docs(voice): record the 雨晴はう licence terms and update the voice docs

- LICENSES.md: voice-model credit cell and the character row now describe 雨晴はう (Amehare
Project, https://amehau.com/?page_id=225, HTTP 403 to non-browser user agents); section 3
carries the VOICEVOX-shipped policy sentence, the five 音声(VOICEVOX_雨晴はう)の利用 lines, the
禁止事項 list and the お願い verbatim, states there is no credit-omission fee clause for this
character, flags the medical-claim clause for Phase 3; verification rows updated; the three
VOICEVOX rows above the character are unchanged
- docs/VOICEVOX-SETUP.md: speaker uuid / style 10, frame-quantisation table re-measured on
雨晴はう (91 / 499 / 666 frames, ratio 1.334669) with the old values kept as dated history,
licence-compliance section rewritten for the 雨晴はう terms, zundamon.vvm named as a
historical filename
- docs/ASSETS.md and docs/ASR-TIERS.md: credit string and a dated note that the ASR clips were
re-synthesised on 2026-09-12 while the A/B table predates it
- deferred-items: test_audio_unlocks_inside_the_gesture on the 6-frame コ under the strict
profile, with the re-run outcome and three levers

.planning/phases/02-japanese-language-core/deferred-items.md

@@ -97,3 +97,27 @@ Out-of-scope discoveries logged during execution, not fixed in the plan that fou
   in `docs/LATENCY.md § Translation` row 8. It is queued for the owner's ten-sentence quality pass
   (02-10 Task 3) and for `docs/LANGUAGE.md § Known translation issues`; it is an upstream OPUS-MT
   property, not a bug in this repo's wiring.
+
+## Logged by quick-260912-eg4 (2026-09-12)
+
+- **`tests/e2e/test_stage_standalone.py::test_audio_unlocks_inside_the_gesture` fails on the
+  regenerated 雨晴はう demo under the strict-autoplay profile** - `the mouth did not move on the
+  running clock: {'aa': 1, 'ih': 1, 'ou': 0, 'ee': 0, 'oh': 0}` (line 362, wants aa/ih/oh all
+  above 0.5). Cause: the only `oh` mora in こんにちは is コ, which the voice switch shortened from
+  **14 frames (0.149 s) to 6 frames (0.064 s)** (`avatar/stage.html` `DEMO_TIMELINE`, row 3).
+  With `avatar/lipsync.js` `ATTACK = 0.05` the weight crosses 0.5 only 25 ms into the mora, so
+  `oh > 0.5` is true for a **39 ms window**; the stage-recorded peak was exactly 0, i.e. no frame
+  was rendered inside the 64 ms mora at all. The profile is visibly starved - speech-start arrived
+  **3,161 ms** after the tap in the whole-directory run (2,964 ms on the re-run). This is the
+  Phase 1 entry "Sub-frame moras under slow rendering" (`01-voice-avatar-loop-skeleton/
+  deferred-items.md`: a 53 ms う was never written at ~15 fps) reached by a 64 ms コ.
+  Evidence it is rendering-rate, not data: `test_demo_timeline_drives_visemes[chromium]` - same
+  WAV, same timeline, same `VISEME_SAMPLER`, same aa/ih/oh > 0.5 assertion, normal profile -
+  **passed** in the same whole-directory run (38 passed / 1 failed / 25 skipped, 350 s).
+  Re-run of the single test once, in isolation (8.7 s): **failed again, identical peaks** - so on
+  this machine's strict-autoplay profile it is 2/2, not a coin flip. The assertion was not
+  loosened, `lipsync.js` was not touched, the demo sentence was not changed. Levers, for whoever
+  owns it: (1) `ATTACK` is a tunable (Phase 1's entry); (2) a demo sentence whose shortest mora is
+  >= 100 ms (Phase 1's entry); (3) assert visemes against frames actually rendered (count the
+  `setExpressionWeights` writes that fell inside each mora) rather than against wall-clock
+  sampling, so a starved renderer reports "no frame landed" instead of "the mouth did not move".

