feat(01-09): add LICENSES.md, prove the credits and the VRM meta on the Space, record the SC-4 run

LICENSES.md (DPLY-04) merges docs/ASSETS.md and docs/VOICEVOX-SETUP.md into one table
row per shipped or runtime-loaded asset - the VRM, VOICEVOX CORE, the VOICEVOX ONNX
Runtime, the voice model, the zundamon character, the Open JTalk dictionary, three.js,
@pixiv /three-vrm, @huggingface/transformers and the Whisper weights - with every cell
filled, the permission each one is used under, the exact credit strings, the model-terms
clause the committed .vvm relies on quoted verbatim, the character terms' fee and
prohibitions, the Nemo swap-in, the NAIST notice reproduced, and a "not used" section
stating that AvatarSample_A/B/C are not CC0. Two facts corrected against primary sources
while writing it: voicevox_core 0.17.0 is MIT (the LGPL/dual licence applied to prebuilt
cores below 0.16), and the whisper-base WEIGHTS are Apache-2.0 per the model card while
the openai/whisper code is MIT.

Two machine-readable blocks make the document and the shipped assets unable to disagree:
test_vrm_meta_matches_licenses reads the deployed page's vrmMeta and compares it with the
vrm_meta block through one normalising helper (VRM 0.0 and 1.0 key names) - 14 fields,
0 mismatches; test_credits_visible reads the required credit strings from the credits
block and asserts them in the footer, the flow-down notice (72 px under Replay, both
terms linked), the About panel and the server-delivered GET /config, with no interaction.
Gradio 6 serves an SSR shell, so the root HTML carries 0 hits by design (docs/HOSTING.md
finding 5); the server-delivered assertion is on /config.

SC-4: the entire tests/e2e/test_avatar_loop.py ran against Space revision c911d74 with
DISABLE_GPU=1 read back from the Space variables in the same script - 15 passed, 0
failed. Recorded under docs/HOSTING.md "SC-4 run record" with the latency numbers the
run produced for plan 01-10.

LICENSES.md

@@ -0,0 +1,232 @@
+# Third-party licences (DPLY-04)
+
+Every third-party asset this project ships, fetches at runtime, or loads into a visitor's
+browser, with the permission each one is used under and how that permission is honoured.
+
+This document **merges** `docs/ASSETS.md` (the VRM, plan 01-01) and `docs/VOICEVOX-SETUP.md`
+(the three VOICEVOX licence layers and the dictionary, plan 01-04); it does not re-research
+either. Every runtime-library licence below was re-verified on 2026-09-06 against the package
+registry and the upstream repository, and the VRM's row is verified automatically:
+`tests/e2e/test_avatar_loop.py::test_vrm_meta_matches_licenses` reads the shipped file's own
+embedded `VRMC_vrm.meta` on the deployed Space and compares it with the `vrm_meta` block in
+this file, so the document and the binary cannot silently disagree.
+
+The repository's own code is MIT (see the Space `README.md` front-matter). Nothing below is
+covered by that licence.
+
+## Summary
+
+| Asset | Version | Author / rights holder | Source URL | Licence / terms URL | Permission relied on | Required credit string | How we comply |
+|---|---|---|---|---|---|---|---|
+| **VRM character** `avatar/assets/tutor.vrm` | `VRM1_Constraint_Twist_Sample` v1.0.1, VRM 1.0 | pixiv Inc. — `(c) 2022 pixiv Inc.` | https://github.com/vrm-c/vrm-specification/tree/master/samples/VRM1_Constraint_Twist_Sample | https://vrm.dev/licenses/1.0/ (VRM Public License 1.0, with per-file flags embedded in the file) | Embedded `VRMC_vrm.meta`: `allowRedistribution: true`, `avatarPermission: everyone`, `modification: allowModificationRedistribution`, `commercialUsage: corporation` | **none** (`creditNotation: unnecessary`); credited voluntarily as `VRM1_Constraint_Twist_Sample (c) 2022 pixiv Inc. - VRM Public License 1.0` | Committed through Git LFS unmodified; credit rendered in the footer, the About panel and the README; the one restriction (`allowAntisocialOrHateUsage: false`) is stated below; sourcing gate re-raised in plan 01-10 |
