependency-light exact-slice probes passed on Python 3.10.19 and 3.11.15 without importing PyTorch; complete exact-current-worktree runs on those interpreters remain pending.

A first final policy command named the nonexistent tests/test_package_imports.py path and collected no tests; it is excluded. The corrected focused reader plus release, distribution, and documentation- policy selection passed 87 tests and 1,502 subtests.

The complete Python 3.12.12 suite passed 2,561 executed tests and 4,082 subtests with 35 warnings in 105.91 seconds. Its 15 skips remain unpassed: three complete-dependency default-runtime checks, three Triton and two CUDA-extension checks, five opt-in ESPNet/NeMo/SenseVoice/SpeechBrain/TEN-VAD asset or oracle checks, and two inaccessible WeNet paths.

The five-record release-alignment check passed. All 68 generated model pages and 59 generated model notebooks remain current. Refreshed inventories remain 68 models (34 TTS, 23 ASR, and 11 VAD), 102 aliases, zero invalid display names, 68 provider pages with no missing or orphaned page, 10 representative page pairs, six shared component/state groups, six public optimization passes, 408 model/pass pairs, eight top-level navigation roots, eight contribution steps, and 261 public exports. The public root and checkpoint-metadata imports remain PyTorch-free.

The strict eleven-language build completed in 33.26 seconds, and the ten-route DOM validator retained all eight ordered navigation roots. The unchanged representative shell retains its preceding Playwright/Axe evidence; this checkpoint data-safety slice makes no new visual-parity claim. A fresh isolated distribution check passed wheel, sdist, and editable installs with 68 models, 81 provenance manifests, 193 compliance files, all required package data, zero runtime-dependency violations, and no eager PyTorch import. It produced a 57,194,195-byte wheel and a 55,460,853-byte sdist.

Committed HEAD 8ea5e941fcbcc0b93e5a4dd180b7a4c15c235930 retains successful remote Continuous Integration, Documentation, and Package CI runs, but those runs do not contain the current dirty worktree and are not exact-candidate evidence. The broader object-by-object public API audit, exact-current-worktree complete Python 3.10/3.11, Linux, Windows, default-runtime, tagged-workflow, publisher, and publication gates, five CUDA/Triton checks, and seven opt-in or inaccessible asset paths remain open. No protected action was taken. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree portable-state config read-safety evidence

This bounded public-lifecycle iteration refreshed Transformers main at commit d09f53a801f45ad73ec3510e17972024234bc0fd. Its revision-pinned 65-line Home, 1,576-line navigation, 500-line Modular Transformers, 30-line Trainer, and 2,906-line TrainingArguments sources retain SHA-256 values e24e0eddadccef8b15e3f25c55acb0df0f65de21a4b5bf709d119d22d0eb555b, f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a, a52af1ee1754e6e658535f071a727fac31a057a3a848c28c18873a5ff371dd96, e7c5368c1223c2b195321468fcd1ac64c5cda52a1e44d4adf854a6473b6c9ee0, and cb55368c4d4b80633a7d289a790c5e4ab9f7ad03b527a969be7b0b0c2c432313. The official Home, navigation, documentation specification, Modular Transformers, and Trainer routes returned current content. Upstream provides the structural pretrained-model workflow; VoiceHub's portable speech-state boundary controls this slice.

The observable contract requires both public pretrained base loaders to apply the shared strict JSON policy to the sibling config.json when restoring a local model_state.pt, including when the caller supplied config= and the ordinary configuration factory was bypassed. Duplicate object keys, NaN, positive or negative Infinity, and exponent overflow must fail before TTS, ASR, or VAD wrapper construction. Diagnostics must identify the artifact and duplicate key or numeric path without exposing a discarded value. A valid saved name_or_path and descriptive token_count metadata must retain the existing lazy restoration behavior.

The pre-implementation focused command exited one with six failing subcases: both base loaders accepted every ambiguous document and reached the guarded constructor. Its two reported top-level passes are not transferred out of the failed run. The two permissive reads now delegate to the existing shared reader; no new abstraction, provider branch, task allowlist, or artifact format was introduced.

The first expanded post-implementation command is also excluded: all six malformed subcases passed, but two valid-artifact assertions compared macOS /var and /private/var spellings without resolving them. After correcting only that test expectation, the focused contract passed two tests and six subtests. The proportional public base, configuration, pipeline, inference, training, Trainer, and portable-state selection passed 221 tests and 181 subtests. Registry, speech-task, and universal-optimization coverage passed 76 tests and 953 subtests. Every applicable selected source/test pre-commit hook passed; the Markdown hook had no selected Markdown file and is not counted. The public TTS, ASR, and VAD base imports remain PyTorch-free.

Direct pytest execution was unavailable in the standalone Python 3.10.19 and 3.11.15 installations because neither interpreter has pytest; those two failed commands collected no tests and are not evidence. Corrected dependency- light scripts then exercised duplicate, constant, and overflow rejection for both base loaders before construction on both interpreters and confirmed that PyTorch remained unimported. Complete exact-current-worktree suites on Python 3.10 and 3.11 remain pending.

The complete Python 3.12.12 suite passed 2,563 executed tests and 4,088 subtests with 35 warnings in 105.16 seconds. Its 15 skips remain unpassed: three complete-dependency default-runtime checks, three Triton and two CUDA-extension checks, five opt-in ESPNet/NeMo/SenseVoice/SpeechBrain/TEN-VAD asset or oracle checks, and two inaccessible WeNet paths.

The five-record release-alignment check passed. All 68 generated model pages and 59 generated model notebooks remain current. Refreshed inventories remain 68 models (34 TTS, 23 ASR, and 11 VAD), 102 aliases, zero invalid display names, 68 provider pages with no missing or orphaned page, 10 representative page pairs, six shared component/state groups, six public optimization passes, 408 model/pass pairs, eight top-level navigation roots, eight contribution steps, and 261 public exports.

The strict eleven-language build completed in 33.88 seconds, and the ten-route DOM validator retained all eight ordered navigation roots. The unchanged representative shell retains its preceding Playwright/Axe evidence; this public data-safety slice makes no new visual-parity claim. A fresh isolated distribution check passed wheel, sdist, and editable installs with 68 models, 81 provenance manifests, 193 compliance files, all required package data, zero runtime-dependency violations, and no eager PyTorch import. It produced a 57,194,160-byte wheel and a 55,460,813-byte sdist.

Committed HEAD 8ea5e941fcbcc0b93e5a4dd180b7a4c15c235930 retains successful remote Continuous Integration, Documentation, and Package CI runs, but those runs do not contain the current dirty worktree and are not exact-candidate evidence. The broader object-by-object public API audit, exact-current-worktree complete Python 3.10/3.11, Linux, Windows, default-runtime, tagged-workflow, publisher, and publication gates, five CUDA/Triton checks, and seven opt-in or inaccessible asset paths remain open. No protected action was taken. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree Hub JSON trust-boundary evidence

This bounded public-artifact iteration refreshed Transformers main at commit d09f53a801f45ad73ec3510e17972024234bc0fd. Its revision-pinned 65-line Home, 1,576-line navigation, 500-line Modular Transformers, 30-line Trainer, and 2,906-line TrainingArguments sources retain the release-ledger SHA-256 fingerprints. The official Home, navigation, documentation specification, Modular Transformers, and Trainer routes returned current content. Upstream provides the structural pretrained-model workflow; VoiceHub owns the native Hub transport and cache policy exercised here.

The observable contract requires remote Hub API payloads to reject duplicate object keys, NaN, positive or negative Infinity, and exponent overflow before a commit or repository tree is interpreted. Ambiguous VoiceHub file metadata and snapshot manifests must be treated as cache misses in offline mode. Diagnostics must identify the source context and duplicate key or numeric path without exposing a discarded value or authentication token. Valid online and offline resolution, redirects, list-valued repository trees, and the dependency-light import boundary must retain their behavior.

The pre-implementation focused command exited one with five failures and one reported pass: all three malformed remote payloads reached the guarded tree request, the ambiguous file metadata returned a commit, and the ambiguous snapshot manifest was reused. No result from that failed command is counted. The strict decoder now lives in a dependency-light shared module used by local artifacts, Safetensors metadata, Hub API payloads, and native Hub cache metadata. Remote ambiguity becomes HubDownloadError; ambiguous cached objects become misses. No provider, model, task, or optimization allowlist was introduced.

The corrected focused contract passed three tests and three subtests. Hub, strict-artifact, and native-checkpoint coverage passed 48 tests and 34 subtests. Public configuration, base, speech-core, inference, and pipeline lifecycle coverage passed 123 tests and 59 subtests. Registry, speech-task, and universal-optimization coverage passed 76 tests and 953 subtests. Every applicable selected source/test pre-commit hook passed. Dependency-light exact probes on Python 3.10.19 and 3.11.15 rejected duplicate, constant, and overflow Hub JSON without leaking the discarded value or runtime token and preserved a PyTorch-free import. Complete exact-current-worktree suites on those two interpreters remain pending.