LICENSES.md

@@ -21,8 +21,8 @@ covered by that licence.
 | **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 and closed in plan 01-10 (kept; `docs/ASSETS.md`) |
 | **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) |
+| **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:雨晴はう` (only 雨晴はう, style 10, is used; the other voices packaged in `0.vvm` — 四国めたん, ずんだもん, 春日部つむぎ — 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 |
+| **雨晴はう** (character voice) | — (voice library shipped inside `0.vvm`) | Amehare Project (©2021-2026 Amehare Project, あめはれくりにっく) | https://amehau.com/ | https://amehau.com/?page_id=225 (雨晴はう 利用規約; returns HTTP 403 to non-browser user agents — read with a Chrome UA, HTTP 200, 13,954 bytes, 2026-09-12); restated by VOICEVOX at https://github.com/VOICEVOX/voicevox_resource/blob/main/character_info/雨晴はう_3474ee95-c274-47f9-aa1a-8322163d96f1/policy.md | 「個人様/企業様ともに商用・非商用問わず利用可能です。ただし音声のR18指定利用はご遠慮ください。」 and 「VOICEVOXのクレジット記載は必須となります。」; the character-name credit is optional: 「キャラクター名のクレジット表記は必須ではありませんが、記載があれば嬉しいです。(例:VOICEVOX:雨晴はう)」 | **`VOICEVOX:雨晴はう`** — the terms' own example, exact, ASCII colon, no spaces; satisfies the mandatory VOICEVOX credit and the welcomed character credit in one string | Persistent footer credit + About panel on first paint (server-delivered, not injected); flow-down notice linking both terms URLs; the audio is generated directly by VOICEVOX CORE, so the voice-changer prohibition does not apply; 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** (incl. its `GLTFLoader` addon) | 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: redistribution permitted with the copyright and permission notice kept | none | **Vendored into this repository** (plan 01-10) as `avatar/vendor/three.mjs` and `avatar/vendor/GLTFLoader.mjs`, the package's `LICENSE` copied alongside as `avatar/vendor/LICENSE-three.txt`; served by the Space itself, no CDN in the render path. Provenance, hashes and the regeneration command: `avatar/vendor/README.md` |
 | **@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: redistribution permitted with the copyright and permission notice kept | none | **Vendored into this repository** (plan 01-10) as `avatar/vendor/three-vrm.mjs` (the build against three@0.185.1), the package's `LICENSE` copied alongside as `avatar/vendor/LICENSE-three-vrm.txt`; served by the Space itself |
@@ -139,25 +139,50 @@ forbid redistributing the *software*; the model terms permit redistributing the
 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.
+**3. Character** — 雨晴はう, Amehare Project (©2021-2026 Amehare Project, あめはれくりにっく),
+https://amehau.com/?page_id=225 (「雨晴はう 利用規約」). The page returns **HTTP 403 to non-browser
+user agents**; it was read with a Chrome UA (HTTP 200, 13,954 bytes) on 2026-09-12 — a verifier
+whose script gets a 403 has hit that wall, not a takedown. VOICEVOX's own restatement, shipped in
+`voicevox_resource` as `character_info/雨晴はう_3474ee95-c274-47f9-aa1a-8322163d96f1/policy.md`,
+reads verbatim:
+「雨晴はうの音声ライブラリを用いて生成した音声は、「VOICEVOX:雨晴はう」とクレジットを記載すれば、商用・非商用で利用可能です。」
+
+The character's own terms, section 音声(VOICEVOX_雨晴はう)の利用, verbatim and in order:
+
+- 「VOICEVOXの利用規約を守ってご使用下さい。」
+- 「個人様/企業様ともに商用・非商用問わず利用可能です。ただし音声のR18指定利用はご遠慮ください。」
+- 「キャラクター名のクレジット表記は必須ではありませんが、記載があれば嬉しいです。(例:VOICEVOX:雨晴はう)」
+- 「VOICEVOXのクレジット記載は必須となります。仕様上で記載が出来ない場合はお問合せください。」
+- 「VOICEVOXで直接生成していない音声及びVOICEVOX出力音声を利用して作成したボイスチェンジャーモデルの利用を禁止。」
+  The terms list approved exceptions to this line (WEB版VOICEVOX / 文章(コメント)読み上げ /
+  ゆかりねっと等). This project's audio is generated directly by VOICEVOX CORE from the `.vvm`, and
+  no voice-changer model is trained or used, so the prohibition does not apply here.
+
+So the VOICEVOX credit is **mandatory** and the character-name credit is optional-but-welcomed;
+the string `VOICEVOX:雨晴はう` — the terms' own example — satisfies both at once, and Phase 1's
+placement standard (a persistent footer beside the avatar **and** the About panel, open on first
+load) is kept unchanged.
+
+禁止事項 (general), verbatim, recorded so later phases can check against them:
+法令に違反する行為又は犯罪行為に関連する行為 / 第三者に不利益、損害、不快感を与える行為 / 第三者に対する詐欺または脅迫行為・公序良俗に著しく反する行為 / 特定の思想運動の勧誘・社会問題の特定の主張に当たる行為団体 / 許可無しに公式イラストを利用した二次配布、商品の販売行為 / キャラクターの著作情報を偽る行為、自作発言 / 故意に医療知識の誤った主張をする及び誤解を生む行為.
+お願い: 倫理的配慮が欠如している作品の公開はお控えください。
+
+The clause Phase 3 must be checked against is 「故意に医療知識の誤った主張をする及び誤解を生む行為」:
+the character is a nurse, and once an LLM produces the avatar's lines, a tutor that is sometimes
+wrong is not in scope of an *intentional* false medical claim; a product designed to mislead would
+be — the same reading the ずんだもん "intentional falsehood" clause was given.
+
+**There is no ¥400,000 credit-omission clause for this character.** The ずんだもん terms this
+section described until 2026-09-12 carried one (per-character licensing at ¥400,000 + tax without
+the credit), which is why that explanation stood here. 雨晴はう's terms contain no such fee. The
+credit is still rendered on first paint rather than behind a click, because the VOICEVOX credit is
+mandatory (「VOICEVOXのクレジット記載は必須となります。仕様上で記載が出来ない場合はお問合せください。」)
+and because consuming the `AudioQuery` requires it on its own (next paragraph).
+
+There is no escalation path and none should be sought: the 雨晴はう terms carry a disclaimer of
+liability, state that they may change at any time, and state that questions about other parties'
+terms are not answered. The posture is unchanged: read the terms (with a browser UA — see the 403
+caveat above), 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
@@ -167,8 +192,9 @@ 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.
+README (https://voicevox.hiroshiba.jp/nemo/term/); there is no third-party rights holder. 雨晴はう
+has no ¥400,000 credit-omission clause either, so what Nemo removes is the third-party rights
+holder, not a fee. The cost is the loss of a recognisable character.
 