+| **VOICEVOX CORE** (software) | `voicevox_core` 0.17.0 (Python wheel, abi3) | Hiroshiba Kazuyuki (ヒホ) / the VOICEVOX project | https://github.com/VOICEVOX/voicevox_core/releases/tag/0.17.0 | https://voicevox.hiroshiba.jp/term/ (VOICEVOX ソフトウェア利用規約); code and build artefacts MIT — https://github.com/VOICEVOX/voicevox_core/blob/0.17.0/LICENSE | Software terms: commercial and non-commercial use permitted; generated audio may be provided onward **subject to the flow-down obligation of clause 3**. MIT for the linked code | `VOICEVOX:<character>` — satisfied by the character row below; plus the terms notice wherever audio is obtainable | Installed from the official release URL pinned in `requirements.txt`; the wheel is **never vendored** (禁止事項 forbids unauthorised redistribution of the software); flow-down notice rendered next to the Replay control |
+| **VOICEVOX ONNX Runtime** | `voicevox_onnxruntime` 1.23.2 | VOICEVOX project build of Microsoft ONNX Runtime | https://github.com/VOICEVOX/onnxruntime-builder/releases/tag/voicevox_onnxruntime-1.23.2 | MIT (both `VOICEVOX/onnxruntime-builder` and `microsoft/onnxruntime`) | MIT | none | Fetched lazily at runtime into the gitignored `voicevox_runtime/` from the official release; never committed or redistributed (`docs/VOICEVOX-SETUP.md` § ONNX Runtime acquisition) |
+| **VOICEVOX voice model** `voicevox/model/zundamon.vvm` | `voicevox_vvm` 0.17.0, file `0.vvm` (SHA-256 `ecd35374d4182cd883cba5040376f7f888cc6ba248b1c2f4cea07cdb34bb1318`) | the VOICEVOX project (model file); voices inside it belong to their character licensors | https://github.com/VOICEVOX/voicevox_vvm/releases/tag/0.17.0 | VOICEVOX 音声モデル 利用規約 — https://github.com/VOICEVOX/voicevox_vvm/blob/main/README.md (shipped as `README.txt` / `TERMS.txt` with the release) | Clause 2: 「アプリケーションに組み込んで再配布することができます」 — embedded redistribution is explicitly permitted; clause 4 flow-down; a credit that shows VOICEVOX was used is required | `VOICEVOX:ずんだもん` (only ずんだもん, style 3, is used; the other voices packaged in `0.vvm` are neither used nor credited) | Committed through Git LFS under clause 2; credit and flow-down notice rendered; the model is used only through VOICEVOX CORE |
+| **ずんだもん** (character voice) | — (voice library shipped inside `0.vvm`) | SSS LLC (合同会社SSS) | https://zunko.jp/ | https://zunko.jp/con_ongen_kiyaku.html (音源利用規約, one document covering nine characters) | Commercial and non-commercial use of generated audio permitted **with credit**; app use requires the credit on an introduction/about screen | **`VOICEVOX:ずんだもん`** — exact, ASCII colon, no spaces | Persistent footer credit + About panel on first paint (server-delivered, not injected); flow-down notice linking both terms URLs; no prohibited-use path exists (see below) |
+| **Open JTalk dictionary** `voicevox/open_jtalk_dic_utf_8-1.11/` | `open_jtalk_dic_utf_8-1.11` | © 2009 **Nara Institute of Science and Technology** (NAIST), Japan | https://downloads.sourceforge.net/project/open-jtalk/Dictionary/open_jtalk_dic-1.11/open_jtalk_dic_utf_8-1.11.tar.gz | BSD-3-Clause — `voicevox/open_jtalk_dic_utf_8-1.11/COPYING` (reproduced below) | BSD-3-Clause redistribution in binary form, with the notice reproduced | none (attribution notice) | `COPYING` is committed alongside the dictionary files; the notice is reproduced in this file; the About panel names NAIST. Open JTalk itself (Nagoya Institute of Technology / HTS Working Group, modified BSD) is linked inside VOICEVOX CORE and is attributed too |
