docs: v1.8.13 release notes — model revert + scoped WhisperSeg default

Updates the v1.8.13 release notes to honestly reflect the post-F4/F6/F7
acceptance testing findings:

Title and theme:
  - Adds "Model Default Revert" to the title alongside the WhisperSeg
    Ensemble/Qwen/Decoupled scoping.
  - Theme paragraph explicitly mentions both the WhisperSeg scope
    reduction and the large-v3 → large-v2 revert.

New "Mode-by-mode default policy" table clarifies which path gets
which segmenter default (Ensemble/Qwen/Decoupled get whisperseg;
simple modes get silero-v3.1) and links to the Known caveats.

Known caveats section gains two entries:

  1. Default Whisper model reverted from large-v3 to large-v2 — full
     empirical context: F4/F6/F7 catastrophic (6-10 entries / 68 GT)
     vs F8 with large-v2 (51 entries / 68 GT). Root cause: v1.8.12
     aggressive preset retune was tuned against large-v2 forensic
     acceptance data. Users wanting v3 can opt in via --model
     large-v3. References the diagnostic suite test artifacts at
     test_media/1813 acceptance/F4/DIAG_FW/chunk_length_test/.

  2. WhisperSeg in simple Transcription Mode deferred to v1.9.0 —
     same explanation as before, retained for context.

Internal changes section gains a "Default model revert" sub-bullet
listing all three modified file locations:
  - config/components/asr/faster_whisper.py:211
  - config/components/asr/stable_ts.py:293
  - webview_gui/assets/index.html:299

v1.9.0 follow-up section adds a new marquee item:
  - Re-tune aggressive sensitivity preset for large-v3 (lifts the
    v1.8.13 model revert). Plan: re-run forensic acceptance suite
    against large-v3, produce per-engine, per-model preset variants.

User-facing how-to-opt-into sections updated:
  - "How to opt into WhisperSeg in simple mode" → use --ensemble
  - "When NOT to use WhisperSeg" — non-Japanese audio, CPU-bound,
    v1.8.12 retest cluster

File: docs/release_notes_v1.8.13.md

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: meizhong986

docs/release_notes_v1.8.13.md

@@ -1,49 +1,84 @@
-# WhisperJAV v1.8.13 — Polish + WhisperSeg Default
+# WhisperJAV v1.8.13 — Polish + WhisperSeg Default (Ensemble / Qwen / Decoupled) + Model Default Revert
 
 > Draft — to be finalized before tag.
 
 ## Theme
 
 A focused polish release that addresses long-standing paper cuts, fixes
-the install regression carried over from v1.8.11, and **promotes WhisperSeg
-from opt-in to default** speech segmenter across all surfaces.
+the install regression carried over from v1.8.11, **promotes WhisperSeg
+to the default speech segmenter on the pipeline paths that fully support
+it** (Ensemble, Qwen, Decoupled), and **reverts the default Whisper model
+from `large-v3` back to `large-v2`** for both the FasterWhisper (BalancedPipeline)
+and Stable-TS (Fast/Faster pipelines) backends, restoring transcription
+quality on continuous-energy non-phonetic JAV content. Simple Transcription
+Mode keeps silero-v3.1 in v1.8.13, pending a config-routing refactor
+scheduled for v1.9.0. See "Known caveats" for full context on the model
+revert.
 
 ---
 
-## Headline change: WhisperSeg is now the default speech segmenter
+## Headline change: WhisperSeg becomes the default for Ensemble / Qwen / Decoupled
 