 ## Open JTalk dictionary — `voicevox/open_jtalk_dic_utf_8-1.11/`
 
@@ -343,7 +369,7 @@ overrule it). **Result: no gaps.**
 | 1 | VRM character | yes | `docs/ASSETS.md` (same source, licence URL and embedded flags); `tests/e2e/test_avatar_loop.py::test_vrm_meta_matches_licenses` green on the deployed file. Sourcing gate closed: the file is final |
 | 2 | VOICEVOX software | yes | `docs/VOICEVOX-SETUP.md` (terms URL `voicevox.hiroshiba.jp/term/`, MIT at tag 0.17.0); `requirements.txt` installs exactly the 0.17.0 release wheel from the official URL |
 | 3 | VOICEVOX voice model (`.vvm`) | yes | `voicevox_vvm` 0.17.0 release in `docs/VOICEVOX-SETUP.md`; `sha256sum voicevox/model/zundamon.vvm` = `ecd35374…1318`, the value in the row; the file is an LFS object |
-| 4 | ずんだもん character terms | yes | `docs/VOICEVOX-SETUP.md` (`zunko.jp/con_ongen_kiyaku.html`); the credit string in the `credits` block is what `test_credits_visible` asserted on the Space |
+| 4 | 雨晴はう character terms | yes | `docs/VOICEVOX-SETUP.md` (`amehau.com/?page_id=225`, browser UA required — 403 otherwise); the credit string in the `credits` block is what `test_credits_visible` asserts; NOT yet re-asserted on the Space (still on 7eb3e94 with ずんだもん until the owner pushes) |
 | 5 | Open JTalk dictionary | yes | `voicevox/open_jtalk_dic_utf_8-1.11/COPYING` first line `Copyright (c) 2009, Nara Institute of Science and Technology, Japan.`; the Nagoya Institute of Technology / HTS Working Group attribution for Open JTalk itself, as `docs/VOICEVOX-SETUP.md` records; the About panel names both institutions |
 | 6 | three.js | yes | `avatar/vendor/README.md` (three@0.185.1 es2022 build, SHA256 recorded, `tests/test_vendor.py` re-checks); `avatar/vendor/LICENSE-three.txt` is the MIT notice © 2010-2026 three.js authors |
 | 7 | @pixiv/three-vrm | yes | `avatar/vendor/README.md` (3.5.5 built against three@0.185.1); `avatar/vendor/LICENSE-three-vrm.txt` is the MIT notice © 2019-2026 pixiv Inc. |
@@ -386,7 +412,7 @@ which is the obligation that a `LICENSES.md` row alone would not discharge.
 | `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 |
+| Terms URLs answer | voicevox.hiroshiba.jp/term/ 200 · zunko.jp/con_ongen_kiyaku.html 200 (historical — the ずんだもん row, not re-fetched since the 2026-09-12 voice switch) · voicevox.hiroshiba.jp/nemo/term/ 200 · vrm.dev/licenses/1.0/ 200 · amehau.com/?page_id=225 403 (non-browser UA) / 200 with a Chrome UA, 13,954 bytes, 2026-09-12 |
 
