test(01-04): capture ground-truth synthesis fixtures for AVTR-02

- make_synth_fixtures.py generates three cases and refuses to write a 'long'
fixture without 20+ moras, a devoiced vowel and a pause mora
- Durations and frame counts come from the WAV header, never from summing the query
- Establish the correct frame-quantisation formula empirically: VOICEVOX quantises
each phoneme at speed 1.0 and then divides the frame count by speedScale, rather
than dividing the length first. Verified 24/24 across 6 speeds and 4 sentences;
the length-first form is right only 8/24 and is off by up to 5 frames
- The generator asserts that formula reproduces each fixture's true frame count,
so the fixtures cannot drift away from the engine unnoticed
- Add a regression guard that fires if a voicevox_core bump changes timings
- Memoise synthesis in the test module: it was doing six identical syntheses and
had pushed the quick loop to ~20s against a 15s budget; now ~11s
- Write fixture JSON with explicit LF so regeneration is byte-identical across OSes

docs/VOICEVOX-SETUP.md

@@ -74,13 +74,72 @@ institutions are genuinely both involved and RESEARCH conflated them.
 Everything else RESEARCH inferred was correct, including the whole `Onnxruntime` → `OpenJtalk` →
 `Synthesizer` → `VoiceModelFile` construction order and the exact Linux wheel filename.
 
-### Bonus: the frame-quantisation algorithm is now empirically verified
+**6. `speedScale` is applied to the frame count, not to the phoneme length.** This is the most
+consequential correction in this document and it has its own section below.
 
-RESEARCH's `build_timeline` pipeline (flatten → wrap in pre/post silence → divide by `speedScale`
-→ quantise each phoneme with `round(sec * 93.75)` using banker's rounding → accumulate) was
-replayed against a real 36-mora synthesis. The accumulated timeline total and the true WAV
-duration agree to **2.7e-15 s** — i.e. exactly, to float representation. Plan 01-06 can implement
-that algorithm as written and expect bit-exact agreement.
+---
+
+## Frame quantisation — read this before writing `visemes.py`
+
+**Plan 01-06 depends on this section.** 01-RESEARCH.md's `build_timeline` gets the order wrong,
+and the error is invisible at normal speed, which is exactly what makes it dangerous.
+
+RESEARCH's pipeline divides each phoneme length by `speedScale` and *then* quantises:
+
+```python
+frames = round(length / speed_scale * 93.75)          # WRONG
+```
+
+VOICEVOX CORE 0.17.0 actually quantises **first**, at speed 1.0, and then divides the resulting
+**frame count** and rounds again:
+
+```python
+frames = round(round(length * 93.75) / speed_scale)   # CORRECT
+```
+
+At `speedScale == 1.0` the two are identical, so RESEARCH's version looks verified. They diverge
+everywhere else.
+
+Measured: 4 sentences × 6 speed values (1.0, 0.9, 0.75, 0.5, 1.25, 1.5), predicted total frames
+compared against the true frame count of the synthesised WAV (`getnframes() // 256`):
+
+| Formula | Correct |
+|---|---|
+| `round(length / speed * 93.75)` (01-RESEARCH.md) | **8 / 24** |
+| `round(round(length * 93.75) / speed)` | **24 / 24** |
+
+Worst observed error from the wrong formula: **5 frames ≈ 53 ms** on the long sentence at
+`speedScale = 1.25`. That is well past the ±1 frame (10.667 ms) tolerance `test_no_drift_long_utterance`
+is specified with, and it is exactly the symptom `01-RESEARCH.md` Pitfall 5 describes — "normal
+speed syncs, slow speed drifts" — reached from the other direction.
+
+Everything else about the pipeline is confirmed and should be implemented as RESEARCH describes:
+
+- Flatten each accent phrase's `moras`, then its `pause_mora` if present — pause **after**.
+- Wrap the whole sequence in `prePhonemeLength` / `postPhonemeLength` silence moras, and scale
+  those by `speedScale` like any other phoneme. They are not exempt.
+- Emit the consonant (mouth closed) before the vowel within a mora.
+- Accumulate the quantised frame counts, never the raw floats. Per-phoneme rounding error is up to
+  ±0.5 frame ≈ ±5.3 ms, and the long fixture has 61 phonemes.
+- Use banker's rounding. Python's built-in `round()` already is; JavaScript's `Math.round()` is
+  round-half-up and would disagree, which is an independent reason the timeline is built in Python
+  and the browser only plays it.
+- Steps involving `pauseLength` / `pauseLengthScale` are no-ops — those fields do not exist here.
+
+The committed fixtures are the proof: `tests/fixtures/make_synth_fixtures.py` asserts the correct
+formula reproduces each fixture's true frame count before it will write anything, so the fixture
+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 |
+
+Note that `slow / long = 1.341085`, **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.
 