+| **three.js** | 0.185.1 | three.js authors (mrdoob et al.) | https://github.com/mrdoob/three.js | MIT — https://github.com/mrdoob/three.js/blob/master/LICENSE | MIT | none | Loaded in the visitor's browser from `https://esm.sh/three@0.185.1` today; plan 01-10 vendors it under `avatar/vendor/` with its LICENSE file |
+| **@pixiv/three-vrm** | 3.5.5 | pixiv Inc. | https://github.com/pixiv/three-vrm | MIT — https://github.com/pixiv/three-vrm/blob/dev/LICENSE | MIT | none | Loaded from `https://esm.sh/@pixiv/three-vrm@3.5.5?deps=three@0.185.1` today; vendored in plan 01-10 with its LICENSE file |
+| **@huggingface/transformers** (transformers.js) | 4.2.0 | Hugging Face | https://github.com/huggingface/transformers.js | Apache-2.0 — https://github.com/huggingface/transformers.js/blob/main/LICENSE | Apache-2.0 | none (NOTICE reproduction if vendored) | Loaded from `https://esm.sh/@huggingface/transformers@4.2.0` today; vendored in plan 01-10 with LICENSE (and NOTICE, if the package ships one) |
+| **Whisper ASR model** `onnx-community/whisper-base` (q4 ONNX) | `onnx-community/whisper-base`, converted from `openai/whisper-base` | OpenAI (model); onnx-community (ONNX conversion) | https://huggingface.co/onnx-community/whisper-base | Weights: Apache-2.0 per the `openai/whisper-base` model card (`license: apache-2.0`; the conversion card declares `base_model: openai/whisper-base` and no licence of its own). Source code of `openai/whisper`: MIT | Apache-2.0 use of the weights in the visitor's browser | none (attribution) | **Not in this repository**: the visitor's browser downloads it from huggingface.co on the first accepted push (135.8 MB); attributed here and in the About panel ("Whisper"). The "better accuracy" model `whisper-large-v3-turbo` is not offered yet |
+
+Nine rows for eight asset classes: the ONNX Runtime is listed separately from VOICEVOX CORE
+because it is fetched by a different mechanism and under a different licence.
+
+## Machine-readable blocks
+
+The deployed suite parses these. Keep the keys and the values exactly as the shipped assets
+and the rendered page have them.
+
+The VRM's embedded `VRMC_vrm.meta`, as read out of `avatar/assets/tutor.vrm` itself (VRM
+1.0 key names; `thumbnailImage` is an image index and is omitted):
+
+```vrm_meta
+name: VRM1_Constraint_Twist_Sample
+version: v1.0.1
+authors: pixiv Inc.
+copyrightInformation: (c) 2022 pixiv Inc.
+licenseUrl: https://vrm.dev/licenses/1.0/
+avatarPermission: everyone
+commercialUsage: corporation
+creditNotation: unnecessary
+allowRedistribution: true
+modification: allowModificationRedistribution
+allowAntisocialOrHateUsage: false
+allowExcessivelySexualUsage: true
+allowExcessivelyViolentUsage: true
+allowPoliticalOrReligiousUsage: true
+```
+
+The credit strings the page must render (`tests/e2e/test_avatar_loop.py::test_credits_visible`
+checks the footer, the About panel and the server-delivered config for them):
+
+```credits
+voice: VOICEVOX:ずんだもん
+avatar: VRM1_Constraint_Twist_Sample (c) 2022 pixiv Inc. - VRM Public License 1.0
+```
+
+## VRM character — `avatar/assets/tutor.vrm`
+
+> **TEMPORARY DEV ASSET.** Legally clean as recorded here, but it ships a spec sample's name.
+> Plan 01-10 must re-raise the sourcing checkpoint (keep it, or author a bespoke VRoid Studio
+> character with a deliberately set `vrm.meta`) before the phase-exit licensing audit. See
+> `docs/ASSETS.md` § Open item for plan 01-10.
+
+VRM Public License 1.0 is a **template, not a fixed grant**: redistribution, avatar use,
+modification and commercial use are per-file flags set by the licensor inside the model's own
+`VRMC_vrm.meta`. Two files citing the same licence URL can grant materially different rights.
+The facts above were therefore parsed out of this file's glTF JSON chunk (10,776,032 bytes,
+SHA-256 `12c2b97e95e700783a6a550dc0eee2d7880aeedccef9ae67bc4c5a2f0f2631a2`), never off a README.