 ### Phase 2 assets (plan 02-11)
 

docs/ASR-TIERS.md

@@ -13,6 +13,10 @@ from a model card.
 | Runtime | `https://esm.sh/@huggingface/transformers@4.2.0` |
 | Clips | `speech_ja.wav`, `speech_ja_long.wav`, `speech_ja_slow.wav` — VOICEVOX ずんだもん, plan 01-04 |
 
+The three clips were re-synthesised with 雨晴はう (VOICEVOX style 10) on 2026-09-12 (quick task
+260912-eg4); every number in this document predates that and was measured on the ずんだもん clips.
+`tests/fixtures/asr_ab_results.json` was deliberately not regenerated.
+
 ---
 
 ## Tier design

docs/ASSETS.md

@@ -132,10 +132,11 @@ oval), い (flat, teeth visible) and う (small, pursed) are clearly distinct, w
 asks for. A bespoke character could widen the five-way distinction; the pipeline drives all five
 presets independently to weight 1.0 (reference sheet `docs/evidence/2026-09-06-viseme-ref-*.png`).
 
-## Voice — VOICEVOX / ずんだもん
+## Voice — VOICEVOX / 雨晴はう
 
 **Already recorded in full by plan 01-04. See `docs/VOICEVOX-SETUP.md`**, which carries all three
-licence layers, the exact `VOICEVOX:ずんだもん` credit string, the placement standard, the flow-down
+licence layers, the exact `VOICEVOX:雨晴はう` credit string (switched from ずんだもん on 2026-09-12),
+the placement standard, the flow-down
 obligation, the Nemo pre-cleared swap, and the Open JTalk dictionary's BSD-3-Clause notice with its
 dual Nara/Nagoya attribution. Not duplicated here; `LICENSES.md` (plan 01-10) should read that file
 directly.

docs/VOICEVOX-SETUP.md

@@ -132,11 +132,13 @@ set cannot silently drift away from the engine.
 
 | Fixture | speedScale | Frames | Duration (s) |
 |---|---|---|---|
-| `short` | 1.0 | 99 | 1.056 |
-| `long` | 1.0 | 516 | 5.504 |
-| `slow` | 0.75 | 692 | 7.381333333333333 |
+| `short` | 1.0 | 91 | 0.9706666666666667 |
+| `long` | 1.0 | 499 | 5.322666666666667 |
+| `slow` | 0.75 | 666 | 7.104 |
 
-Note that `slow / long = 1.341085`, **not** exactly `1/0.75 = 1.333333`. Re-quantisation after
+(Re-measured on 雨晴はう, 2026-09-12; the ずんだもん values were 99/516/692, ratio 1.341085.)
+
+Note that `slow / long = 1.334669`, **not** exactly `1/0.75 = 1.333333`. Re-quantisation after
 scaling means the realised ratio lands near the requested one, not on it. A test that asserts
 "exactly 1/0.75× longer" against *durations* will fail; assert the timeline matches the WAV
 instead, which is the property that actually matters for lip-sync.
@@ -177,15 +179,21 @@ phrase keys (`moras`, `accent`, `pause_mora`, `is_interrogative`) pass through u
 | Voice model | `voicevox/model/zundamon.vvm` | 59,308,488 B | `voicevox_vvm` 0.17.0 asset, original filename **`0.vvm`** | **yes** — root `.gitattributes` `*.vvm` |
 | VOICEVOX ONNX Runtime | *not in the repo* — fetched to `voicevox_runtime/` (gitignored) | 8.2 MB (linux-x64 CPU `.tgz`) | `github.com/VOICEVOX/onnxruntime-builder` release `voicevox_onnxruntime-1.23.2` | **NO — never vendored** (option **c**, below) |
 