 ---
 

tests/fixtures/audio_query_long.json

@@ -0,0 +1,355 @@
+{
+  "accent_phrases": [
+    {
+      "accent": 1,
+      "is_interrogative": false,
+      "moras": [
+        {
+          "consonant": "ky",
+          "consonant_length": 0.1269165426492691,
+          "pitch": 5.891244888305664,
+          "text": "キョ",
+          "vowel": "o",
+          "vowel_length": 0.10178578644990921
+        },
+        {
+          "consonant": null,
+          "consonant_length": null,
+          "pitch": 5.944635391235352,
+          "text": "オ",
+          "vowel": "o",
+          "vowel_length": 0.08867011219263077
+        },
+        {
+          "consonant": "w",
+          "consonant_length": 0.06420903652906418,
+          "pitch": 5.655149459838867,
+          "text": "ワ",
+          "vowel": "a",
+          "vowel_length": 0.13292233645915985
+        }
+      ],
+      "pause_mora": null
+    },
+    {
+      "accent": 1,
+      "is_interrogative": false,
+      "moras": [
+        {
+          "consonant": null,
+          "consonant_length": null,
+          "pitch": 5.756601810455322,
+          "text": "イ",
+          "vowel": "i",
+          "vowel_length": 0.09258951991796494
+        },
+        {
+          "consonant": null,
+          "consonant_length": null,
+          "pitch": 6.103951930999756,
+          "text": "イ",
+          "vowel": "i",
+          "vowel_length": 0.0974084809422493
+        }
+      ],
+      "pause_mora": null
+    },
+    {
+      "accent": 1,
+      "is_interrogative": false,
+      "moras": [
+        {
+          "consonant": "t",
+          "consonant_length": 0.0657225176692009,
+          "pitch": 6.139893531799316,
+          "text": "テ",
+          "vowel": "e",
+          "vowel_length": 0.1257137954235077
+        },
+        {
+          "consonant": null,
+          "consonant_length": null,
+          "pitch": 6.0533294677734375,
+          "text": "ン",
+          "vowel": "N",
+          "vowel_length": 0.07083306461572647
+        },
+        {
+          "consonant": "k",
+          "consonant_length": 0.05886319279670715,
+          "pitch": 5.938107967376709,
+          "text": "キ",
+          "vowel": "i",
+          "vowel_length": 0.059421248733997345
+        },
+        {
+          "consonant": "d",
+          "consonant_length": 0.05466592311859131,
+          "pitch": 5.8197479248046875,
+          "text": "デ",
+          "vowel": "e",
+          "vowel_length": 0.09957551956176758
+        },
+        {
+          "consonant": "s",
+          "consonant_length": 0.03665236383676529,
+          "pitch": 0.0,
+          "text": "ス",
+          "vowel": "U",
+          "vowel_length": 0.06477490812540054
+        },
+        {
+          "consonant": "k",
+          "consonant_length": 0.07497167587280273,
+          "pitch": 5.6952314376831055,
+          "text": "カ",
+          "vowel": "a",
+          "vowel_length": 0.09379018098115921
+        },
+        {
+          "consonant": "r",
+          "consonant_length": 0.03401697427034378,
+          "pitch": 5.5977702140808105,
+          "text": "ラ",
+          "vowel": "a",
+          "vowel_length": 0.14935533702373505
+        }
+      ],
+      "pause_mora": {
+        "consonant": null,
+        "consonant_length": null,
+        "pitch": 0.0,
+        "text": "、",
+        "vowel": "pau",
+        "vowel_length": 0.2590501606464386
+      }
+    },
+    {
+      "accent": 5,
+      "is_interrogative": false,
+      "moras": [
+        {
+          "consonant": "k",
+          "consonant_length": 0.08955883234739304,
+          "pitch": 5.693568229675293,
+          "text": "コ",
+          "vowel": "o",
+          "vowel_length": 0.09677395969629288
+        },
+        {
+          "consonant": null,
+          "consonant_length": null,
+          "pitch": 5.938162803649902,
+          "text": "オ",
+          "vowel": "o",
+          "vowel_length": 0.12056692689657211
+        },
+        {
+          "consonant": null,
+          "consonant_length": null,
+          "pitch": 6.016718864440918,
+          "text": "エ",
+          "vowel": "e",
+          "vowel_length": 0.147250235080719
+        },
+        {
+          "consonant": null,
+          "consonant_length": null,
+          "pitch": 6.010623931884766,
+          "text": "ン",
+          "vowel": "N",
+          "vowel_length": 0.08805551379919052
+        },
+        {
+          "consonant": null,
+          "consonant_length": null,
+          "pitch": 6.007989883422852,
+          "text": "オ",
+          "vowel": "o",
+          "vowel_length": 0.10105019062757492
+        }
+      ],
+      "pause_mora": null
+    },
+    {
+      "accent": 3,
+      "is_interrogative": false,
+      "moras": [
+        {
+          "consonant": "s",