+
+Permissions relied on, and what each one covers here:
+
+| Flag | Value | What it licenses in this project |
+|---|---|---|
+| `allowRedistribution` | `true` | Committing the file to this public repository and serving it from the Space |
+| `avatarPermission` | `everyone` | Any visitor uses it as the tutor avatar |
+| `modification` | `allowModificationRedistribution` | The rest pose applied at mount, and any future retargeting, still redistributable |
+| `commercialUsage` | `corporation` | The broadest tier; a public portfolio Space is covered unambiguously |
+| `creditNotation` | `unnecessary` | No credit is required. One is rendered anyway |
+| `allowAntisocialOrHateUsage` | **`false`** | The one restriction. A Japanese-language tutor has no such use path; stated so it is not silently dropped |
+
+## VOICEVOX — three licence layers, read separately
+
+The voice output involves three rights holders, and their terms differ in one place that looks
+like a contradiction until both are read.
+
+**1. Software** — VOICEVOX ソフトウェア利用規約, https://voicevox.hiroshiba.jp/term/. Commercial
+and non-commercial use permitted. Clause 3 attaches a **flow-down obligation**: when audio
+generated here is made available to others, those others must be bound to the same terms.
+禁止事項 forbids redistributing the software in whole or in part without authorisation, so the
+`voicevox_core` wheel and the ONNX Runtime binary are always **referenced** from their official
+release URLs and never copied into this repository. Separately from the terms of use, the
+VOICEVOX CORE **code and build artefacts** are MIT-licensed as of 0.17.0 (`LICENSE`, © 2021
+Hiroshiba Kazuyuki; README § ライセンス). *Correction to earlier project notes:* prebuilt cores
+below version 0.16 were under a different (LGPL v3 / commercial dual) licence, and that is the
+version some of this project's planning text still describes; the shipped 0.17.0 is MIT.
+
+**2. Voice model** — VOICEVOX 音声モデル 利用規約, in the `voicevox_vvm` README and shipped as
+`README.txt` / `TERMS.txt` with the release. Clause 2 reads
+「アプリケーションに組み込んで再配布することができます」 — **embedded redistribution is explicitly
+permitted**, which is why `voicevox/model/zundamon.vvm` is committed (through Git LFS) while the
+software is not. That asymmetry is deliberate on VOICEVOX's side and on ours: the software terms
+forbid redistributing the *software*; the model terms permit redistributing the *model* inside an
+application. Clause 4 carries the same flow-down obligation, and a credit showing that VOICEVOX
+was used is required.
+
+**3. Character** — ずんだもん, SSS LLC, https://zunko.jp/con_ongen_kiyaku.html. Audio generated
+with the ずんだもん voice library may be used commercially and non-commercially **provided the
+credit `VOICEVOX:ずんだもん` is displayed**. Without the credit, per-character licensing is
+**¥400,000 (+ tax) per character** — which is why the credit is not optional and why it is
+rendered on first paint rather than behind a click. Placement (clause 2): for app use, on an
+introduction/about screen, somewhere findable with a little looking
+(「アプリの紹介画面などに記載をお願いします。（少し探せばわかる場所に）」); Phase 1 renders it in a
+persistent footer beside the avatar **and** in the About panel, open on first load.
+
+Prohibited uses under the character terms, recorded so later phases can check against them:
+公序良俗に反する利用; 政治的・宗教的活動; 情報商材; **意図的な**虚偽情報の作成・拡散; 風俗営業;
+反社会的勢力による利用. The falsehood clause is scoped to *intentional* creation or spread —
+relevant once an LLM produces the avatar's lines (Phase 3): a tutor that is sometimes wrong is not
+in scope of it; a product designed to mislead would be.
+
+There is no escalation path and none should be sought: SSS LLC's 免責条項 2 states that they do
+not, as a rule, answer "does my use qualify" questions
+(「本ガイドラインに該当するかどうかのご質問については、原則お答えしておりません。」). The posture is:
+read the guideline, comply visibly, document here.
+
+The credit obligation survives an engine swap: the official VOICEVOX Q&A answers that using the
+intermediate `AudioQuery` with another synthesiser still requires the credit
+(「必要です。」). The viseme timeline in this project is built from the `AudioQuery`, so the credit
+is owed on that basis alone.