The complete Python 3.12.12 suite passed 2,566 executed tests and 4,091 subtests with 35 warnings in 105.91 seconds. Its 15 skips remain unpassed: three complete-dependency default-runtime checks, three Triton and two CUDA-extension checks, five opt-in ESPNet/NeMo/SenseVoice/SpeechBrain/TEN-VAD asset or oracle checks, and two inaccessible WeNet paths.

The five-record release-alignment check passed. All 68 generated model pages and 59 generated model notebooks remain current. Refreshed inventories remain 68 models (34 TTS, 23 ASR, and 11 VAD), 102 aliases, zero invalid display names, 68 provider pages with no missing or orphaned page, 10 representative page pairs, six shared component/state groups, six public optimization passes, 408 model/pass pairs, eight top-level navigation roots, eight contribution steps, and 261 public exports.

The strict eleven-language build completed in 35.30 seconds, and the ten-route DOM validator retained all eight ordered navigation roots. The unchanged representative shell retains its preceding Playwright/Axe evidence; this artifact-safety slice makes no new visual-parity claim. A fresh isolated distribution check passed wheel, sdist, and editable installs with 68 models, 81 provenance manifests, 193 compliance files, all required package data, zero runtime-dependency violations, and no eager PyTorch import. It produced a 57,194,648-byte wheel and a 55,461,196-byte sdist.

Committed HEAD 8ea5e941fcbcc0b93e5a4dd180b7a4c15c235930 retains successful remote Continuous Integration, Documentation, and Package CI runs, but those runs do not contain the current dirty worktree and are not exact-candidate evidence. The broader object-by-object public API audit, exact-current-worktree complete Python 3.10/3.11, Linux, Windows, default-runtime, tagged-workflow, publisher, and publication gates, five CUDA/Triton checks, and seven opt-in or inaccessible asset paths remain open. No protected action was taken. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree tokenizer JSON trust-boundary evidence

This bounded public-artifact iteration refreshed Transformers main at commit d09f53a801f45ad73ec3510e17972024234bc0fd. Its revision-pinned 65-line Home, 1,576-line navigation, 500-line Modular Transformers, 30-line Trainer, and 2,906-line TrainingArguments sources retain SHA-256 values e24e0eddadccef8b15e3f25c55acb0df0f65de21a4b5bf709d119d22d0eb555b, f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a, a52af1ee1754e6e658535f071a727fac31a057a3a848c28c18873a5ff371dd96, e7c5368c1223c2b195321468fcd1ac64c5cda52a1e44d4adf854a6473b6c9ee0, and cb55368c4d4b80633a7d289a790c5e4ab9f7ad03b527a969be7b0b0c2c432313. The official Home, navigation, documentation specification, Modular Transformers, and Trainer routes returned HTTP 200. Upstream supplies the pretrained tokenizer workflow; VoiceHub owns the bounded declarative tokenizer parsers exercised here.

The observable contract requires both shared tokenizer JSON entry points—the Hugging Face Byte-BPE loader and the SentencePiece-BPE loader—to reject duplicate object keys, NaN, positive or negative Infinity, and exponent overflow before interpreting the model graph, vocabulary, merges, or added tokens. Diagnostics must identify the source path and duplicate key or numeric path without exposing a discarded value. Valid tokenizer loading and saving, byte/token/merge/nesting/node limits, registered consumers, and PyTorch-free imports must retain their behavior.

The pre-implementation focused command exited one with six failing subcases and two reported top-level passes. The Byte-BPE loader accepted a duplicate model and exponent overflow and omitted source context for NaN; the SentencePiece-BPE loader omitted source context for its existing duplicate and constant rejection and accepted exponent overflow. No result from that failed command is counted. Both loaders now delegate decoding to the existing shared strict JSON boundary before applying their tokenizer-specific graph and size validators. The redundant SentencePiece duplicate/constant hooks and the Byte-BPE constant hook were removed. No provider, model, task, or optimization allowlist was introduced.

The corrected focused contract passed two tests and six subtests. Core tokenizer, Moonshine, VoxCPM2, and shared-artifact coverage passed 59 tests and 27 subtests with four warnings. Eleven additional native consumer suites passed 166 tests and 27 subtests. Registry, speech-task, universal-optimization, model-page/navigation, and native dependency-policy coverage passed 159 tests and 2,429 subtests. The first selected-file pre-commit command exited one after the end-of-file hook repaired one missing final newline and is excluded; the complete corrected selected-file command then passed every applicable hook. The Markdown hook found no selected Markdown file and is not counted.

The first standalone Python 3.10.19 and 3.11.15 probe imported a test helper that requires unavailable PyTorch, so both invocations failed before exercising the contract and are excluded. Corrected dependency-light probes constructed both valid tokenizer graphs directly, loaded them successfully, rejected the three ambiguous variants for each loader with source-aware diagnostics, and confirmed PyTorch-free imports on both interpreters. Complete exact-current- worktree suites on Python 3.10 and 3.11 remain pending.

The complete Python 3.12.12 suite passed 2,568 executed tests and 4,097 subtests with 35 warnings in 110.46 seconds. Its 15 skips remain unpassed: three complete-dependency default-runtime checks, three Triton and two CUDA-extension checks, five opt-in ESPNet/NeMo/SenseVoice/SpeechBrain/TEN-VAD asset or oracle checks, and two inaccessible WeNet paths.

The five-record release-alignment check passed. All 68 generated model pages and 59 generated model notebooks remain current. Refreshed inventories remain 68 models (34 TTS, 23 ASR, and 11 VAD), 102 aliases, zero invalid display names, 68 provider pages with no missing or orphaned page, 10 representative page pairs, six shared component/state groups, six public optimization passes, 408 model/pass pairs, eight top-level navigation roots, eight contribution steps, and 261 public exports.

The strict eleven-language build completed in 35.92 seconds, and the ten-route DOM validator retained all eight ordered navigation roots. The unchanged representative shell retains its preceding Playwright/Axe evidence; this artifact-safety slice makes no new visual-parity claim. A fresh isolated distribution check passed wheel, sdist, and editable installs with 68 models, 81 provenance manifests, 193 compliance files, all required package data, zero runtime-dependency violations, and no eager PyTorch import. It produced a 57,194,432-byte wheel and a 55,462,589-byte sdist.

Committed HEAD 8ea5e941fcbcc0b93e5a4dd180b7a4c15c235930 retains successful remote Continuous Integration, Documentation, and Package CI runs, but those runs do not contain the current dirty worktree and are not exact-candidate evidence. Newer successful runs for unrelated head 8fee3d7 likewise do not cover this worktree. The broader object-by-object public API audit, exact- current-worktree complete Python 3.10/3.11, Linux, Windows, default-runtime, tagged-workflow, publisher, and publication gates, five CUDA/Triton checks, and seven opt-in or inaccessible asset paths remain open. No protected action was taken. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree speech manifest JSON trust-boundary evidence

This bounded public-dataset iteration refreshed Transformers main at commit d09f53a801f45ad73ec3510e17972024234bc0fd. Its revision-pinned 65-line Home, 1,576-line navigation, 500-line Modular Transformers, 30-line Trainer, and 2,906-line TrainingArguments sources retain SHA-256 values e24e0eddadccef8b15e3f25c55acb0df0f65de21a4b5bf709d119d22d0eb555b, f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a, a52af1ee1754e6e658535f071a727fac31a057a3a848c28c18873a5ff371dd96, e7c5368c1223c2b195321468fcd1ac64c5cda52a1e44d4adf854a6473b6c9ee0, and cb55368c4d4b80633a7d289a790c5e4ab9f7ad03b527a969be7b0b0c2c432313. The official Home, navigation, documentation specification, Modular Transformers, and Trainer routes returned HTTP 200. VoiceHub owns the speech manifest parsers exercised here.

The observable contract requires public TTS and ASR .json, JSON Lines, and embedded CSV/TSV JSON values to reject duplicate keys, NaN, positive or negative Infinity, and exponent overflow before dataset construction. Diagnostics must name the file and, where available, line and field without exposing a discarded value. Valid object/list manifests, NeMo JSON Lines fallback, tabular scalar coercion, credential checks, portable round trips, and lazy imports must retain their behavior.

The pre-implementation focused command exited one with 18 failing subcases and two reported top-level passes; none of that command is counted. Both dataset readers now use the shared strict JSON decoder for whole JSON files, JSON Lines records, and JSON-shaped tabular fields. Tabular syntax errors keep the established string fallback. ASR .json falls back to NeMo JSON Lines only for a syntax error, never for a duplicate or non-finite value. No provider, model, task, or optimization allowlist was introduced.