+          "consonant_length": 0.0925176665186882,
+          "pitch": 6.003979206085205,
+          "text": "サ",
+          "vowel": "a",
+          "vowel_length": 0.1288621425628662
+        },
+        {
+          "consonant": null,
+          "consonant_length": null,
+          "pitch": 6.026742935180664,
+          "text": "ン",
+          "vowel": "N",
+          "vowel_length": 0.0830387994647026
+        },
+        {
+          "consonant": "p",
+          "consonant_length": 0.054095521569252014,
+          "pitch": 6.099804878234863,
+          "text": "ポ",
+          "vowel": "o",
+          "vowel_length": 0.1004747673869133
+        }
+      ],
+      "pause_mora": null
+    },
+    {
+      "accent": 4,
+      "is_interrogative": false,
+      "moras": [
+        {
+          "consonant": "sh",
+          "consonant_length": 0.021739566698670387,
+          "pitch": 0.0,
+          "text": "シ",
+          "vowel": "I",
+          "vowel_length": 0.04718482494354248
+        },
+        {
+          "consonant": "t",
+          "consonant_length": 0.05912453681230545,
+          "pitch": 6.083391189575195,
+          "text": "テ",
+          "vowel": "e",
+          "vowel_length": 0.08641223609447479
+        },
+        {
+          "consonant": "k",
+          "consonant_length": 0.05805716663599014,
+          "pitch": 6.061047554016113,
+          "text": "カ",
+          "vowel": "a",
+          "vowel_length": 0.0868045762181282
+        },
+        {
+          "consonant": "r",
+          "consonant_length": 0.03226054459810257,
+          "pitch": 5.876008987426758,
+          "text": "ラ",
+          "vowel": "a",
+          "vowel_length": 0.14383982121944427
+        }
+      ],
+      "pause_mora": {
+        "consonant": null,
+        "consonant_length": null,
+        "pitch": 0.0,
+        "text": "、",
+        "vowel": "pau",
+        "vowel_length": 0.2888171374797821
+      }
+    },
+    {
+      "accent": 5,
+      "is_interrogative": false,
+      "moras": [
+        {
+          "consonant": "k",
+          "consonant_length": 0.0854499563574791,
+          "pitch": 5.536592483520508,
+          "text": "カ",
+          "vowel": "a",
+          "vowel_length": 0.12839564681053162
+        },
+        {
+          "consonant": null,
+          "consonant_length": null,
+          "pitch": 5.821318626403809,
+          "text": "イ",
+          "vowel": "i",
+          "vowel_length": 0.07545183598995209
+        },
+        {
+          "consonant": "m",
+          "consonant_length": 0.05912942439317703,
+          "pitch": 6.023046493530273,
+          "text": "モ",
+          "vowel": "o",
+          "vowel_length": 0.0851563960313797
+        },
+        {
+          "consonant": "n",
+          "consonant_length": 0.05413999408483505,
+          "pitch": 6.078396797180176,
+          "text": "ノ",
+          "vowel": "o",
+          "vowel_length": 0.08871601521968842
+        },
+        {
+          "consonant": "n",
+          "consonant_length": 0.07090981304645538,
+          "pitch": 6.0716552734375,
+          "text": "ニ",
+          "vowel": "i",
+          "vowel_length": 0.0862032100558281
+        }
+      ],
+      "pause_mora": null
+    },
+    {
+      "accent": 3,
+      "is_interrogative": false,
+      "moras": [
+        {
+          "consonant": null,
+          "consonant_length": null,
+          "pitch": 6.005620956420898,
+          "text": "イ",
+          "vowel": "i",
+          "vowel_length": 0.09876994043588638
+        },
+        {
+          "consonant": "k",
+          "consonant_length": 0.07177173346281052,
+          "pitch": 6.0334930419921875,
+          "text": "キ",
+          "vowel": "i",
+          "vowel_length": 0.057893283665180206
+        },
+        {
+          "consonant": "m",
+          "consonant_length": 0.058485276997089386,
+          "pitch": 6.074601173400879,
+          "text": "マ",
+          "vowel": "a",
+          "vowel_length": 0.09547590464353561
+        },
+        {
+          "consonant": "sh",
+          "consonant_length": 0.030789220705628395,
+          "pitch": 0.0,
+          "text": "シ",
+          "vowel": "I",
+          "vowel_length": 0.06114967167377472
+        },
+        {
+          "consonant": "t",
+          "consonant_length": 0.05993541330099106,
+          "pitch": 5.849719524383545,
+          "text": "タ",
+          "vowel": "a",
+          "vowel_length": 0.1888434886932373
+        }
+      ],
+      "pause_mora": null
+    }
+  ],
+  "intonationScale": 1.0,
+  "kana": "キョ'オワ/イ'イ/テ'ンキデ_スカラ、コオエンオ'/サンポ'/_シテカラ'、カイモノニ'/イキマ'_シタ",
+  "outputSamplingRate": 24000,
+  "outputStereo": false,
+  "pitchScale": 0.0,
+  "postPhonemeLength": 0.10000000149011612,
+  "prePhonemeLength": 0.10000000149011612,
+  "speedScale": 1.0,
+  "volumeScale": 1.0
+}