+
+**Pre-cleared swap-in: VOICEVOX Nemo.** If the character terms ever become inconvenient,
+`n0.vvm` from the same `voicevox_vvm` 0.17.0 release (nine voices, style IDs 10000–10008) is a
+two-line change. Credit is just `VOICEVOX Nemo`; the terms live entirely in the `voicevox_vvm`
+README (https://voicevox.hiroshiba.jp/nemo/term/); there is no third-party rights holder and no
+¥400,000 credit-omission clause. The cost is the loss of a recognisable character.
+
+## Open JTalk dictionary — `voicevox/open_jtalk_dic_utf_8-1.11/`
+
+BSD-3-Clause. The shipped `COPYING` reads, verbatim:
+
+> Copyright (c) 2009, Nara Institute of Science and Technology, Japan.
+> All rights reserved.
+>
+> Redistribution and use in source and binary forms, with or without modification, are permitted
+> provided that the following conditions are met: Redistributions of source code must retain the
+> above copyright notice, this list of conditions and the following disclaimer. Redistributions in
+> binary form must reproduce the above copyright notice, this list of conditions and the following
+> disclaimer in the documentation and/or other materials provided with the distribution. Neither
+> the name of the Nara Institute of Science and Technology (NAIST) nor the names of its
+> contributors may be used to endorse or promote products derived from this software without
+> specific prior written permission.
+>
+> THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
+> IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+> FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
+> CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+> CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+> SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+> THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
+> OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+> POSSIBILITY OF SUCH DAMAGE.
+
+Two institutions are genuinely involved, and earlier research conflated them:
+
+- The **dictionary data** committed here (`open_jtalk_dic_utf_8-1.11`, NAIST Japanese Dictionary /
+  IPAdic lineage) is © 2009 **Nara Institute of Science and Technology** — the notice above is the
+  one that must be reproduced for the files actually redistributed in this repository.
+- **Open JTalk itself** — the engine whose dictionary format this is, linked inside
+  `voicevox_core` — is by the **Nagoya Institute of Technology** and the **HTS Working Group**,
+  under a modified BSD licence. It is not redistributed by this repository; it is attributed
+  because the About panel names the engine.
+
+## Runtime libraries loaded into the visitor's browser
+
+| Library | Version | Licence | Copyright | Loaded from (today) |
+|---|---|---|---|---|
+| three.js | 0.185.1 | MIT | © 2010-2026 three.js authors | `https://esm.sh/three@0.185.1` (+ `examples/jsm/loaders/GLTFLoader.js`) |
+| @pixiv/three-vrm | 3.5.5 | MIT | © 2019-2026 pixiv Inc. | `https://esm.sh/@pixiv/three-vrm@3.5.5?deps=three@0.185.1` |
+| @huggingface/transformers | 4.2.0 | Apache-2.0 | © Hugging Face | `https://esm.sh/@huggingface/transformers@4.2.0` |
+
+All three are fetched by the visitor's browser from esm.sh at runtime in this revision; **plan
+01-10 vendors them** under `avatar/vendor/` with each package's LICENSE (and NOTICE where one
+exists) alongside, at which point this table's last column changes to the vendored path. MIT and
+Apache-2.0 both permit that redistribution with the notice preserved.
+
+The ASR model (`onnx-community/whisper-base`, q4) is downloaded by the browser from
+huggingface.co on the first accepted push-to-talk utterance and cached by the browser; it is
+never served from this repository or the Space. `docs/ASR-TIERS.md` records why this model and
+this quantisation.
+
+## Not used, and why
+
+- **`AvatarSample_A`, `AvatarSample_B` and `AvatarSample_C` are NOT CC0.** Copyright is retained
+  and VRoid Hub's conditions apply; VRoid's own FAQ prohibits redistributing them under a CC0
+  designation. The claim is repeated confidently across tutorials and blog posts, and it is
+  false. Recorded here so it is not reintroduced into this repository.
+- CC0 **cannot be set** on VRoid Hub at all, so any "CC0 on VRoid Hub" claim is false by
+  construction; and enabling "Allow third-party usage" on VRoid Hub **overwrites** a model's
+  embedded licence with the Hub's own conditions, so a Hub-sourced file's embedded flags may not
+  say what its author intended. Both are reasons the VRM row above reads the flags out of the
+  shipped file.
+- **Hosted ASR on accelerated hardware (tier D)** is deliberately not built: it would burn the
+  visitor's ZeroGPU quota, contradicting SC-4. `tests/test_transport_seam.py` fails if one appears.