The corrected focused contract passed two tests and 18 subtests. Complete TTS and ASR dataset coverage passed 81 tests and 134 subtests. Proportional core training, Trainer, adapter, collator, and speech-training coverage passed 168 tests and 212 subtests. Registry, speech-task, universal-optimization, model-page/navigation, model-scaffold, and native dependency-policy coverage passed 179 tests and 2,464 subtests. The first selected-file pre-commit command exited one after YAPF changed the files and is excluded; the corrected command then passed every applicable hook. The Markdown hook found no applicable file and is not counted.

Dependency-light probes on Python 3.10.19 and 3.11.15 each rejected all 18 task/format/ambiguity paths and retained PyTorch-free public imports. Complete exact-current-worktree suites on Python 3.10 and 3.11 remain pending. The complete Python 3.12.12 suite passed 2,570 executed tests and 4,115 subtests with 35 warnings in 110.64 seconds. Its 15 skips remain unpassed: three complete-dependency default-runtime checks, three Triton and two CUDA-extension checks, five opt-in ESPNet/NeMo/SenseVoice/SpeechBrain/TEN-VAD asset or oracle checks, and two inaccessible WeNet paths.

The five-record release-alignment check passed. All 68 generated model pages and 59 generated model notebooks remain current. Refreshed inventories remain 68 models (34 TTS, 23 ASR, and 11 VAD), 102 aliases, zero invalid display names, 68 provider pages with no missing or orphaned page, 10 representative page pairs, six shared component/state groups, six public optimization passes, 408 model/pass pairs, eight top-level navigation roots, eight contribution steps, and 261 public exports.

The strict eleven-language build passed, and the ten-route DOM validator retained all eight ordered navigation roots. The unchanged representative shell retains its preceding Playwright/Axe evidence; this dataset-safety slice makes no new visual-parity claim. A fresh isolated distribution check passed wheel, sdist, and editable installs with 68 models, 81 provenance manifests, 193 compliance files, all required package data, zero runtime-dependency violations, and no eager PyTorch import. It produced a 57,194,588-byte wheel and a 55,464,587-byte sdist.

Committed HEAD 8ea5e941fcbcc0b93e5a4dd180b7a4c15c235930 retains successful remote Continuous Integration, Documentation, and Package CI runs, but those runs do not contain the current dirty worktree and are not exact-candidate evidence. Newer successful runs for unrelated head 8fee3d7 likewise do not cover this worktree. The broader object-by-object public API audit, exact- current-worktree complete Python 3.10/3.11, Linux, Windows, default-runtime, tagged-workflow, publisher, and publication gates, five CUDA/Triton checks, and seven opt-in or inaccessible asset paths remain open. No protected action was taken. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree external LLM JSON response trust-boundary evidence

This bounded public-serving iteration refreshed Transformers main at commit d09f53a801f45ad73ec3510e17972024234bc0fd. Its revision-pinned 65-line Home, 1,576-line navigation, 500-line Modular Transformers, 30-line Trainer, and 2,906-line TrainingArguments sources retain SHA-256 values e24e0eddadccef8b15e3f25c55acb0df0f65de21a4b5bf709d119d22d0eb555b, f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a, a52af1ee1754e6e658535f071a727fac31a057a3a848c28c18873a5ff371dd96, e7c5368c1223c2b195321468fcd1ac64c5cda52a1e44d4adf854a6473b6c9ee0, and cb55368c4d4b80633a7d289a790c5e4ab9f7ad03b527a969be7b0b0c2c432313. The official Home, navigation, documentation specification, Modular Transformers, and Trainer routes returned current content. VoiceHub owns the external speech-serving transport exercised here.

The observable contract requires bounded vLLM and SGLang JSON responses to reject duplicate keys, NaN, positive or negative Infinity, and exponent overflow before token IDs, usage metadata, or other protocol fields are interpreted. Diagnostics must identify the backend route and offending key or numeric path without exposing a discarded value. Valid responses, request serialization, response-size limits, redirect rejection, credential redaction, backend capability dispatch, and lazy imports must retain their behavior.

The first pre-implementation command is excluded because a test-control-flow error produced a secondary AttributeError after the expected duplicate-key failure. After correcting only that test structure, the valid red command exited one with all three ambiguity subcases failing and one reported top-level pass; none of that command is counted. The HTTP response boundary now delegates decoding to the shared strict JSON parser and wraps its source-aware failure in the existing sanitized LLMBackendRequestError. No provider, model, task, or optimization allowlist was introduced.

The corrected focused contract passed one test and three subtests. The complete LLM-serving suite passed 53 tests and 57 subtests. Proportional diffusion- serving, inference, base-model, speech-core, registry, speech-task, universal-optimization, native-optimization, and native-architecture coverage passed 237 tests and 1,185 subtests. Release, distribution, and documentation policy coverage passed 77 tests and 1,476 subtests. The selected-file pre-commit command passed every applicable hook; the Markdown hook found no applicable file and is not counted.

Dependency-light probes on Python 3.10.19 and 3.11.15 each rejected all three ambiguity forms, accepted one valid response, and retained PyTorch-free public imports. Complete exact-current-worktree suites on Python 3.10 and 3.11 remain pending. The complete Python 3.12.12 suite passed 2,571 executed tests and 4,118 subtests with 35 warnings in 109.45 seconds. Its 15 skips remain unpassed: three complete-dependency default-runtime checks, three Triton and two CUDA-extension checks, five opt-in ESPNet/NeMo/SenseVoice/SpeechBrain/ TEN-VAD asset or oracle checks, and two inaccessible WeNet paths.

The five-record release-alignment check passed. All 68 generated model pages and 59 generated model notebooks remain current. Refreshed inventories remain 68 models (34 TTS, 23 ASR, and 11 VAD), 102 aliases, zero invalid display names, 68 provider pages with no missing or orphaned page, 10 representative page pairs, six shared component/state groups, six public optimization passes, 408 model/pass pairs, eight top-level navigation roots, eight contribution steps, and 261 public exports.

The strict eleven-language build and ten-route DOM validator passed and retained all eight ordered navigation roots. The unchanged representative shell retains its preceding Playwright/Axe evidence; this serving-safety slice makes no new visual-parity claim. A fresh isolated distribution check passed wheel, sdist, and editable installs with 68 models, 81 provenance manifests, 193 compliance files, all required package data, zero runtime-dependency violations, and no eager PyTorch import. It produced a 57,194,624-byte wheel and a 55,465,049-byte sdist.

Committed HEAD 8ea5e941fcbcc0b93e5a4dd180b7a4c15c235930 retains successful remote Continuous Integration, Documentation, and Package CI runs, but those runs do not contain the current dirty worktree and are not exact-candidate evidence. Newer successful runs for unrelated head 8fee3d7 likewise do not cover this worktree. The broader object-by-object public API audit, exact- current-worktree complete Python 3.10/3.11, Linux, Windows, default-runtime, tagged-workflow, publisher, and publication gates, five CUDA/Triton checks, and seven opt-in or inaccessible asset paths remain open. No protected action was taken. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree model-integration JSON trust-boundary evidence

This bounded contribution-path iteration refreshed Transformers main at commit d09f53a801f45ad73ec3510e17972024234bc0fd. Its revision-pinned 65-line Home, 1,576-line navigation, 500-line Modular Transformers, 30-line Trainer, and 2,906-line TrainingArguments sources retain SHA-256 values e24e0eddadccef8b15e3f25c55acb0df0f65de21a4b5bf709d119d22d0eb555b, f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a, a52af1ee1754e6e658535f071a727fac31a057a3a848c28c18873a5ff371dd96, e7c5368c1223c2b195321468fcd1ac64c5cda52a1e44d4adf854a6473b6c9ee0, and cb55368c4d4b80633a7d289a790c5e4ab9f7ad03b527a969be7b0b0c2c432313. The official Home, navigation, documentation specification, Modular Transformers, and Trainer routes returned current content. VoiceHub owns the model-contribution metadata boundary exercised here.

The observable contract requires activated package-local model-integration.json files and their required source/SOURCE.json records to reject duplicate keys, NaN, positive or negative Infinity, and exponent overflow before registry or training-profile construction. The standalone scaffold checker and catalog renderer must enforce the same ambiguity boundary without importing VoiceHub or PyTorch. Diagnostics must identify the file and offending key or numeric path without exposing a discarded value. Inactive work-in-progress manifests must remain undiscovered, and valid zero-central- edit TTS, ASR, and VAD discovery must remain unchanged.