tests/fixtures/audio_query_short.json

@@ -0,0 +1,60 @@
+{
+  "accent_phrases": [
+    {
+      "accent": 5,
+      "is_interrogative": false,
+      "moras": [
+        {
+          "consonant": "k",
+          "consonant_length": 0.09125985950231552,
+          "pitch": 5.689502716064453,
+          "text": "コ",
+          "vowel": "o",
+          "vowel_length": 0.1464204043149948
+        },
+        {
+          "consonant": null,
+          "consonant_length": null,
+          "pitch": 5.868537902832031,
+          "text": "ン",
+          "vowel": "N",
+          "vowel_length": 0.07110694795846939
+        },
+        {
+          "consonant": "n",
+          "consonant_length": 0.032809868454933167,
+          "pitch": 5.991527080535889,
+          "text": "ニ",
+          "vowel": "i",
+          "vowel_length": 0.09959950298070908
+        },
+        {
+          "consonant": "ch",
+          "consonant_length": 0.08417012542486191,
+          "pitch": 5.978631973266602,
+          "text": "チ",
+          "vowel": "i",
+          "vowel_length": 0.07654783874750137
+        },
+        {
+          "consonant": "w",
+          "consonant_length": 0.07711347192525864,
+          "pitch": 6.040231227874756,
+          "text": "ワ",
+          "vowel": "a",
+          "vowel_length": 0.1825464963912964
+        }
+      ],
+      "pause_mora": null
+    }
+  ],
+  "intonationScale": 1.0,
+  "kana": "コンニチワ'",
+  "outputSamplingRate": 24000,
+  "outputStereo": false,
+  "pitchScale": 0.0,
+  "postPhonemeLength": 0.10000000149011612,
+  "prePhonemeLength": 0.10000000149011612,
+  "speedScale": 1.0,
+  "volumeScale": 1.0
+}