+- **Style-Bert-VITS2** (AGPL-3.0), **pykakasi** (GPL-3.0) and **edge-tts** (undocumented
+  endpoint) were rejected at stack selection for licence reasons and do not appear anywhere in
+  the tree. Python dependencies that are not assets (`gradio`, `spaces`, `huggingface_hub`,
+  `numpy`) are ordinary Apache-2.0 / BSD packages installed from PyPI and are not redistributed.
+
+## Verification record
+
+| Check | Result (2026-09-06) |
+|---|---|
+| VRM `VRMC_vrm.meta` parsed from the shipped file | matches the `vrm_meta` block above, key for key |
+| `voicevox_core` 0.17.0 licence | README § ライセンス at tag 0.17.0: MIT; `LICENSE` © 2021 Hiroshiba Kazuyuki; GitHub API `MIT` |
+| `voicevox_vvm` clause 2 | present verbatim in the `main` README (line 21) |
+| `VOICEVOX/onnxruntime-builder`, `microsoft/onnxruntime` | GitHub API: MIT, MIT |
+| `three@0.185.1`, `@pixiv/three-vrm@3.5.5`, `@huggingface/transformers@4.2.0` | npm registry: MIT, MIT, Apache-2.0 (all three are also the current `latest`) |
+| `openai/whisper-base` model card | HF API `cardData.license = apache-2.0`; `openai/whisper` repo MIT |
+| `onnx-community/whisper-base` card | no licence field; `base_model: openai/whisper-base` |
+| Open JTalk `COPYING` | first line: `Copyright (c) 2009, Nara Institute of Science and Technology, Japan.` |
+| Terms URLs answer | voicevox.hiroshiba.jp/term/ 200 · zunko.jp/con_ongen_kiyaku.html 200 · voicevox.hiroshiba.jp/nemo/term/ 200 · vrm.dev/licenses/1.0/ 200 |

docs/HOSTING.md

@@ -149,6 +149,34 @@ python -c "from huggingface_hub import HfApi; print(HfApi().get_space_variables(
 `AVATAR_TRANSPORT` is deliberately **not** set, so the Space uses the component's default,
 `inline` — the transport the spike is testing. Setting it to `iframe` is the entire fallback.
 
+### SC-4 run record — the whole loop with the GPU disabled (plan 01-09)
+
+The platform forces one `@spaces.GPU` function to exist (`app.py::zerogpu_probe`, finding 4
+below), so "no GPU function exists" is not an available proof. The honest proof is that the
+whole turn loop completes without ever reaching one — and reaching it would raise, because it
+checks `DISABLE_GPU` on every call.
+
+| | |
+|---|---|
+| Date | 2026-09-06, 03:27–03:32 UTC |
+| Space revision | `c911d74b378981e6ba0d0111eb570ed3b20927db` (plan 01-08 code, deployed by the owner) |
+| Runtime at run time | `RUNNING`, hardware `zero-a10g`, read from `space_info()` in the same script |
+| `DISABLE_GPU` | **`1`**, read back from `get_space_variables()` in the same script (set 2026-09-05 by plan 01-05) |
+| Command | `pytest tests/e2e/test_avatar_loop.py -q --space-url https://wolfdavid-japanese-learning-avatar.hf.space` |
+| Result | **15 passed, 0 failed** (14 test functions; `test_silence_rejected` runs twice) — exit 0 |
+| Turns that completed | typed turn ×5 (incl. 20 wave-5 interactions in `test_no_remount`), push-to-talk ×2 (WASM tier), replay, slower re-synthesis; silence and cafe noise produced 0 turns |
+
+Every row in `01-VALIDATION.md`'s Per-Task Verification Map that names a
+`tests/e2e/test_avatar_loop.py` node was green in this run. The two ASR rows ran on the WASM
+tier because headless Chromium has no WebGPU adapter; neither tier touches the server.
+
+Numbers this run recorded for plan 01-10's latency harness (all on the Space's CPU, one
+visitor): `synthesis_ms` 3088–6724 for こんにちは across four typed turns in the session,
+10194–10292 for the 36-mora long sentence (normal / speedScale 0.75); dispatch→speech-start
+4170–7252 ms for こんにちは; replay 0 ms and 0 requests; page→ready 4.0–16.8 s with
+`tutor.vrm` 3.0–7.4 s of that. Wake was 0.2 s every time — the Space was already running, so
+true cold-from-sleep remains 01-10's manual row.