The pre-implementation command exited one with six failures, two reported passes, and three reported subtests; that entire command is excluded. Activated runtime discovery now strictly reparses its manifest and provenance record, while the initial bounded parse retains the documented inactive-draft escape hatch. The standalone checker and catalog renderer use an equivalent local strict decoder to preserve their no-package-import boundary. No provider, model, task, or optimization allowlist was introduced.

The corrected focused contract passed two tests and nine subtests; the complete scaffold suite passed 22 tests and 44 subtests. Proportional scaffold, registry, speech-task, training, documentation, release, universal- optimization, native-optimization, and distribution coverage passed 241 tests and 2,659 subtests. The selected-file pre-commit command passed every applicable hook; the Markdown hook found no applicable file and is not counted.

Dependency-light probes on Python 3.10.19 and 3.11.15 each rejected all six active artifact/ambiguity combinations, accepted one valid activated manifest, and retained VoiceHub- and PyTorch-free standalone-tool imports. Complete exact-current-worktree suites on Python 3.10 and 3.11 remain pending. The complete Python 3.12.12 suite passed 2,573 executed tests and 4,127 subtests with 35 warnings in 109.63 seconds. Its 15 skips remain unpassed: three complete-dependency default-runtime checks, three Triton and two CUDA-extension checks, five opt-in ESPNet/NeMo/SenseVoice/SpeechBrain/TEN-VAD asset or oracle checks, and two inaccessible WeNet paths.

The five-record release-alignment check passed. All 68 generated model pages and 59 generated model notebooks remain current. Refreshed inventories remain 68 models (34 TTS, 23 ASR, and 11 VAD), 102 aliases, zero invalid display names, 68 provider pages with no missing or orphaned page, 10 representative page pairs, six shared component/state groups, six public optimization passes, 408 model/pass pairs, eight top-level navigation roots, eight contribution steps, and 261 public exports.

The strict eleven-language build and ten-route DOM validator passed and retained all eight ordered navigation roots. The unchanged representative shell retains its preceding Playwright/Axe evidence; this contribution-safety slice makes no new visual-parity claim. A fresh isolated distribution check passed wheel, sdist, and editable installs with 68 models, 81 provenance manifests, 193 compliance files, all required package data, zero runtime- dependency violations, and no eager PyTorch import. It produced a 57,194,699-byte wheel and a 55,466,554-byte sdist.

Committed HEAD 8ea5e941fcbcc0b93e5a4dd180b7a4c15c235930 retains successful remote Continuous Integration, Documentation, and Package CI runs, but those runs do not contain the current dirty worktree and are not exact-candidate evidence. Newer successful runs for unrelated head 8fee3d7 likewise do not cover this worktree. The broader object-by-object public API audit, exact- current-worktree complete Python 3.10/3.11, Linux, Windows, default-runtime, tagged-workflow, publisher, and publication gates, five CUDA/Triton checks, and seven opt-in or inaccessible asset paths remain open. No protected action was taken. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree shared JSON artifact byte-bound evidence

This bounded public-artifact iteration refreshed Transformers main at commit d09f53a801f45ad73ec3510e17972024234bc0fd. Its revision-pinned 65-line Home, 1,576-line navigation, 500-line Modular Transformers, 30-line Trainer, and 2,906-line TrainingArguments sources retain SHA-256 values e24e0eddadccef8b15e3f25c55acb0df0f65de21a4b5bf709d119d22d0eb555b, f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a, a52af1ee1754e6e658535f071a727fac31a057a3a848c28c18873a5ff371dd96, e7c5368c1223c2b195321468fcd1ac64c5cda52a1e44d4adf854a6473b6c9ee0, and cb55368c4d4b80633a7d289a790c5e4ab9f7ad03b527a969be7b0b0c2c432313. The official Home, navigation, documentation specification, Modular Transformers, and Trainer routes returned current content. VoiceHub owns the shared local and downloaded artifact boundary exercised here.

The observable contract requires every JSON object routed through read_json_file() to have a validated positive integer byte ceiling before decoding. The default ceiling is 64 MiB, and an owning internal loader may choose a smaller limit. A document already over the limit, or one that grows past it between metadata inspection and reading, must fail before JSON parsing. Diagnostics must identify the source and actual size or configured limit without exposing document content. Exact-limit finite JSON, existing strict duplicate/non-finite rejection, and lazy dependency behavior must remain unchanged.

The pre-implementation focused command exited one with six failures and one reported pass; that entire command is excluded. The shared reader now validates the caller's limit, rejects an already oversized file from metadata, and reads at most one byte beyond the ceiling so concurrent growth also fails closed. Configuration, auto-loading, processing, training, checkpoint, and native- artifact callers inherit the default without provider branches or duplicated reader logic.

The corrected focused contract passed four tests and four subtests; the complete shared JSON artifact suite passed 11 tests and 22 subtests. Proportional public configuration, auto-loading, pipeline, inference, speech-core, checkpoint, Trainer, registry, task, optimization, documentation, release, and distribution coverage passed 409 tests and 2,663 subtests. The first selected- file pre-commit run is excluded because YAPF reformatted the new code. The formatted rerun passed every applicable hook; the Markdown hook found no applicable file and is not counted.

Dependency-light probes on Python 3.10.19 and 3.11.15 each passed four exact- slice checks and retained PyTorch-free imports. Complete exact-current- worktree suites on Python 3.10 and 3.11 remain pending. The complete Python 3.12.12 suite passed 2,577 executed tests and 4,131 subtests with 35 warnings in 114.47 seconds. Its 15 skips remain unpassed: three complete-dependency default-runtime checks, three Triton and two CUDA-extension checks, five opt-in ESPNet/NeMo/SenseVoice/SpeechBrain/TEN-VAD asset or oracle checks, and two inaccessible WeNet paths.

The inventory suite passed 198 tests and 2,608 subtests. The five-record release-alignment check passed. All 68 generated model pages and 59 generated model notebooks remain current. Refreshed inventories remain 68 models (34 TTS, 23 ASR, and 11 VAD), 102 aliases, zero invalid display names, 68 provider pages with no missing or orphaned page, 10 representative page pairs, six shared component/state groups, six public optimization passes, 408 model/pass pairs, eight top-level navigation roots, eight contribution steps, and 261 public exports.

The strict eleven-language build completed in 34.41 seconds, and the ten-route DOM validator retained all eight ordered navigation roots. The unchanged representative shell retains its preceding Playwright/Axe evidence; this artifact-safety slice makes no new visual-parity claim. A fresh isolated distribution check passed wheel, sdist, and editable installs with 68 models, 81 provenance manifests, 193 compliance files, all required package data, zero runtime-dependency violations, and no eager PyTorch import. It produced a 57,194,971-byte wheel and a 55,467,121-byte sdist.

Committed HEAD 8ea5e941fcbcc0b93e5a4dd180b7a4c15c235930 retains successful remote Continuous Integration, Documentation, and Package CI runs, but those runs do not contain the current dirty worktree and are not exact-candidate evidence. Newer successful runs for unrelated head 8fee3d7 likewise do not cover this worktree. The broader object-by-object public API audit, exact- current-worktree complete Python 3.10/3.11, Linux, Windows, default-runtime, tagged-workflow, publisher, and publication gates, five CUDA/Triton checks, and seven opt-in or inaccessible asset paths remain open. No protected action was taken. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree supported-Python macOS evidence

This bounded platform iteration refreshed Transformers main at commit d09f53a801f45ad73ec3510e17972024234bc0fd. Its revision-pinned 65-line Home, 1,576-line navigation, 500-line Modular Transformers, 30-line Trainer, and 2,906-line TrainingArguments sources retain SHA-256 values e24e0eddadccef8b15e3f25c55acb0df0f65de21a4b5bf709d119d22d0eb555b, f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a, a52af1ee1754e6e658535f071a727fac31a057a3a848c28c18873a5ff371dd96, e7c5368c1223c2b195321468fcd1ac64c5cda52a1e44d4adf854a6473b6c9ee0, and cb55368c4d4b80633a7d289a790c5e4ab9f7ad03b527a969be7b0b0c2c432313. The official Home, navigation, documentation specification, Modular Transformers, and Trainer routes returned current content. VoiceHub's supported-Python declaration and current test corpus control this slice.

The observable contract requires complete exact-worktree pytest execution on each supported interpreter. Each non-default interpreter must run in a fresh temporary virtual environment, import VoiceHub from this checkout, use the declared test extra, and avoid every lockfile operation. Dependency setup, test execution, platform, versions, duration, warnings, and skips must be recorded separately. A setup failure, partial run, or skip is not a pass.

uv 0.11.21 created isolated environments with CPython 3.10.19 and 3.11.15. UV_TORCH_BACKEND=cpu uv pip install --python <environment> -e '.[test]' resolved and installed the declared test dependencies without using uv.lock. Both environments imported /Users/kadirnar/Documents/voicehub/voicehub/__init__.py. They used macOS 26.5.2 arm64, PyTorch 2.8.0, Transformers 5.14.1, and pytest 9.1.1. The Python 3.10 environment resolved 93 packages and the Python 3.11 environment resolved 90 packages under their respective compatibility constraints.