tests/fixtures/audio_query_slow.json

@@ -0,0 +1,355 @@
+{
+  "accent_phrases": [
+    {
+      "accent": 1,
+      "is_interrogative": false,
+      "moras": [
+        {
+          "consonant": "ky",
+          "consonant_length": 0.1269165426492691,
+          "pitch": 5.891244888305664,
+          "text": "キョ",
+          "vowel": "o",
+          "vowel_length": 0.10178578644990921
+        },
+        {
+          "consonant": null,
+          "consonant_length": null,
+          "pitch": 5.944635391235352,
+          "text": "オ",
+          "vowel": "o",
+          "vowel_length": 0.08867011219263077
+        },
+        {
+          "consonant": "w",
+          "consonant_length": 0.06420903652906418,
+          "pitch": 5.655149459838867,
+          "text": "ワ",
+          "vowel": "a",
+          "vowel_length": 0.13292233645915985
+        }
+      ],
+      "pause_mora": null
+    },
+    {
+      "accent": 1,
+      "is_interrogative": false,
+      "moras": [
+        {
+          "consonant": null,
+          "consonant_length": null,
+          "pitch": 5.756601810455322,
+          "text": "イ",
+          "vowel": "i",
+          "vowel_length": 0.09258951991796494
+        },
+        {
+          "consonant": null,
+          "consonant_length": null,
+          "pitch": 6.103951930999756,
+          "text": "イ",
+          "vowel": "i",
+          "vowel_length": 0.0974084809422493
+        }
+      ],
+      "pause_mora": null
+    },
+    {
+      "accent": 1,
+      "is_interrogative": false,
+      "moras": [
+        {
+          "consonant": "t",
+          "consonant_length": 0.0657225176692009,
+          "pitch": 6.139893531799316,
+          "text": "テ",
+          "vowel": "e",
+          "vowel_length": 0.1257137954235077
+        },
+        {
+          "consonant": null,
+          "consonant_length": null,
+          "pitch": 6.0533294677734375,
+          "text": "ン",
+          "vowel": "N",
+          "vowel_length": 0.07083306461572647
+        },
+        {
+          "consonant": "k",
+          "consonant_length": 0.05886319279670715,
+          "pitch": 5.938107967376709,
+          "text": "キ",
+          "vowel": "i",
+          "vowel_length": 0.059421248733997345
+        },
+        {
+          "consonant": "d",
+          "consonant_length": 0.05466592311859131,
+          "pitch": 5.8197479248046875,
+          "text": "デ",
+          "vowel": "e",
+          "vowel_length": 0.09957551956176758
+        },
+        {
+          "consonant": "s",
+          "consonant_length": 0.03665236383676529,
+          "pitch": 0.0,
+          "text": "ス",
+          "vowel": "U",
+          "vowel_length": 0.06477490812540054
+        },
+        {
+          "consonant": "k",
+          "consonant_length": 0.07497167587280273,
+          "pitch": 5.6952314376831055,
+          "text": "カ",
+          "vowel": "a",
+          "vowel_length": 0.09379018098115921
+        },
+        {
+          "consonant": "r",
+          "consonant_length": 0.03401697427034378,
+          "pitch": 5.5977702140808105,
+          "text": "ラ",
+          "vowel": "a",
+          "vowel_length": 0.14935533702373505
+        }
+      ],
+      "pause_mora": {
+        "consonant": null,
+        "consonant_length": null,
+        "pitch": 0.0,
+        "text": "、",
+        "vowel": "pau",
+        "vowel_length": 0.2590501606464386
+      }
+    },
+    {
+      "accent": 5,
+      "is_interrogative": false,
+      "moras": [
+        {
+          "consonant": "k",
+          "consonant_length": 0.08955883234739304,
+          "pitch": 5.693568229675293,
+          "text": "コ",
+          "vowel": "o",
+          "vowel_length": 0.09677395969629288
+        },
+        {
+          "consonant": null,
+          "consonant_length": null,
+          "pitch": 5.938162803649902,
+          "text": "オ",
+          "vowel": "o",
+          "vowel_length": 0.12056692689657211
+        },
+        {
+          "consonant": null,
+          "consonant_length": null,
+          "pitch": 6.016718864440918,
+          "text": "エ",
+          "vowel": "e",
+          "vowel_length": 0.147250235080719
+        },
+        {
+          "consonant": null,
+          "consonant_length": null,
+          "pitch": 6.010623931884766,
+          "text": "ン",
+          "vowel": "N",
+          "vowel_length": 0.08805551379919052
+        },
+        {
+          "consonant": null,
+          "consonant_length": null,
+          "pitch": 6.007989883422852,
+          "text": "オ",
+          "vowel": "o",
+          "vowel_length": 0.10105019062757492
+        }
+      ],
+      "pause_mora": null
+    },
+    {
+      "accent": 3,
+      "is_interrogative": false,
+      "moras": [
+        {
+          "consonant": "s",