 Introduced in v1.8.12 as opt-in, **WhisperSeg ONNX is now the default
-speech segmenter** in v1.8.13. On the Netflix-GT 283s JAV reference
+speech segmenter** for Ensemble Mode (pass1 and pass2), the Qwen pipeline,
+and the Decoupled pipeline. On the Netflix-GT 283s JAV reference
 benchmark it scored F1=0.787 — the strongest result among the four
 shipped backends:
 
 | Segmenter | F1 (Netflix-GT JAV) |
 |---|---:|
-| **WhisperSeg (new default)** | **0.787** |
+| **WhisperSeg (new default for Ensemble/Qwen/Decoupled)** | **0.787** |
 | TEN VAD | 0.698 |
 | Silero v6.2 | 0.654 |
-| Silero v3.1 (previous default) | 0.625 |
+| Silero v3.1 (default for simple Transcription Mode) | 0.625 |
+
+### Mode-by-mode default policy in v1.8.13
+
+| Path | Default segmenter | Why |
+|---|---|---|
+| **GUI Ensemble Mode** (pass1 + pass2) | WhisperSeg | Pass-worker routes grouping params correctly |
+| **`--ensemble` CLI** | WhisperSeg | Same path as above |
+| **`--mode qwen`** (ChronosJAV) | WhisperSeg | QwenPipeline forwards `chunk_threshold_s` / `max_group_duration_s` explicitly |
+| **`--pipeline decoupled`** | WhisperSeg | DecoupledPipeline kwargs path |
+| **GUI Transcription Mode** | **silero-v3.1** | Routes through legacy ASR constructors with a known param-routing bug; fix is v1.9.0 |
+| **`--mode balanced` / `--mode fidelity`** (without `--ensemble`) | **silero-v3.1** | Same — see Known caveats below |
 
 ### What changes for you
 
-- First transcription after upgrade downloads ~119 MB WhisperSeg model
-  from HuggingFace (one time). Cached at `~/.cache/huggingface/` (Linux/macOS)
-  or `%USERPROFILE%\.cache\huggingface\` (Windows).
-- ChronosJAV pipeline's segmenter default also flipped (was silero-v6.2).
-- Colab + Kaggle expert notebook defaults flipped to match.
+- **Ensemble Mode users** get WhisperSeg out of the box on first run. First
+  transcription downloads ~119 MB WhisperSeg model from HuggingFace
+  (one time). Cached at `~/.cache/huggingface/` (Linux/macOS) or
+  `%USERPROFILE%\.cache\huggingface\` (Windows).
+- **Simple Transcription Mode users** see no behavior change vs v1.8.12 —
+  silero-v3.1 stays the default.
+- ChronosJAV (`--mode qwen`) segmenter default flipped to WhisperSeg.
+- Colab + Kaggle expert notebook defaults flipped to match (Ensemble path).
+
+### How to opt into WhisperSeg in simple mode
 
-### When NOT to use WhisperSeg
+Until the v1.9.0 routing fix lands, the recommended path to use WhisperSeg
+is **Ensemble Mode** (which also gives you the dual-pass merging that has
+become WhisperJAV's strongest result on JAV content). In the GUI, switch
+from the "Transcription Mode" tab to the "Ensemble Mode" tab. On the CLI,
+add `--ensemble` plus pass1/pass2 settings.
+
+If you explicitly pass `--speech-segmenter whisperseg` (or `ten`, `nemo`,
+`whisper-vad`) to a simple `--mode balanced` / `--mode fidelity` run,
+v1.8.13 will warn loudly and downgrade to silero-v3.1 to avoid catastrophic
+empty output. The warning includes a pointer to `--ensemble`.
+
+### When NOT to use WhisperSeg (Ensemble path)
 
 - **Non-Japanese audio** (Korean, Chinese, English): WhisperSeg is trained
   on Japanese ASMR. For other languages, override:
-  - **CLI**: `--speech-segmenter silero-v3.1`
+  - **CLI**: `--pass1-speech-segmenter silero-v3.1 --pass2-speech-segmenter silero-v3.1`
   - **GUI Ensemble tab**: pick "Silero v3.1" from the speech segmenter
-    dropdown
+    dropdown for each pass
 - **CPU-bound users**: WhisperSeg is ~2-3× slower than silero/ten on CPU.
   If you have no GPU (or onnxruntime-gpu isn't installed), silero-v3.1
   may produce faster total runtime.
 - **v1.8.12 retest cluster** (#294, #302, #287, #297): if you're validating
-  the v1.8.12 ASR fixes, keep `--speech-segmenter silero-v3.1` for clean
+  the v1.8.12 ASR fixes, keep silero-v3.1 in both passes for clean
   attribution of which fix helped your case.
 