The complete Python 3.10.19 suite passed 2,577 executed tests and 4,131 subtests with 35 warnings in 173.72 seconds. The complete Python 3.11.15 suite passed the same 2,577 executed tests and 4,131 subtests with 35 warnings in 163.92 seconds. The existing exact-current Python 3.12.12 suite passed those same counts in 114.47 seconds with the same PyTorch, Transformers, and pytest versions. Each interpreter reported 15 skips, which remain unpassed: three complete-dependency default-runtime checks, three Triton and two CUDA-extension checks, five opt-in ESPNet/NeMo/SenseVoice/SpeechBrain/TEN-VAD asset or oracle checks, and two inaccessible WeNet paths.

Only this evidence record changed after the complete Python 3.10 and 3.11 runs. The release, distribution, and documentation-policy selection then passed 77 tests and 1,476 subtests independently on Python 3.10.19, 3.11.15, and 3.12.12. The strict eleven-language build completed in 35.55 seconds, and the ten-route DOM validator retained all eight ordered navigation roots. A fresh distribution check passed isolated wheel, sdist, and editable installs with 68 models, 81 provenance manifests, 193 compliance files, required package data, zero runtime-dependency violations, and no eager PyTorch import. It produced a 57,194,971-byte wheel and a 55,466,862-byte sdist.

No runtime, test, workflow, package, generated page, or generated notebook source changed in this platform-only slice. The preceding inventory evidence therefore remains 68 models (34 TTS, 23 ASR, and 11 VAD), 102 aliases, zero invalid display names, 68 current provider pages, 59 current notebooks, 10 representative page pairs, six shared component/state groups, six public optimization passes, 408 model/pass pairs, eight top-level navigation roots, eight contribution steps, 261 public exports, and five benchmark records.

Committed HEAD 8ea5e941fcbcc0b93e5a4dd180b7a4c15c235930 retains successful remote Continuous Integration, Documentation, and Package CI runs, but those runs do not contain the current dirty worktree and are not exact-candidate evidence. Newer successful runs for unrelated head 8fee3d7 likewise do not cover this worktree. Exact-current-worktree Linux, Windows, default-runtime, tagged-workflow, publisher, and publication gates, five CUDA/Triton checks, seven opt-in or inaccessible asset paths, and the broader object-by-object public API audit remain open. No protected action was taken. The untracked uv.lock stayed unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact current-worktree full-dependency default-runtime evidence

This bounded release iteration refreshed Transformers main at commit d09f53a801f45ad73ec3510e17972024234bc0fd. Its revision-pinned Home, navigation, Modular Transformers, Trainer, and TrainingArguments sources retained the fingerprints recorded in the parity ledger. The official navigation, documentation specification, Modular Transformers, and Trainer routes returned current content; the documentation Home endpoint returned an HTTP 429 during this retrieval and is not reported as passed.

The observable gate requires a fresh Python 3.12 environment to install the checkout directly with all three declared extras, keep the lock file untouched, pass dependency validation, execute the focused packaging and default-runtime contracts, and complete the entire suite with VOICEHUB_FULL_RUNTIME_TEST=1. A setup failure, skip, or partial run is not a pass.

The first direct uv pip install -e ".[test,training,docs]" resolved the broad librosa transitive range to numba 0.53.1 and failed because that release rejects Python 3.12.12. That setup is recorded as failed evidence. The test extra now declares the Python-3.12-compatible lower bound numba>=0.59 while keeping numba outside the default inference installation. A focused metadata contract protects that boundary.

In the same still-empty temporary environment, uv 0.11.21 then installed all 135 packages directly from the current checkout with its CPU PyTorch resolver and without synchronization or a lockfile operation. uv pip check reported all packages compatible. VoiceHub imported from this checkout on macOS 26.5.2 arm64 with Python 3.12.12, PyTorch 2.8.0, Transformers 5.14.1, and pytest 9.1.1. The focused metadata file passed 11 tests and 76 subtests, and the activated default-runtime file passed five tests and 138 subtests.

The complete activated suite passed 2,581 tests and 4,269 subtests with 35 warnings in 176.53 seconds. Its 12 skips remain unpassed:

• Three Triton and two compiled CUDA-extension checks require unavailable CUDA hardware or a CUDA toolkit.
• The ESPNet, NeMo QuartzNet, SenseVoice, SpeechBrain, and TEN-VAD artifact or oracle checks remain opt-in and were not executed in this run.
• The WeNet checkpoint and tokenizer artifacts remain inaccessible.

Proportional registry, task, public-API, optimization, model-scaffold, documentation, release, and packaging coverage passed 205 tests and 2,556 subtests. The release inventory retained five benchmark records and 68 documented providers: 34 TTS, 23 ASR, and 11 VAD. All 68 generated model pages and 59 generated model notebooks remain current. The strict eleven-language build completed in 35.90 seconds, and the ten-route DOM validator retained all eight ordered navigation roots. This slice changes no rendered component and makes no new visual-parity claim.

A fresh distribution check passed isolated wheel, source-distribution, and editable installs with 68 models, 81 provenance manifests, 193 compliance files, required package data, zero runtime-dependency violations, and no eager PyTorch import. It produced a 57,194,978-byte wheel and a 55,467,389-byte source distribution. The release report itself changed after this build, so the source-distribution size is execution evidence rather than a final tagged- artifact fingerprint.

This closes the locally executable full-dependency/default-runtime gate. It does not substitute for exact-current-worktree Linux, Windows, or remote default-runtime CI, and it does not refresh any asset, hardware, tagged- workflow, publisher, or publication gate. No protected action was taken. The untracked uv.lock remained unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Current generated package-root public API evidence

This public-contract iteration refreshed Transformers main at commit d09f53a801f45ad73ec3510e17972024234bc0fd. The revision-pinned Home, navigation, Modular Transformers, Trainer, and TrainingArguments sources retain the fingerprints recorded in the parity ledger. The official Transformers Models main-class page was rendered as the structural reference at 1440 x 900, 1024 x 768, and 390 x 844.

The observable contract requires every unique name in voicehub.__all__ to resolve and appear exactly once in a generated public reference. Every entry must record its kind, canonical repository module, source file and line, callable signature or explicit constant/type-alias marker, summary, and root- level lazy state. Duplicate, unresolved, undocumented, source-less, stale, or version-dependent inventory output must fail. A fresh root import must remain PyTorch-free.

The pre-implementation focused test failed during collection because the generator did not exist and is excluded. Two intermediate generator runs then failed while locating a facade constant and tuple assignment; those failures led to generic facade re-export and assignment-target resolution and are also excluded. Before the slice, 148 of 261 root exports were not individually named in the two hand-written API pages.

scripts/generate_public_api.py now statically validates root export/source metadata, follows direct and declarative facade re-exports, resolves canonical repository definitions, and produces docs/reference/public-api.md. Its 261 records comprise 131 classes, 73 callables, 39 enums, 11 exceptions, four constants, and three type aliases. The seven speech-domain groups contain 94 training, 81 optimization/codec, 29 configuration/factory/model, 28 inference/ serving, 20 input/output, eight policy/error/utility, and one package-metadata entry. Every row has a local source line and a usable signature or explicit marker; no entry uses an unavailable-signature fallback.

The first fresh Python 3.10 focused run exposed 39 Enum signatures whose standard-library introspection differs from Python 3.11/3.12. It reported 66 passes, two failures, and 1,741 subtests and is not a pass. Enum rows now use the stable public recovery contract (value). The corrected public-API and documentation selection passed 68 tests and 2,002 subtests independently in fresh Python 3.10.19, 3.11.15, and Python 3.12.12 environments. The final Python 3.12.12 complete suite passed 2,582 tests and 4,657 subtests with 15 skips and 35 warnings in 123.52 seconds. The skips remain unpassed: three default-runtime checks, three Triton and two CUDA-extension checks, five opt- in asset/oracle checks, and two inaccessible WeNet paths.

Proportional registry, task, optimization, model-scaffold, documentation, release, packaging, and distribution-compliance coverage passed 190 tests and 3,075 subtests. All 68 generated model pages and 59 generated notebooks remain current. The release inventory retained five benchmark records and 68 documented providers: 34 TTS, 23 ASR, and 11 VAD. The inventory remains 102 aliases, zero invalid display names, six public passes over 408 model/pass pairs, eight top-level navigation roots, ten representative page pairs, six shared component/state groups, eight contribution steps, and 261 public exports.