+          "consonant_length": 0.0925176665186882,
+          "pitch": 6.003979206085205,
+          "text": "サ",
+          "vowel": "a",
+          "vowel_length": 0.1288621425628662
+        },
+        {
+          "consonant": null,
+          "consonant_length": null,
+          "pitch": 6.026742935180664,
+          "text": "ン",
+          "vowel": "N",
+          "vowel_length": 0.0830387994647026
+        },
+        {
+          "consonant": "p",
+          "consonant_length": 0.054095521569252014,
+          "pitch": 6.099804878234863,
+          "text": "ポ",
+          "vowel": "o",
+          "vowel_length": 0.1004747673869133
+        }
+      ],
+      "pause_mora": null
+    },
+    {
+      "accent": 4,
+      "is_interrogative": false,
+      "moras": [
+        {
+          "consonant": "sh",
+          "consonant_length": 0.021739566698670387,
+          "pitch": 0.0,
+          "text": "シ",
+          "vowel": "I",
+          "vowel_length": 0.04718482494354248
+        },
+        {
+          "consonant": "t",
+          "consonant_length": 0.05912453681230545,
+          "pitch": 6.083391189575195,
+          "text": "テ",
+          "vowel": "e",
+          "vowel_length": 0.08641223609447479
+        },
+        {
+          "consonant": "k",
+          "consonant_length": 0.05805716663599014,
+          "pitch": 6.061047554016113,
+          "text": "カ",
+          "vowel": "a",
+          "vowel_length": 0.0868045762181282
+        },
+        {
+          "consonant": "r",
+          "consonant_length": 0.03226054459810257,
+          "pitch": 5.876008987426758,
+          "text": "ラ",
+          "vowel": "a",
+          "vowel_length": 0.14383982121944427
+        }
+      ],
+      "pause_mora": {
+        "consonant": null,
+        "consonant_length": null,
+        "pitch": 0.0,
+        "text": "、",
+        "vowel": "pau",
+        "vowel_length": 0.2888171374797821
+      }
+    },
+    {
+      "accent": 5,
+      "is_interrogative": false,
+      "moras": [
+        {
+          "consonant": "k",
+          "consonant_length": 0.0854499563574791,
+          "pitch": 5.536592483520508,
+          "text": "カ",
+          "vowel": "a",
+          "vowel_length": 0.12839564681053162
+        },
+        {
+          "consonant": null,
+          "consonant_length": null,
+          "pitch": 5.821318626403809,
+          "text": "イ",
+          "vowel": "i",
+          "vowel_length": 0.07545183598995209
+        },
+        {
+          "consonant": "m",
+          "consonant_length": 0.05912942439317703,
+          "pitch": 6.023046493530273,
+          "text": "モ",
+          "vowel": "o",
+          "vowel_length": 0.0851563960313797
+        },
+        {
+          "consonant": "n",
+          "consonant_length": 0.05413999408483505,
+          "pitch": 6.078396797180176,
+          "text": "ノ",
+          "vowel": "o",
+          "vowel_length": 0.08871601521968842
+        },
+        {
+          "consonant": "n",
+          "consonant_length": 0.07090981304645538,
+          "pitch": 6.0716552734375,
+          "text": "ニ",
+          "vowel": "i",
+          "vowel_length": 0.0862032100558281
+        }
+      ],
+      "pause_mora": null
+    },
+    {
+      "accent": 3,
+      "is_interrogative": false,
+      "moras": [
+        {
+          "consonant": null,
+          "consonant_length": null,
+          "pitch": 6.005620956420898,
+          "text": "イ",
+          "vowel": "i",
+          "vowel_length": 0.09876994043588638
+        },
+        {
+          "consonant": "k",
+          "consonant_length": 0.07177173346281052,
+          "pitch": 6.0334930419921875,
+          "text": "キ",
+          "vowel": "i",
+          "vowel_length": 0.057893283665180206
+        },
+        {
+          "consonant": "m",
+          "consonant_length": 0.058485276997089386,
+          "pitch": 6.074601173400879,
+          "text": "マ",
+          "vowel": "a",
+          "vowel_length": 0.09547590464353561
+        },
+        {
+          "consonant": "sh",
+          "consonant_length": 0.030789220705628395,
+          "pitch": 0.0,
+          "text": "シ",
+          "vowel": "I",
+          "vowel_length": 0.06114967167377472
+        },
+        {
+          "consonant": "t",
+          "consonant_length": 0.05993541330099106,
+          "pitch": 5.849719524383545,
+          "text": "タ",
+          "vowel": "a",
+          "vowel_length": 0.1888434886932373
+        }
+      ],
+      "pause_mora": null
+    }
+  ],
+  "intonationScale": 1.0,
+  "kana": "キョ'オワ/イ'イ/テ'ンキデ_スカラ、コオエンオ'/サンポ'/_シテカラ'、カイモノニ'/イキマ'_シタ",
+  "outputSamplingRate": 24000,
+  "outputStereo": false,
+  "pitchScale": 0.0,
+  "postPhonemeLength": 0.10000000149011612,
+  "prePhonemeLength": 0.10000000149011612,
+  "speedScale": 0.75,
+  "volumeScale": 1.0
+}