+
 ## Why not the alternatives
 
 - **`cpu-basic` (free)** — no longer available. Attempting it returned **HTTP 402 Payment Required**

tests/e2e/test_avatar_loop.py

@@ -18,6 +18,8 @@ or a decoded ``AudioBuffer.duration``, never a timeline inference.
 
 from __future__ import annotations
 
+import json
+import re
 import time
 from pathlib import Path
 
@@ -872,3 +874,154 @@ def test_asr_wasm_fallback(
         assert page_errors == [], f"an uncaught exception reached the page: {page_errors}"
         assert errors == [], f"the facade emitted error events on the fallback path: {errors}"
         speech_events.wait_for(page, "speech-end", timeout_ms=SPEECH_END_TIMEOUT_MS)
+
+
+# ================================================================ credits and licences
+
+# The exact flow-down sentence rendered next to the Replay control (blocks.py).
+FLOW_DOWN_SENTENCE = "by using it you agree to comply with them"
+VOICEVOX_TERMS_HOST = "voicevox.hiroshiba.jp/term"
+ZUNDAMON_TERMS_HOST = "zunko.jp/con_ongen_kiyaku.html"
+# The notice sits directly under the Replay / Slower row; this is a generous bound on the
+# vertical gap between the bottom of the Replay button and the top of the notice.
+NOTICE_MAX_GAP_PX = 160
+
+# VRM 0.0 meta uses different key names for the same facts. Normalised in ONE place so the
+# assertions below never need an `or` fallback.
+VRM0_TO_VRM1_KEYS = {
+    "title": "name",
+    "author": "authors",
+    "otherLicenseUrl": "licenseUrl",
+    "reference": "references",
+}
+
+
+def licenses_block(name: str) -> dict[str, str]:
+    """Parse a ```<name> fenced block of ``key: value`` lines out of LICENSES.md."""
+    text = LICENSES_MD.read_text(encoding="utf-8")
+    match = re.search(rf"```{re.escape(name)}\n(.*?)\n```", text, re.S)
+    assert match, f"LICENSES.md has no ```{name} block"
+    block: dict[str, str] = {}
+    for line in match.group(1).splitlines():
+        if not line.strip() or line.lstrip().startswith("#"):
+            continue
+        key, sep, value = line.partition(":")
+        assert sep, f"LICENSES.md ```{name} line is not `key: value`: {line!r}"
+        block[key.strip()] = " ".join(value.split())
+    return block
+
+
+def normalise_vrm_meta(meta: dict) -> dict[str, str]:
+    """Flatten a VRM meta block - VRM 0.0 or 1.0 key names - to VRM 1.0 names with
+    string values: lists joined with ' | ', booleans as 'true'/'false', whitespace
+    collapsed, nulls dropped. The documented helper the assertions compare through."""
+    out: dict[str, str] = {}
+    for key, value in meta.items():
+        if value is None:
+            continue
+        name = VRM0_TO_VRM1_KEYS.get(key, key)
+        if isinstance(value, list):
+            text = " | ".join(str(v) for v in value)
+        elif isinstance(value, bool):
+            text = "true" if value else "false"
+        else:
+            text = str(value)
+        out[name] = " ".join(text.split())
+    return out
+
+
+@pytest.mark.deployed
+def test_credits_visible(page, space_url, warm_space):
+    """DPLY-04. On first load, with NO interaction: the VOICEVOX:ずんだもん credit is visible,
+    the flow-down notice sits by the Replay control with both terms linked, the About panel
+    carries both terms URLs and the VRM credit, and the credit is delivered by the server
+    rather than injected by this project's JavaScript.
+
+    The strings asserted are the ones LICENSES.md declares as required, parsed from its
+    ```credits block, so the document and the page cannot disagree.