The final strict eleven-language build passed. The ten-route DOM validator retained all eight navigation roots, and a focused rendered check found the new page's seven tables, 261 repository-source links, active navigation, and inventory statement. The complete Playwright 1.62.0/Axe 4.12.1 matrix then passed 60 base cases, 60 screenshot signatures, 342 keyboard cases, and 4,613 focus steps. Its first run failed because the Models page's expected next footer still named Bark; this failed run is excluded, and the corrected contract names Public exports. A separate corrected comparison passed the new page in both palettes at all three mapped viewports with zero overflow and rendered the official Transformers Models main-class page at those same viewports. Three setup probes that used a missing script path, a nonexistent upstream group route, and an incorrect heading selector did not reach the comparison and are excluded.

A fresh isolated distribution check passed wheel, source-distribution, and editable installs with 68 models, 81 provenance manifests, 193 compliance files, required package data, zero runtime-dependency violations, and no eager PyTorch import. It produced a 57,194,978-byte wheel and a 55,468,189-byte source distribution.

This closes the broader object-by-object package-root public API audit. Exact- current complete Python 3.10/3.11, Linux, Windows, remote default-runtime, tagged-workflow, publisher, and publication gates remain pending. Five CUDA/ Triton checks and seven opt-in or inaccessible asset paths remain explicitly unpassed. No protected action was taken. The untracked uv.lock remained unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact-current final supported-Python macOS evidence

This evidence-only platform iteration refreshed Transformers main at commit d09f53a801f45ad73ec3510e17972024234bc0fd. Its 65-line Home, 1,576-line navigation, 500-line Modular Transformers, 30-line Trainer, and 2,906-line TrainingArguments sources retain SHA-256 values e24e0eddadccef8b15e3f25c55acb0df0f65de21a4b5bf709d119d22d0eb555b, f7d0504e36cd7c312968b549af4fe02b6ee7b3c23d8023986e6a5824680c8f3a, a52af1ee1754e6e658535f071a727fac31a057a3a848c28c18873a5ff371dd96, e7c5368c1223c2b195321468fcd1ac64c5cda52a1e44d4adf854a6473b6c9ee0, and cb55368c4d4b80633a7d289a790c5e4ab9f7ad03b527a969be7b0b0c2c432313. The official Home, documentation specification, Modular Transformers, and Trainer routes returned HTTP 200.

The observable contract requires a fresh direct installation and complete CPU-safe suite on every supported interpreter after the final generated-public- API slice. Each environment must import the checkout rather than an installed copy, pass dependency validation, retain a PyTorch-free package-root import, and avoid every lockfile synchronization path. A partial run or skip is not a pass.

uv 0.11.21 created fresh CPython 3.10.19 and 3.11.15 environments. Direct UV_TORCH_BACKEND=cpu uv pip install --python <environment> -e '.[test]' commands installed 93 and 90 compatible packages respectively without a lockfile argument or synchronization; uv pip check passed in both environments. Both resolved VoiceHub to /Users/kadirnar/Documents/voicehub/voicehub/__init__.py, and isolated root- import probes resolved all 261 public exports without importing PyTorch. The host was macOS 26.5.2 arm64 with PyTorch 2.8.0, Transformers 5.14.1, and pytest 9.1.1.

The complete Python 3.10.19 suite passed 2,582 tests and 4,657 subtests with 15 skips and 35 warnings in 229.03 seconds. The complete Python 3.11.15 suite passed the same counts in 214.94 seconds. Together with the exact-current Python 3.12.12 result of the same counts in 123.52 seconds, all three supported interpreters now pass the complete CPU-safe suite on the exact current macOS worktree.

The 15 skips reproduced independently and remain unpassed:

• Three default-runtime import checks require the explicit full-runtime CI environment.
• Three Triton kernel checks require VOICEHUB_TEST_TRITON_KERNELS=1 on a Triton CUDA host.
• Two compiled CUDA-extension checks require VOICEHUB_TEST_CUDA_EXTENSIONS=1 on a CUDA toolkit host.
• ESPNet, NeMo QuartzNet, SenseVoice, SpeechBrain, and TEN-VAD require their pinned opt-in configuration, archive, tokenizer, or ONNX oracle assets.
• The WeNet conversion checkpoint and tokenizer assets remain inaccessible.

After the complete runs, the exact public-API, documentation, release, packaging, and distribution-policy selection passed 92 tests and 2,078 subtests independently on Python 3.10.19, 3.11.15, and 3.12.12. The refreshed inventories remain 68 models (34 TTS, 23 ASR, and 11 VAD), 102 aliases, zero invalid display names, 68 current provider pages, 59 current notebooks, six public optimization passes over 408 model/pass pairs, ten representative page pairs, eight top- level navigation roots, eight contribution steps, 261 public exports, and five benchmark records.

The final strict eleven-language build completed in 36.87 seconds, and the ten- route DOM validator retained all eight ordered navigation roots. The unchanged representative shell retains its preceding complete visual evidence; this platform-only slice makes no new visual-parity claim. A fresh isolated distribution check passed wheel, source-distribution, and editable installs with 68 models, 81 provenance manifests, 193 compliance files, required package data, zero runtime-dependency violations, and no eager PyTorch import. It produced a 57,194,978-byte wheel and a 55,468,722-byte source distribution. This evidence record changed afterward, so the sizes are execution evidence rather than final tagged-artifact fingerprints.

No runtime, test, workflow, generated page, or generated notebook source changed in this evidence-only slice. Exact-current Linux, Windows, remote default-runtime, tagged-workflow, publisher, and publication gates remain pending. The five CUDA/Triton gates and seven opt-in or inaccessible asset paths remain explicitly unpassed. No protected action was taken. The untracked uv.lock remained unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact-current pinned release-asset evidence

This checkpoint iteration refreshed Transformers main at commit d09f53a801f45ad73ec3510e17972024234bc0fd; the revision-pinned reference fingerprints remain unchanged. The observable contract requires every pinned release asset to match its immutable source revision, declared byte size, and SHA-256 before its behavioral, conversion, or differential test can pass. A download failure, digest mismatch, skipped test, partial conversion, or missing oracle dependency is not a pass.

The three small Hugging Face assets passed in one exact-current command:

• ESPNet repository espnet/shinji-watanabe-librispeech_asr_train_asr_transformer_e18_raw_bpe_sp_valid.acc.best at revision bc6bbd771cec698f070640ee677a66719181f0a2 supplied the 82,131-byte config.yaml with SHA-256 16351b9bf79631d1df0a4645a858dc330c40434cf03470408c9c8fd446b6ea19; its extracted token list matched 48ec6eedbee6a22e2a9b51adeb425af3c39db23128086c015240f591601a3ea3.
• FunAudioLLM/SenseVoiceSmall revision 3847d57b6bdf2dd8875cb1508d2af43d80a16bf7 supplied the 377,341-byte tokenizer with SHA-256 aa87f86064c3730d799ddf7af3c04659151102cba548bce325cf06ba4da4e6a8; the published text/control vectors and semantic labels matched.
• speechbrain/asr-crdnn-rnnlm-librispeech revision 979a53a7a3f6c9291c02c040fd8ebfb2471cf8a3 supplied the 253,217-byte tokenizer.ckpt with SHA-256 37a6cba34cd520b33fd83612d5efc8ba7e351166541eb2726642bb3032234d31; its published encode/decode vector matched.

That focused command passed three tests with 49 deselections in 1.90 seconds. Each cached file was independently re-read and reproduced its declared URL, revision, size, and digest.

The official TEN-VAD ONNX graph came from immutable source revision 22a3bcd4509d0faaa8eef4881e8af5f39c178950. The downloaded file was 315,449 bytes with SHA-256 e10b98a0cab1c98e847fbdda14cb3d45a38336d47535a3f63a0fb6c4e0f4cdf4. On macOS 26.5.2 arm64 with Python 3.12.12, PyTorch 2.8.0, and pinned ONNX Runtime 1.22.1, the converter wrote the native Safetensors artifact, strict- loaded it, and matched ONNX Runtime across 25 recurrent steps. The focused test passed once with seven deselections in 0.91 seconds.

The official NVIDIA NGC stt_en_quartznet15x5 1.0.0rc1 archive downloaded with identity encoding at 70,993,538 bytes and SHA-256 1b9b7b87a9277e6fef164d8f99d1226f0511af154423bbf919b920421ac9602f. The exact-current converter reproduced tensor fingerprint 47c098414f58e8380868692db82cf0e4cde3b2777be1cdfd557cb7c5865ef37e; the focused test passed once with eight deselections in 1.91 seconds.

One combined command then activated all five paths simultaneously and passed five tests with 64 deselections in 2.80 seconds. Broader native-architecture, registry, task, optimization, release, packaging, and distribution-compliance coverage passed 169 tests and 1,029 subtests with all five assets active. No runtime, test, workflow, generated page, or generated notebook source changed in this evidence-only slice.