tests/fixtures/make_synth_fixtures.py

@@ -0,0 +1,218 @@
+"""Regenerate the committed VOICEVOX synthesis fixtures.
+
+    uv run --extra voice python tests/fixtures/make_synth_fixtures.py
+
+This is a **maintainer action**, not part of CI. The outputs are the ground truth for every
+AVTR-02 assertion in plan 01-06, so every duration here is read from the WAV header and never
+summed from the query - a summed query is the very thing the no-drift test exists to check.
+
+Three cases, chosen deliberately:
+
+===== ======================= ===== ==============================================================
+Case  Text                    Speed Why this text
+===== ======================= ===== ==============================================================
+short こんにちは              1.0   Contains ``N`` (ん) between vowels - exercises the closed-mouth
+                                    path. Doubles as the push-to-talk fixture for VOIC-02.
+long  (see LONG_TEXT)         1.0   です devoices to ``d e s U`` and した to ``sh I t a``, so the
+                                    fixture exercises the uppercase devoiced vowels that a
+                                    lowercase-only lookup silently drops. ~36 moras, long enough
+                                    that float-accumulation drift would be plainly visible.
+slow  same as long            0.75  VOIC-03's mechanism; the timeline must be exactly 1/0.75x
+                                    longer, pre/post silence included.
+===== ======================= ===== ==============================================================
+
+The generator refuses to write a ``long`` fixture that does not actually contain a devoiced vowel,
+a pause mora and 20+ moras. A fixture that does not exercise those branches is the wrong fixture,
+and finding that out here is much cheaper than finding it out in plan 01-06.
+"""
+
+from __future__ import annotations
+
+import datetime as dt
+import json
+import sys
+import wave
+from pathlib import Path
+
+FIXTURES = Path(__file__).resolve().parent
+REPO_ROOT = FIXTURES.parents[1]
+sys.path.insert(0, str(REPO_ROOT / "src"))
+
+SHORT_TEXT = "こんにちは"
+LONG_TEXT = "今日はいい天気ですから、公園を散歩してから、買い物に行きました。"
+
+VOICEVOX_CORE_VERSION = "0.17.0"
+VOICEVOX_VVM_VERSION = "0.17.0"
+
+DEVOICED_VOWELS = frozenset("AIUEO")
+MIN_LONG_MORAS = 20
+
+#: 24000 Hz / 256 samples. Every phoneme is quantised to a whole number of these.
+FRAMERATE = 93.75
+
+CASES = [
+    ("short", SHORT_TEXT, 1.0, "speech_ja.wav"),
+    ("long", LONG_TEXT, 1.0, "speech_ja_long.wav"),
+    ("slow", LONG_TEXT, 0.75, "speech_ja_slow.wav"),
+]
+
+
+def flatten_moras(query: dict) -> list[dict]:
+    """Every mora in utterance order, each accent phrase's pause mora following its moras."""
+    moras: list[dict] = []
+    for phrase in query["accent_phrases"]:
+        moras.extend(phrase["moras"])
+        if phrase["pause_mora"]:
+            moras.append(phrase["pause_mora"])
+    return moras
+
+
+def phoneme_lengths(query: dict) -> list[float]:
+    """Unscaled phoneme durations in seconds, wrapped in the pre/post silence, in order."""
+    silence = {"vowel": "pau", "consonant": None, "consonant_length": None}
+    moras = [
+        {**silence, "vowel_length": query["prePhonemeLength"]},
+        *flatten_moras(query),
+        {**silence, "vowel_length": query["postPhonemeLength"]},
+    ]
+    lengths: list[float] = []
+    for mora in moras:
+        if mora["consonant"] is not None:
+            lengths.append(mora["consonant_length"])
+        lengths.append(mora["vowel_length"])
+    return lengths
+
+
+def write_wav(path: Path, wav_bytes: bytes) -> tuple[float, int]:
+    """Write the WAV; return its true duration and frame count from the header just written.
+
+    ``frame_count`` is in VOICEVOX frames (256 samples at 24000 Hz = 93.75 fps), not PCM samples,
+    because that is the unit plan 01-06's +/-1 frame no-drift tolerance is expressed in.
+    """
+    path.write_bytes(wav_bytes)
+    with wave.open(str(path), "rb") as handle:
+        assert handle.getframerate() == 24000, handle.getframerate()
+        assert handle.getnchannels() == 1, handle.getnchannels()
+        assert handle.getsampwidth() == 2, handle.getsampwidth()
+        samples = handle.getnframes()
+        assert samples % 256 == 0, samples
+        return samples / float(handle.getframerate()), samples // 256
+
+
+def main() -> int:
+    try:
+        import voicevox_core  # noqa: F401
+    except ImportError:
+        print(
+            "voicevox_core is not installed, so the fixtures cannot be regenerated.\n"
+            "Install it with `uv sync --extra dev --extra voice`; see docs/VOICEVOX-SETUP.md.\n"
+            "Regeneration is a maintainer action - CI does not need it.",
+            file=sys.stderr,
+        )
+        return 1
+
+    from japanese_avatar.voice.tts import SPEAKER_STYLE_ID, synthesize
+
+    meta: dict = {
+        "generated": dt.datetime.now(dt.UTC).date().isoformat(),
+        "voicevox_core": VOICEVOX_CORE_VERSION,
+        "voicevox_vvm": VOICEVOX_VVM_VERSION,
+        "style_id": SPEAKER_STYLE_ID,
+        "cases": {},
+    }
+
+    for case, text, speed, wav_name in CASES:
+        result = synthesize(text, speed=speed)
+        query = result.audio_query
+
+        query_name = f"audio_query_{case}.json"
+        (FIXTURES / query_name).write_text(
+            json.dumps(query, indent=2, ensure_ascii=False, sort_keys=True) + "\n",
+            encoding="utf-8",
+            newline="\n",
+        )
+
+        duration, frame_count = write_wav(FIXTURES / wav_name, result.wav_bytes)
+        # The WAV we just wrote must agree with what synthesize() reported, bit for bit.
+        assert duration == result.duration, (case, duration, result.duration)
+
+        # Ground truth for plan 01-06. See the "Frame quantisation" section of
+        # docs/VOICEVOX-SETUP.md: quantise at speed 1.0 FIRST, then divide the frame count by
+        # speedScale and round again. Dividing the length by speedScale before quantising is a
+        # different, wrong answer that only coincides at speed 1.0.
+        predicted = sum(
+            round(round(length * FRAMERATE) / speed) for length in phoneme_lengths(query)
+        )
+        assert predicted == frame_count, (case, predicted, frame_count)
+
+        moras = flatten_moras(query)
+        vowels = [m["vowel"] for m in moras]
+        pause_moras = sum(1 for p in query["accent_phrases"] if p["pause_mora"])
+
+        print(
+            f"{case:>5}: moras={len(moras):<3} pause_moras={pause_moras} frames={frame_count:<5} "
+            f"duration={duration!r} vowels={sorted(set(vowels))}"
+        )
+
+        if case == "long":
+            if len(moras) < MIN_LONG_MORAS:
+                print(
+                    f"FAIL: the long fixture has {len(moras)} moras, fewer than "
+                    f"{MIN_LONG_MORAS}. Pick a longer sentence and record which one.",
+                    file=sys.stderr,
+                )
+                return 1
+            devoiced = sorted(set(vowels) & DEVOICED_VOWELS)
+            if not devoiced:
+                print(
+                    "FAIL: the long fixture contains no devoiced vowel (A/I/U/E/O). Its entire "
+                    "job is to exercise that branch - pick a different sentence.",
+                    file=sys.stderr,
+                )
+                return 1
+            if not pause_moras:
+                print(
+                    "FAIL: the long fixture produced no pause_mora. Pick a sentence with a "
+                    "comma so the pause-ordering branch is covered.",
+                    file=sys.stderr,
+                )
+                return 1
+            print(f"       devoiced vowels present: {devoiced}")
+
+        meta["cases"][case] = {
+            "text": text,
+            "speed_scale": speed,
+            "query": query_name,
+            "wav": wav_name,
+            "duration_seconds": duration,
+            "frame_count": frame_count,
+            "mora_count": len(moras),
+            "pause_mora_count": pause_moras,
+            "vowel_symbols": vowels,
+        }
+
+    # Not exactly 1/0.75: VOICEVOX re-quantises every phoneme after scaling, so the realised
+    # ratio lands within a fraction of a percent of it rather than on it. 2% is the tolerance
+    # the fixture contract is written against.
+    ratio = meta["cases"]["slow"]["duration_seconds"] / meta["cases"]["long"]["duration_seconds"]
+    print(f"slow/long duration ratio = {ratio:.6f} (expected ~{1 / 0.75:.6f})")
+    if abs(ratio - 1 / 0.75) >= 0.02:
+        print(
+            f"FAIL: speedScale=0.75 did not lengthen the audio by ~1/0.75x: {ratio}",
+            file=sys.stderr,
+        )
+        return 1
+
+    (FIXTURES / "synth_meta.json").write_text(
+        json.dumps(meta, indent=2, ensure_ascii=False, sort_keys=True) + "\n",
+        encoding="utf-8",
+        # Explicit LF: the default translates to CRLF on Windows, which would make the committed
+        # fixtures differ byte-for-byte depending on which OS last regenerated them.
+        newline="\n",
+    )
+    print(f"wrote {FIXTURES / 'synth_meta.json'}")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