+The on-disk name `zundamon.vvm` is historical (chosen when ずんだもん was the speaker); the file is
+upstream `0.vvm` unchanged and carries all four voices, so the 2026-09-12 voice switch did not
+rename it.
+
 ```text
 SHA256 zundamon.vvm: ecd35374d4182cd883cba5040376f7f888cc6ba248b1c2f4cea07cdb34bb1318
-speaker_uuid:        388f246b-8c41-4ac1-8e2d-5d79f3ff56d9
-style_id (ノーマル):  3
+speaker_uuid:        3474ee95-c274-47f9-aa1a-8322163d96f1  (雨晴はう)
+style_id (ノーマル):  10
 ```
 
-`style_id` **3** is the integer constant `SPEAKER_STYLE_ID` in `src/japanese_avatar/voice/tts.py`.
+ずんだもん (`388f246b-8c41-4ac1-8e2d-5d79f3ff56d9`, style 3) was the speaker until 2026-09-12.
+
+`style_id` **10** is the integer constant `SPEAKER_STYLE_ID` in `src/japanese_avatar/voice/tts.py`.
 
-`0.vvm` also carries 四国めたん (2/0/6/4), 春日部つむぎ (8) and 雨晴はう (10). Only ずんだもん is
+`0.vvm` also carries 四国めたん (2/0/6/4), ずんだもん (3) and 春日部つむぎ (8). Only 雨晴はう is
 credited and used; the others are present because the model file is packaged that way upstream.
 Confirmed against the `metas.json` inside the downloaded file and against the character↔style
 table in the release's own `README.txt`.
@@ -195,7 +203,8 @@ table in the release's own `README.txt`.
 The `voicevox_vvm` 0.17.0 release ships 29 assets — 25 talk models `0.vvm`..`24.vvm`, `n0.vvm`
 (Nemo), `s0.vvm` (song), plus `README.txt` and `TERMS.txt`. Each talk model is 57-67 MB. The
 character↔file index lives in `README.txt` (section 音声モデル(.vvm)ファイルと声…の対応表), which
-resolves ずんだもん ノーマル to **`0.vvm`, style ID 3**. Only that one file was downloaded.
+resolves 雨晴はう ノーマル to **`0.vvm`, style ID 10** (the file was originally downloaded for
+ずんだもん ノーマル, `0.vvm` style ID 3 — the same file). Only that one file was downloaded.
 
 ---
 
@@ -245,7 +254,7 @@ the `StyleType::StreamingTalk` variant were **introduced in core 0.17.0**. So:
 
 Evidence, executed: loading `voicevox/model/zundamon.vvm` into a 0.17.0 `Synthesizer` succeeded and
 `create_audio_query("こんにちは", 3)` → `synthesis(...)` returned a 50,732-byte RIFF WAV at
-24000 Hz / mono / 16-bit, duration 1.056 s.
+24000 Hz / mono / 16-bit, duration 1.056 s (measured with style 3 in plan 01-04).
 
 Pin both to **0.17.0**. If either is bumped, regenerate the fixtures with
 `tests/fixtures/make_synth_fixtures.py` — `tests/test_tts_contract.py::test_fixtures_match_current_engine`
@@ -271,7 +280,8 @@ Measured on Windows / Python 3.12.12, warm disk, runtime already present:
 | **Total import-to-ready (`warmup()`)** | **1.73 s** |
 | Second `warmup()` call (cache hit) | 0.000001 s |
 
-Per-turn, once warm: `create_audio_query` **1.2 ms**, `synthesis` **1215 ms** for 「こんにちは」.
+Per-turn, once warm: `create_audio_query` **1.2 ms**, `synthesis` **1215 ms** for 「こんにちは」
+(measured with style 3 in plan 01-04).
 Both are recorded through `TurnTimings`, so plan 01-08's latency harness gets them for free.
 
 **The lazy runtime download is exercised, not assumed.** Pointing `VOICEVOX_ORT_DIR` at an empty
@@ -302,28 +312,34 @@ The VRM's equivalent lives in `docs/ASSETS.md`, owned by plan 01-01. Plan 01-09
   as `README.txt` / `TERMS.txt` alongside the release assets. 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 what licenses committing `zundamon.vvm` to this repo.
-- **Character** — ずんだもん, SSS LLC, <https://zunko.jp/con_ongen_kiyaku.html>. One document covers
-  nine characters. Commercial and non-commercial use permitted *with credit*; without credit,
-  per-character licensing is ¥400,000 (+ tax).
+- **Character** — 雨晴はう, Amehare Project (©2021-2026 Amehare Project, あめはれくりにっく),
+  <https://amehau.com/?page_id=225> (雨晴はう 利用規約; the page returns HTTP 403 to non-browser
+  user agents — read it with a browser UA). Verbatim:
+  「個人様/企業様ともに商用・非商用問わず利用可能です。ただし音声のR18指定利用はご遠慮ください。」 and
+  「VOICEVOXのクレジット記載は必須となります。仕様上で記載が出来ない場合はお問合せください。」
+  Switched from ずんだもん on 2026-09-12 (quick task 260912-eg4); speaker uuid
+  `3474ee95-c274-47f9-aa1a-8322163d96f1`, style 10.
 