The release-policy selection passed 92 tests and 2,078 subtests. The strict eleven-language build completed in 38.12 seconds, and the ten-route DOM validator retained all eight ordered navigation roots. The unchanged rendered shell retains its preceding complete visual evidence; this checkpoint-only slice makes no new visual-parity claim. Fresh isolated wheel, source- distribution, and editable installs passed with 68 models, 81 provenance manifests, 193 compliance files, required package data, zero dependency violations, and no eager PyTorch import. They produced a 57,194,978-byte wheel and a 55,468,269-byte source distribution. This evidence record changed after the build, so those sizes are execution evidence rather than final tagged- artifact fingerprints.

This closes the five pinned opt-in release-asset/oracle gates for the exact current macOS worktree. The two WeNet asset paths remain inaccessible and are not passed. Three Triton and two compiled CUDA-extension gates remain hardware- limited. Exact-current Linux, Windows, remote default-runtime, tagged-workflow, publisher, and publication gates also remain pending. No protected action was taken, and the untracked uv.lock remained unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact-current WeNet release-asset evidence

This release slice refreshed Transformers main at commit d09f53a801f45ad73ec3510e17972024234bc0fd; all five pinned reference fingerprints remain unchanged. Its observable contract requires an accessible immutable source for the exact audited WeNet archive, independent size and SHA-256 verification before parsing, successful restricted conversion and tokenizer behavior, reproducible Package CI and tagged-workflow gates, and no weakening of the pickle trust or checkpoint-license boundaries.

The original UCloud HTTP object still returns 404, its HTTPS host still fails certificate-hostname validation, and the current WeNet download portal has an expired certificate and returns NoSuchKey for the documented route. These endpoints remain failed evidence. The public openspeech/wenet-models Hugging Face mirror at immutable revision 90acd57d17169a15d5ceab462c6e7db3bd003921 exposes gigaspeech_u2pp_conformer_exp.tar.gz with linked size 503,845,602 bytes and linked SHA-256 061ccfa51d64ebe7ea091a5a13ae31e37d9c36f4eface5c7bafc80bd4a06b26e. An independent download reproduced both values and contained the five exact audited source members. This makes the mirror byte-identical to the previously fingerprinted upstream archive; it does not make the mirror the model's upstream source or establish new checkpoint license terms.

With the downloaded archive and its extracted tokenizer members activated, the real checkpoint conversion and tokenizer selection passed two tests with 18 deselections in 6.69 seconds. The converter verified the archive and all five member digests, used PyTorch's restricted weights_only=True reader behind the existing explicit trust flag, matched the 670-tensor, 136,225,077-value namespace fingerprint, wrote native Safetensors, and strict-loaded the result. The tokenizer reproduced the 4,999-entry unit inventory and the pinned HELLO WORLD encode/decode behavior.

Runtime metadata now separates the failed original URL from the immutable mirror repository, revision, filename, URL, size, and digest. SOURCE.json records both the failed upstream routes and verified mirror without changing the undeclared checkpoint-license status. The generated WeNet page keeps a local converted-artifact quickstart and explains the trust-gated conversion. Package CI and the tagged release build now download the exact mirror revision, check its SHA-256 before extraction, and run only the two isolated WeNet gates; the source pickle remains outside VoiceHub distributions.

The focused WeNet and workflow regression passed 22 tests in 7.11 seconds. The activated WeNet, architecture, registry, task, optimization, documentation, scaffold, release, packaging, and distribution-policy suite passed 224 tests and 2,621 subtests in 17.89 seconds. The selected pre-commit sequence passed every applicable hook. A post-evidence selection passed 88 documentation, release, packaging, and distribution-policy tests plus 1,556 subtests; all public-API, model-page, notebook, benchmark, version, and provenance inventories remained current. Both workflow files parsed successfully as YAML.

The strict eleven-language build completed in 38.98 seconds, and the ten-route DOM validator retained all eight ordered navigation roots. The unchanged representative shell retains its preceding complete visual evidence; this release-asset slice makes no new visual-parity claim. Fresh isolated wheel, source-distribution, and editable installs passed with 68 models, 81 provenance manifests, 193 compliance files, required package data, zero dependency violations, and no eager PyTorch import. They produced a 57,195,308-byte wheel and a 55,470,162-byte source distribution. This evidence record changed after the build, so those sizes are execution evidence rather than final tagged- artifact fingerprints.

This closes both previously inaccessible WeNet asset paths on the exact current macOS worktree. Three Triton and two compiled CUDA-extension paths remain hardware-limited and unpassed. Exact-current Linux, Windows, remote default- runtime, tagged-workflow, publisher, and publication gates remain pending. No protected action was taken, and the untracked uv.lock remained unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact-current full-dependency runtime refresh

This release-evidence iteration refreshed Transformers main at commit d09f53a801f45ad73ec3510e17972024234bc0fd; the five pinned reference fingerprints remain unchanged. The observable gate requires a fresh Python 3.12 environment installed directly from this worktree with all declared test, training, and documentation extras, a clean dependency check, all focused default-runtime tests executed rather than skipped, and a complete activated suite. It does not transfer a macOS result to another interpreter or platform.

The repository .venv is not release evidence: uv pip check found an unrelated installed trainer distribution that declares Python <3.12, so the first command stopped before executing any tests. That failed command is not a pass, and the existing environment was preserved without modification.

uv 0.11.21 created a fresh CPython 3.12.12 environment and installed 135 compatible packages directly from the current checkout through its CPU PyTorch resolver without synchronizing or reading the repository lock file. uv pip check passed. VoiceHub resolved to this checkout on macOS 26.5.2 arm64 with PyTorch 2.8.0, Transformers 5.14.1, and pytest 9.1.1.

With VOICEHUB_FULL_RUNTIME_TEST=1, the focused default-runtime file passed all five tests and 138 registry subtests in 9.34 seconds. The same environment then passed the complete activated suite: 2,585 tests, 4,795 subtests, 12 skips, and 35 warnings in 181.72 seconds. The three default-runtime tests that the normal suite skips all executed and passed.

The 12 residual skips remain separately classified and are not counted by this command:

• Three Triton and two compiled CUDA-extension checks require unavailable CUDA hardware or a CUDA toolkit and remain unpassed.
• ESPNet, NeMo QuartzNet, SenseVoice, SpeechBrain, TEN-VAD, and the two WeNet checks require their opt-in assets or oracle environment. All seven have separate exact-current passing evidence in the immediately preceding release sections, but they were not executed in this full-runtime command.

The post-evidence documentation, release, packaging, and distribution-policy selection passed 88 tests and 1,556 subtests. All public-API, model-page, notebook, benchmark, version, and provenance inventories remained current, and the selected pre-commit sequence passed. The strict eleven-language build completed in 39.99 seconds, and the ten-route DOM validator retained all eight ordered navigation roots. This evidence-only slice changes no rendered component and makes no new visual-parity claim.

Fresh isolated wheel, source-distribution, and editable installs passed with 68 models, 81 provenance manifests, 193 compliance files, required package data, zero dependency violations, and no eager PyTorch import. They produced a 57,195,308-byte wheel and a 55,471,307-byte source distribution. This evidence record changed after the build, so those sizes are execution evidence rather than final tagged-artifact fingerprints.

This closes the exact-current local Python 3.12 full-dependency/default-runtime gate after the WeNet slice. It does not refresh Python 3.10 or 3.11 after those source/test changes and does not substitute for exact-current Linux, Windows, or remote default-runtime CI. Tagged-workflow artifacts, publisher configuration, publication approval, and all five hardware gates remain open. No protected action was taken, and the untracked uv.lock remained unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact-current final supported-Python refresh

This release-evidence iteration refreshed Transformers main at commit d09f53a801f45ad73ec3510e17972024234bc0fd; all five pinned reference fingerprints remain unchanged. The observable gate requires fresh Python 3.10 and 3.11 environments to install this exact worktree directly, pass dependency validation and the changed WeNet/workflow/documentation contracts, keep the package-root import PyTorch-free, and complete the CPU-safe suite. A result from the source state before the WeNet slice is not transferred.

uv 0.11.21 created CPython 3.10.19 and 3.11.15 environments and installed 93 and 90 compatible packages respectively through its CPU PyTorch resolver. Both installs used direct uv pip install commands without synchronization or a lockfile argument, and both passed uv pip check. Each resolved VoiceHub to this checkout and resolved all 261 package-root exports without importing PyTorch. The host was macOS 26.5.2 arm64 with PyTorch 2.8.0, Transformers 5.14.1, and pytest 9.1.1.