tests/fixtures/speech_ja.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:726ee8497999a5aac84567edbc53fb6751529348520bb1643c5a98a548e9f1fc
+size 50732

tests/fixtures/speech_ja_long.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ccbd3f1d0cd7714e1d7ed12518978087be6c251992c9c5c59600742d45291160
+size 264236

tests/fixtures/speech_ja_slow.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c91341fd06952c7e2bab243dffc77cfa4d1db36c636ddc0e74999b309238d544
+size 354348

tests/fixtures/synth_meta.json

@@ -0,0 +1,121 @@
+{
+  "cases": {
+    "long": {
+      "duration_seconds": 5.504,
+      "frame_count": 516,
+      "mora_count": 36,
+      "pause_mora_count": 2,
+      "query": "audio_query_long.json",
+      "speed_scale": 1.0,
+      "text": "今日はいい天気ですから、公園を散歩してから、買い物に行きました。",
+      "vowel_symbols": [
+        "o",
+        "o",
+        "a",
+        "i",
+        "i",
+        "e",
+        "N",
+        "i",
+        "e",
+        "U",
+        "a",
+        "a",
+        "pau",
+        "o",
+        "o",
+        "e",
+        "N",
+        "o",
+        "a",
+        "N",
+        "o",
+        "I",
+        "e",
+        "a",
+        "a",
+        "pau",
+        "a",
+        "i",
+        "o",
+        "o",
+        "i",
+        "i",
+        "i",
+        "a",
+        "I",
+        "a"
+      ],
+      "wav": "speech_ja_long.wav"
+    },
+    "short": {
+      "duration_seconds": 1.056,
+      "frame_count": 99,
+      "mora_count": 5,
+      "pause_mora_count": 0,
+      "query": "audio_query_short.json",
+      "speed_scale": 1.0,
+      "text": "こんにちは",
+      "vowel_symbols": [
+        "o",
+        "N",
+        "i",
+        "i",
+        "a"
+      ],
+      "wav": "speech_ja.wav"
+    },
+    "slow": {
+      "duration_seconds": 7.381333333333333,
+      "frame_count": 692,
+      "mora_count": 36,
+      "pause_mora_count": 2,
+      "query": "audio_query_slow.json",
+      "speed_scale": 0.75,
+      "text": "今日はいい天気ですから、公園を散歩してから、買い物に行きました。",
+      "vowel_symbols": [
+        "o",
+        "o",
+        "a",
+        "i",
+        "i",
+        "e",
+        "N",
+        "i",
+        "e",
+        "U",
+        "a",
+        "a",
+        "pau",
+        "o",
+        "o",
+        "e",
+        "N",
+        "o",
+        "a",
+        "N",
+        "o",
+        "I",
+        "e",
+        "a",
+        "a",
+        "pau",
+        "a",
+        "i",
+        "o",
+        "o",
+        "i",
+        "i",
+        "i",
+        "a",
+        "I",
+        "a"
+      ],
+      "wav": "speech_ja_slow.wav"
+    }
+  },
+  "generated": "2026-08-27",
+  "style_id": 3,
+  "voicevox_core": "0.17.0",
+  "voicevox_vvm": "0.17.0"
+}

tests/test_tts_contract.py

@@ -23,8 +23,27 @@ LONG_TEXT = "今日はいい天気ですから、公園を散歩してから、
 
 
 @pytest.fixture(scope="module")
-def short_result():
-    return synthesize(SHORT_TEXT)
+def synth():
+    """Memoised synthesis, keyed by (text, speed).
+
+    The engine is deterministic, so synthesising the same pair twice is pure waste - and this
+    module would otherwise do it six times over, pushing the quick loop past the 15 s feedback
+    budget in ``01-VALIDATION.md``. Three distinct syntheses cover every test below.
+    """
+    cache: dict[tuple[str, float], object] = {}
+
+    def _synth(text: str, speed: float = 1.0):
+        key = (text, speed)
+        if key not in cache:
+            cache[key] = synthesize(text, speed=speed)
+        return cache[key]
+
+    return _synth
+
+
+@pytest.fixture(scope="module")
+def short_result(synth):
+    return synth(SHORT_TEXT)
 
 
 def test_synthesize_returns_nonempty_wav(short_result):
@@ -49,13 +68,13 @@ def test_audio_query_is_plain_json_roundtrippable(short_result):
         assert key in query, key
 
 
-def test_moras_carry_timings(short_result):
+def test_moras_carry_timings(synth):
     """Per-mora timings exist, and consonant/consonant_length are genuinely optional.
 
     The optionality is asserted rather than assumed: plan 01-06's timeline builder must not index
     into a key that is absent for every vowel-only mora (ア) and for every pause.
     """
-    result = synthesize(LONG_TEXT)
+    result = synth(LONG_TEXT)
     moras = []
     for phrase in result.audio_query["accent_phrases"]:
         moras.extend(phrase["moras"])
@@ -74,9 +93,9 @@ def test_moras_carry_timings(short_result):
     assert without_consonant, "expected at least one mora with no consonant at all"
 
 
-def test_speed_scale_lengthens_audio():
-    normal = synthesize(LONG_TEXT, speed=1.0)
-    slow = synthesize(LONG_TEXT, speed=0.75)
+def test_speed_scale_lengthens_audio(synth):
+    normal = synth(LONG_TEXT, speed=1.0)
+    slow = synth(LONG_TEXT, speed=0.75)
     assert slow.duration > normal.duration
     ratio = slow.duration / normal.duration
     assert abs(ratio - 1 / 0.75) < 0.02 * (1 / 0.75), ratio
@@ -89,6 +108,18 @@ def test_output_sampling_rate_is_24000(short_result):
     assert SPEAKER_STYLE_ID == 3
 
 
+def test_fixtures_match_current_engine(synth_meta, synth):
+    """Regression guard: if a voicevox_core bump changes timings, this fails loudly
+    rather than test_golden_timeline failing mysteriously in plan 01-06."""
+    for case, meta in synth_meta["cases"].items():
+        r = synth(meta["text"], speed=meta["speed_scale"])
+        assert abs(r.duration - meta["duration_seconds"]) < 0.011, (
+            f"{case}: engine now produces {r.duration}s vs recorded {meta['duration_seconds']}s. "
+            "Regenerate with tests/fixtures/make_synth_fixtures.py and re-review the "
+            "golden timeline."
+        )
+
+
 def test_no_gpu_imports_on_synthesis_path():
     modules = sorted(VOICE_PKG.glob("*.py"))
     assert modules, f"no modules found under {VOICE_PKG}"