 ### The credit string
 
 ```text
-VOICEVOX:ずんだもん
+VOICEVOX:雨晴はう
 ```
 
 Exact, including the ASCII colon and no spaces. Later plans render it; this document is where the
 string is defined.
 
-**Placement requirement** (character terms clause 2):
-「アプリなどでの利用の場合は、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）」
-— an app introduction / about screen, in a place findable with a little looking. Phase 1 satisfies
-this with a persistently visible footer line next to the avatar **plus** an About/Credits surface.
+**Credit requirement** (character terms, 音声(VOICEVOX_雨晴はう)の利用): the VOICEVOX credit is
+mandatory and the character credit is optional —
+「キャラクター名のクレジット表記は必須ではありませんが、記載があれば嬉しいです。(例:VOICEVOX:雨晴はう)」
+— so `VOICEVOX:雨晴はう`, the terms' own example, satisfies both in one string. The placement
+standard is kept unchanged from Phase 1: a persistently visible footer line next to the avatar
+**plus** an About/Credits surface, open on first load.
 
 **Flow-down obligation** (software clause 3 / voice-model clause 4): when audio generated here is
 made available to others, those others must be bound to the same terms. Anywhere audio is
 user-obtainable (playback, replay, download), display a terms notice to the effect of *"Synthesised
-audio is provided under the VOICEVOX and VOICEVOX:ずんだもん terms of use; by using it you agree to
+audio is provided under the VOICEVOX and VOICEVOX:雨晴はう terms of use; by using it you agree to
 comply with them"*, linking both URLs above.
 
 **The credit obligation survives an engine swap.** The official Q&A answers
@@ -333,11 +349,13 @@ timeline consumes — triggers the credit requirement on its own.
 
 ### There is no "ask the rights holder" step, ever
 
-SSS LLC 免責条項 2: 「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。
-特に無償利用の範囲についての問い合わせについては回答いたしかねます。…できるかぎりガイドラインを自身で
-読み込んだ上での判断をお願いいたします。」 They decline "does my use qualify?" questions as a matter
-of policy. The compliance posture is: read the guideline (done, above), comply visibly, document
-here. Recorded so it is never re-litigated in a later phase.
+The 雨晴はう terms carry a disclaimer of liability, state that they may change at any time, and
+state that questions about other parties' terms are not answered. The one contact they invite is
+「仕様上で記載が出来ない場合はお問合せください」, which does not arise here because the credit is
+rendered. The terms page answers HTTP 403 to non-browser user agents (read with a Chrome UA, HTTP
+200, 13,954 bytes, 2026-09-12) — a scripted 403 is not a takedown. The compliance posture is
+unchanged: read the terms (done, above), comply visibly, document here. Recorded so it is never
+re-litigated in a later phase.
 
 ### Pre-cleared swap-in: VOICEVOX Nemo
 
@@ -348,9 +366,11 @@ the swap is a two-line change (`SPEAKER_STYLE_ID` plus the model path):
   10000-10008 (男声1-3, 女声1-6), all ノーマル.
 - Credit string is just `VOICEVOX Nemo` — no character name, no third-party rights holder.
 - Terms: <https://voicevox.hiroshiba.jp/nemo/term/>, restated in the `voicevox_vvm` README.
-- No ¥400,000 credit-omission clause, because there is no separate character licensor.
+- No ¥400,000 credit-omission clause — and 雨晴はう has none either; what Nemo removes is the
+  third-party rights holder, not a fee.
 
-Cost of the swap: the loss of a recognisable character identity. Recommendation remains ずんだもん.
+Cost of the swap: the loss of a recognisable character identity. Recommendation remains 雨晴はう
+(switched from ずんだもん on 2026-09-12 for a more natural voice).
 
 ### Open JTalk dictionary licence