 ---
@@ -122,13 +157,52 @@ Re-open the notebook from the [latest commit](https://github.com/meizhong986/Whi
 
 ## Known caveats
 
-- **Whisperseg field testing is thin**. v1.8.12's WhisperSeg release got
-  zero user reports back. v1.8.13's promotion to default rests on the
-  Netflix-GT bench plus internal smoke tests. If you observe regressions
-  vs v1.8.12, override to silero-v3.1 and report on the issue tracker.
-
-- **Cold-start cost**: first transcription after upgrade pauses for the
-  119 MB WhisperSeg model download. Subsequent runs use the cached model.
+- **Default Whisper model reverted from `large-v3` to `large-v2`.** Late in
+  v1.8.13 acceptance testing (F4/F6/F7 vs F8 side-by-side on the same
+  293-second JAV reference clip), the v1.8.12 aggressive ASR preset retune
+  (`no_speech_threshold=0.84`, `beam_size=3`, `best_of=2`,
+  `temperature=[0.0, 0.17]`, `compression_ratio_threshold=2.6`,
+  `repetition_penalty=1.3`, `no_repeat_ngram_size=3`, `chunk_length=30`)
+  was found to interact pathologically with `large-v3` on continuous-energy
+  non-phonetic content (JAV moaning), producing **6–10 SRT entries out of
+  68 ground-truth** (≈85–90% loss) under simple Transcription Mode. The
+  same audio + same preset values + `large-v2` produced **51 entries**
+  (≈75% capture). Root cause: the v1.8.12 retune was tuned against
+  large-v2 forensic acceptance data; large-v3's slightly different
+  encoder/decoder behavior makes the same gate values too strict on this
+  content distribution. Until the preset is re-tuned for large-v3 in
+  v1.9.x, v1.8.13 reverts the default to large-v2 across both
+  FasterWhisper (`config/components/asr/faster_whisper.py`) and Stable-TS
+  (`config/components/asr/stable_ts.py`) backends. Users who want
+  large-v3 can opt in with `--model large-v3` (CLI) or via the GUI model
+  override checkbox. OpenAI-Whisper backend (FidelityPipeline) was
+  already on large-v2 in v1.8.12 and is unaffected.
+
+- **WhisperSeg in simple Transcription Mode is deferred to v1.9.0.**
+  During F4/F6 acceptance testing, WhisperSeg invoked through the simple
+  `--mode balanced` path produced catastrophic empty output (10 of 68
+  ground-truth subtitles) on JAV moaning content, while the same audio +
+  same backend through Ensemble Mode produced 52/68. Root cause is a
+  config-routing path: BalancedPipeline / FidelityPipeline use a
+  CONSTRUCTOR FIREWALL in their ASR modules that strips backend-agnostic
+  grouping params (`chunk_threshold_s`, `max_group_duration_s`) for
+  non-Silero backends, causing WhisperSeg to fall back to its 29-second
+  default group duration and trigger a Whisper repetition pathology.
+  v1.8.13 ships the SAFE scope (Ensemble + Qwen + Decoupled keep
+  WhisperSeg, simple modes keep silero-v3.1). v1.9.0 lands the proper
+  fix (split SileroVADOptions, introduce SegmenterGroupingOptions,
+  eliminate the firewall pattern).
+
+- **WhisperSeg field testing is thin**. v1.8.12's WhisperSeg release got
+  zero user reports back. v1.8.13's promotion to default for the Ensemble
+  path rests on F5 acceptance test (52/68 GT), the Netflix-GT bench, plus
+  internal smoke tests. If you observe regressions vs v1.8.12, switch
+  back to silero-v3.1 in the Ensemble dropdown and report on the issue
+  tracker.
+
+- **Cold-start cost**: first Ensemble transcription after upgrade pauses
+  for the 119 MB WhisperSeg model download. Subsequent runs use the
+  cached model.
 