The focused WeNet metadata, Package CI, tagged-workflow, checkpoint- documentation, and packaging contracts passed 15 tests and 76 subtests on each interpreter: Python 3.10 in 18.03 seconds and Python 3.11 in 17.83 seconds. The complete Python 3.10 suite then passed 2,582 tests and 4,657 subtests with 15 skips and 35 warnings in 226.26 seconds. Python 3.11 passed the same counts in 212.37 seconds.

After this evidence update, the public-API, documentation, release, packaging, and distribution-policy selection passed 92 tests and 2,078 subtests on both interpreters: Python 3.10 in 16.64 seconds and Python 3.11 in 15.65 seconds. All 261 exports, 68 model pages, 59 notebooks, five benchmark records, version metadata, and source-provenance inventories remained current.

The selected pre-commit sequence passed. The strict eleven-language documentation build completed in 38.06 seconds, and the DOM validator retained all ten representative routes and eight ordered navigation roots. Fresh isolated wheel, source-distribution, and editable installs passed with 68 models, 81 provenance manifests, 193 compliance files, all required package data, zero dependency violations, and no eager PyTorch import. They produced a 57,195,308-byte wheel and a 55,469,928-byte source distribution. This report changed after the build, so those sizes are execution evidence rather than final tagged-artifact fingerprints.

The 15 skips remain explicit and are not counted by either complete command:

• Three complete-dependency default-runtime imports require VOICEHUB_FULL_RUNTIME_TEST=1; the exact-current Python 3.12 runtime section above executes and passes them separately.
• Three Triton and two compiled CUDA-extension paths require unavailable CUDA hardware or a CUDA toolkit and remain unpassed.
• ESPNet, NeMo QuartzNet, SenseVoice, SpeechBrain, TEN-VAD, and the two WeNet paths require their opt-in assets or oracle variables. All seven retain separate exact-current passing evidence, but neither supported-version command executed them.

Together with the exact-current Python 3.12.12 full-dependency result, the current macOS worktree now passes the complete suite on every supported Python version. This does not substitute for exact-current remote Linux, macOS, or Windows execution. Tagged-workflow artifacts, publisher configuration, publication approval, and all five hardware gates remain open. No protected action was taken, and the untracked uv.lock remained unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Exact-current repository-wide lint refresh

This release-evidence iteration retained Transformers main commit d09f53a801f45ad73ec3510e17972024234bc0fd and the five pinned reference fingerprints. Its observable gate is stricter than pre-commit run --all-files on a dirty candidate: every tracked file and every candidate untracked file must enter one configured-hook invocation, while the protected untracked uv.lock must not enter the argument set or change.

The resulting candidate set contained 4,675 files after explicitly excluding uv.lock. End-of-file, trailing-whitespace, case-conflict, private-key, AWS-credential, pyupgrade, isort, YAPF, Markdown-formatting, Flake8, and docformatter hooks all passed. SHA-256 manifests captured before and after the complete invocation were identical, so no hook rewrote a checked file. An initial audit-wrapper attempt stopped before invoking pre-commit because its NUL-delimited file list retained an empty sentinel; it changed no repository file and is not counted as lint evidence.

The post-lint registry, public-API, documentation, optimization, scaffold, release, and packaging selection passed 157 tests and 2,797 subtests in 20.25 seconds before the evidence edit and again in 21.62 seconds afterward. All 261 exports, 68 model pages, 59 notebooks, five benchmark records, version metadata, and source-provenance inventories remained current. The strict eleven-language build completed in 39.31 seconds, and the DOM validator retained all ten representative routes and eight ordered navigation roots. Fresh isolated wheel, source-distribution, and editable installs passed with 68 models, 81 provenance manifests, 193 compliance files, all required package data, zero dependency violations, and no eager PyTorch import. They produced a 57,195,308-byte wheel and a 55,471,136-byte source distribution. This report changed after the build, so those sizes are execution evidence rather than final tagged-artifact fingerprints. The slice changes no runtime or rendered component and makes no new visual-parity claim.

This closes the exact-current local repository-wide formatting-and-lint gate. It does not substitute for exact-current remote lint or Linux, macOS, and Windows execution. Tagged-workflow artifacts, publisher configuration, publication approval, and the five hardware gates remain open. No protected action was taken, and the untracked uv.lock remained unchanged at SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

Pull-request release-candidate verification

This final candidate iteration retrieved Transformers main at commit ff2421c67f35cc83a0fbabbc2633c96734685918 on 2026-08-04; the official Transformers documentation route returned HTTP 200. VoiceHub implementation commit aead0611b9eafa0e20d32900568c063073976741 is the reviewed head of pull request 73 on codex/voicehub-0.3-rc. The worktree contained only the protected untracked uv.lock, which retained SHA-256 48f7d98d6eab756580348b081e8fc891d3a5dd2847433e41766b3b45854a70b1.

The local focused Quickstart source regression passed two tests, three subtests, and 62 deselections. The complete documentation source suite passed 64 tests and 1,480 subtests. The final Playwright 1.62.0/Axe 4.12.1 run passed 60 rendered cases, 60 screenshot signatures, 60 accessibility cases, 342 keyboard cases, and 4,591 focus steps across ten representative routes, three viewports, and two palettes. Every applicable pre-commit hook passed for the two changed checker files. Two preceding complete visual commands timed out in the Quickstart interaction after exposing state left by the preceding full focus cycle; they are failed diagnostic runs and are not counted. The passing checker reloads the route before its independent content-tab interaction, uses native radio-group ArrowRight activation, waits for focus and selection to settle, and retains a one-CSS-pixel fractional viewport tolerance.

Exact-head GitHub Actions then passed every required pull-request job:

• Documentation run 30875378385 passed in 45 minutes 34 seconds. Its Python 3.12.13 Linux job passed 64 documentation-source tests, the strict eleven-language build in 74.81 seconds, the ten-route DOM validator with all eight ordered navigation roots, 60 Linux screenshot signatures, 60 accessibility cases, 342 keyboard cases, and 4,587 focus steps. The Pages deployment job was skipped on the pull request and is not counted as passed.
• Package run 30875378415 passed in 2 minutes 43 seconds. Wheel, source-distribution, and editable validation each found 68 models, 81 provenance manifests, 193 compliance files, all required package data, zero runtime-dependency violations, and no eager PyTorch import. The exact artifacts were a 57,195,308-byte wheel and a 55,457,609-byte source distribution. All seven pinned asset/oracle gates, including both WeNet paths, executed and passed separately in this run.
• Continuous Integration run 30875378456 passed lint, training, full-dependency runtime, both cross-platform runtime smokes, and the full suite on Python 3.10, 3.11, and 3.12 on Ubuntu, macOS, and Windows. Every ordinary matrix job executed 2,581 tests and 4,657 subtests with 16 explicit skips. The full-dependency job executed 2,584 tests and 4,795 subtests with 13 explicit skips. The separate pre-commit.ci check also passed on the exact head.

The refreshed generated inventories remain 68 models (34 TTS, 23 ASR, and 11 VAD), 102 aliases, zero invalid display names, 68 model pages, 59 model notebooks, six public optimizations over 408 model/pass pairs, ten representative page pairs, eight top-level navigation roots, eight contribution steps, 261 public exports, and five benchmark records. The direct model-page and model-notebook generator checks passed. A bare local public-API generator probe lacked the installed VoiceHub/PyTorch runtime and failed before validation; it is not a pass and is superseded by the exact-head full-runtime, package, and complete-suite jobs above.

Five native-kernel paths remain explicitly unpassed: three Triton tests require VOICEHUB_TEST_TRITON_KERNELS=1 on a Triton CUDA host, and two compiled CUDA-extension tests require VOICEHUB_TEST_CUDA_EXTENSIONS=1 on a CUDA toolkit host. No tag or tagged-release workflow was created, the pull request was not merged, publisher configuration was not changed, and no GitHub release or PyPI publication was attempted. Those maintainer-controlled actions remain outside release-candidate verification and require explicit approval.

One-time publisher configuration

Before the first 0.3 publication:

1. In PyPI project settings, add a GitHub Trusted Publisher for repository kadirnar/voicehub, workflow release.yml, and environment pypi.
2. In GitHub, create the pypi environment and require a maintainer reviewer.
3. Do not add a long-lived PyPI API token. Only the publish job receives id-token: write.

PyPI's default per-file limit is 100 MB. scripts/check_release.py --dist-dir rejects oversized or unexpected files before the protected publish job begins.

Publish and post-publish verification

After every local and cross-platform gate is green, create the signed tag and manually dispatch the release workflow with its publish confirmation enabled. Approve the protected pypi environment only after reviewing the build job and artifact hashes.

When PyPI finishes indexing the release, verify external parity:

python scripts/check_release.py \
  --tag v0.3.0 \
  --require-tag-at-head \
  --pypi-policy published

Only then create or finalize the matching GitHub release. If any external gate fails, leave the candidate unpublished and record the exact blocker here.