+    """
+    credits = licenses_block("credits")
+    voice, avatar = credits["voice"], credits["avatar"]
+    assert voice == "VOICEVOX:ずんだもん", f"LICENSES.md declares the voice credit as {voice!r}"
+
+    page.goto(space_url, timeout=STAGE_ATTACHED_TIMEOUT_MS + 30_000)
+    page.wait_for_selector("#credits", state="visible", timeout=STAGE_ATTACHED_TIMEOUT_MS)
+
+    footer = page.locator("#credits")
+    assert footer.is_visible()
+    footer_text = footer.inner_text()
+    assert voice in footer_text, f"#credits reads {footer_text!r}"
+    assert avatar in footer_text, f"#credits reads {footer_text!r}"
+
+    notice = page.locator("#terms-notice")
+    assert notice.is_visible()
+    notice_text = notice.inner_text()
+    assert FLOW_DOWN_SENTENCE in notice_text, f"#terms-notice reads {notice_text!r}"
+    assert voice in notice_text
+    notice_links = notice.locator("a").evaluate_all("els => els.map((e) => e.href)")
+    assert any(VOICEVOX_TERMS_HOST in h for h in notice_links), notice_links
+    assert any(ZUNDAMON_TERMS_HOST in h for h in notice_links), notice_links
+
+    replay = page.locator("#replay-button")
+    assert replay.is_visible()
+    replay_box = replay.bounding_box()
+    notice_box = notice.bounding_box()
+    assert replay_box and notice_box
+    gap = notice_box["y"] - (replay_box["y"] + replay_box["height"])
+    print(
+        f"[deployed] credits: footer {footer_text!r}; notice {gap:.0f}px below Replay; "
+        f"notice links {notice_links}"
+    )
+    assert -1 <= gap <= NOTICE_MAX_GAP_PX, (
+        f"#terms-notice is {gap:.0f}px from the Replay button; it must sit with the controls"
+    )
+
+    about = page.locator("#about-panel")
+    assert about.count() == 1
+    about_text = about.inner_text()
+    about_links = about.locator("a").evaluate_all("els => els.map((e) => e.href)")
+    assert any(VOICEVOX_TERMS_HOST in h for h in about_links), about_links
+    assert any(ZUNDAMON_TERMS_HOST in h for h in about_links), about_links
+    assert voice in about_text
+    assert avatar in about_text, f"the About panel does not carry the VRM credit: {about_text!r}"
+
+    # Server-delivered, not injected by avatar/host.js: Gradio 6 serves an SSR shell and
+    # ships the component tree - including every server-rendered gr.HTML value - from
+    # GET /config (docs/HOSTING.md, first-deploy finding 5). The root HTML is reported for
+    # the record; the assertion is on what the server sends before any project script runs.
+    config = requests.get(f"{space_url}/config", timeout=30)
+    config.raise_for_status()
+    served = json.dumps(config.json(), ensure_ascii=False)
+    root_hits = requests.get(space_url, timeout=30).text.count(voice)
+    print(f"[deployed] credit in GET /config: {served.count(voice)}; in the SSR shell: {root_hits}")
+    assert voice in served, "the VOICEVOX credit is not in the server-delivered config"
+    assert avatar in served
+
+
+@pytest.mark.deployed
+def test_vrm_meta_matches_licenses(page, space_url, warm_space, wait_for_avatar_ready):
+    """DPLY-04. The shipped VRM's own embedded licence metadata agrees with LICENSES.md.
+
+    Read from the deployed page's ``vrmMeta`` (the stage copies the scalar fields of
+    ``vrm.meta`` out of the loaded file) and compared with the ```vrm_meta block, key for
+    key, through one normalising helper. If either side changes, this fails.
+    """
+    expected = licenses_block("vrm_meta")
+    for key in ("name", "authors", "licenseUrl"):
+        assert key in expected, f"LICENSES.md vrm_meta block lacks {key}"
+
+    wait_for_avatar_ready(page, space_url)
+    debug = read_debug(page)
+    meta = debug.get("vrmMeta") if debug else None
+    assert meta, "vrmMeta is missing from the deployed debug surface; the VRM loaded without it"
+    actual = normalise_vrm_meta(meta)
+    mismatches = {k: (v, actual.get(k)) for k, v in expected.items() if actual.get(k) != v}
+    print(
+        f"[deployed] vrmMeta ({len(actual)} fields, metaVersion {actual.get('metaVersion')!r}) "
+        f"vs LICENSES.md ({len(expected)} fields): {len(mismatches)} mismatch(es) {mismatches}"
+    )
+    assert not mismatches, (
+        "the shipped VRM's embedded meta disagrees with LICENSES.md "
+        f"(documented, actual): {mismatches}"
+    )
+    assert debug["vrmMetaTitle"] == expected["name"]