 - **Customize Parameters modal**: the Anime-Whisper model dropdown shows
   v0.3 as an option but defaults to litagin/anime-whisper. v0.3 is
@@ -155,14 +229,25 @@ Re-open the notebook from the [latest commit](https://github.com/meizhong986/Whi
 
 ## Internal changes
 
-- WhisperSeg promotion touched 11 locations across resolver, ASR module
+- WhisperSeg promotion touched 11+ locations across resolver, ASR module
   fallbacks, qwen pipeline, CLI argparse help, GUI HTML, decoupled YAML,
-  and Colab + Kaggle notebooks. Architectural note added inline in
-  `legacy.py` explaining why `LEGACY_PIPELINES["balanced"|"fidelity"]["vad"]`
-  field stays at "silero-v3.1" (it names a Pydantic VAD *component* that
-  defines preset values, not the runtime segmenter backend selector — the
-  firewall in `whisper_pro_asr.py:71-77` clears those silero presets at
-  runtime when whisperseg is selected).
+  GUI app.js per-pipeline presets, and Colab + Kaggle notebooks.
+
+- **Default model revert (large-v3 → large-v2)** in three locations:
+  `config/components/asr/faster_whisper.py:211`,
+  `config/components/asr/stable_ts.py:293`, and
+  `webview_gui/assets/index.html:299` (GUI model dropdown default selection).
+  Inline comments reference `faster_whisper.py` for full empirical context.
+  OpenAI-Whisper backend (`config/components/asr/openai_whisper.py:178`)
+  was already on large-v2 and unchanged.
+
+- **v1.8.13 default-flip scope reduction** (post-F4/F6): `main.py`'s
+  `--speech-segmenter` default-resolution gained an explicit allow-list of
+  paths that route segmenter grouping params correctly to non-Silero
+  backends (`--ensemble`, `--pipeline decoupled`, `--mode qwen`). Other
+  paths default to silero-v3.1. Explicit `--speech-segmenter whisperseg`
+  on a non-allow-listed path emits a hard warning and downgrades to
+  silero-v3.1. See inline comment at `main.py:1840-1898`.
 
 - 6 prepared post-release replies in
   `docs/release_v1.8.13_reply_drafts.md` will be posted after the GitHub
@@ -174,6 +259,26 @@ Re-open the notebook from the [latest commit](https://github.com/meizhong986/Whi
 
 Marquee features for v1.9.0 (per `docs/plans/PRODUCT_VISION_AND_ROADMAP_v1.9_v2.md`):
 
+- **Re-tune aggressive sensitivity preset for `large-v3`** (lifts the
+  v1.8.13 model revert). The current v1.8.12 aggressive preset values
+  (`no_speech_threshold=0.84`, `beam_size=3`, etc.) were tuned against
+  large-v2 forensic acceptance data and produce catastrophic empty
+  output on JAV content with large-v3. v1.9.x will re-run the forensic
+  acceptance suite against large-v3 and produce a per-engine, per-model
+  preset variant (e.g., `aggressive_v3` with relaxed gates) so large-v3
+  can return as the default. Reference: F4/F6/F7 vs F8 acceptance test
+  artifacts in `test_media/1813 acceptance/`.
+
+- **Unified segmenter param routing** (lifts the v1.8.13 simple-mode
+  WhisperSeg restriction). Plan: split `SileroVADOptions` into
+  `SileroVADOptions` (Silero-specific) + `SegmenterGroupingOptions`
+  (backend-agnostic: `chunk_threshold_s`, `max_group_duration_s`,
+  `max_speech_duration_s`); resolver places grouping params in a canonical
+  location all consumers read from; eliminate the constructor firewall in
+  `faster_whisper_pro_asr.py` and `whisper_pro_asr.py`. After this,
+  WhisperSeg becomes the default everywhere including simple Transcription
+  Mode. Reference: F4/F6 acceptance test artifacts and the diagnostic
+  suite runs at `test_media/1813 acceptance/F4/DIAG_FW/`.
 - **GUI redesign**: 5→4 tabs, eliminate Advanced, add Utilities tab
 - **Standalone Merge utility** (GUI + `whisperjav-merge` CLI) — preview at
   Section 8.5 of the roadmap doc