tools(fw-diag): add G_PROD_CL30 + H_PROD_NOCL chunk_length isolation variants

Adds two new PARAM_VARIANTS to fw_diagnostic_suite.py for isolating
the chunk_length=30 parameter effect on Faster-Whisper transcription.

Background: during v1.8.13 acceptance testing, a hypothesis emerged
that chunk_length=30 (set only on the aggressive preset since the
2025-12-19 ctranslate2 divide-by-zero crash fix, commit 9dbbdb1)
might be contributing to the catastrophic empty output on JAV
content. The existing 8 variants (A1, A2, A, B, C, D, E, F) all
hard-coded chunk_length=30 in _base_params(), so the without-chunk_length
case had never been measured.

New variants:

  G_PROD_CL30 — v1.8.13 production aggressive preset values verbatim
                (faster_whisper.py:304-343), with chunk_length=30.
                Acts as the baseline for the isolation test.

  H_PROD_NOCL — IDENTICAL to G_PROD_CL30 except chunk_length=None.
                Tests whether removing chunk_length helps or hurts.

Both use the documented v1.8.13 production aggressive params:
  beam_size=3, best_of=2, patience=2.0
  temperature=[0.0, 0.17]
  compression_ratio_threshold=2.6
  log_prob_threshold=-1.0
  no_speech_threshold=0.84
  repetition_penalty=1.3, no_repeat_ngram_size=3

Empirical result (whisperseg + large-v2, same audio):
  G_PROD_CL30 (chunk_length=30):   33 segments / 12 of 21 groups (42.86% empty)
  H_PROD_NOCL (chunk_length=None):  0 segments /  0 of 21 groups (100%   empty)

Conclusion: chunk_length=30 is NOT a bug. Removing it produces 100%
empty output. The v1.8.10 ctranslate2 crash fix value is correct
as-is. This empirical result is preserved in this diagnostic
infrastructure so future investigations can re-run quickly.

Side benefit: G_PROD_CL30's success (33 segments) at large-v2 led
directly to the discovery that the actual v1.8.13 regression was
the model version (large-v3 default), not chunk_length. See
fix(asr): revert default model commit (ad22a6e).

Test artifacts: test_media/1813 acceptance/F4/DIAG_FW/chunk_length_test/

File: tools/fw_diagnostic_suite.py (PARAM_VARIANTS dict additions
after Run F)

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

Author: meizhong986

tools/fw_diagnostic_suite.py

@@ -394,6 +394,41 @@ def _base_params(language: str = DEFAULT_LANGUAGE) -> Dict[str, Any]:
         "no_speech_threshold": 0.6,
         "compression_ratio_threshold": 2.4,
     },
+    # Run G_PROD_CL30 — v1.8.13 production aggressive preset EXACTLY,
+    # with chunk_length=30 (current production value). This is the baseline
+    # for the chunk_length isolation test (compare against H_PROD_NOCL).
+    # Source: whisperjav/config/components/asr/faster_whisper.py:304-343
+    # (FasterWhisperOptions["aggressive"]).
+    "G_PROD_CL30": {
+        **_base_params(),
+        "beam_size": 3,                            # v1.8.12 retune
+        "best_of": 2,                              # v1.8.12.post1 fix
+        "patience": 2.0,
+        "temperature": [0.0, 0.17],                # v1.8.10-hf3 light fallback
+        "compression_ratio_threshold": 2.6,
+        "log_prob_threshold": -1.0,
+        "no_speech_threshold": 0.84,               # v1.8.12 retune
+        "repetition_penalty": 1.3,
+        "no_repeat_ngram_size": 3,
+        "chunk_length": 30,                        # v1.8.10 ctranslate2 crash fix (this variant)
+    },
+    # Run H_PROD_NOCL — IDENTICAL to G_PROD_CL30 except chunk_length=None.
+    # Tests the user's hypothesis that chunk_length=30 contributes to the
+    # empty-output pathology. If H produces meaningfully more output than G,
+    # chunk_length is implicated.
+    "H_PROD_NOCL": {
+        **_base_params(),
+        "beam_size": 3,
+        "best_of": 2,
+        "patience": 2.0,
+        "temperature": [0.0, 0.17],
+        "compression_ratio_threshold": 2.6,
+        "log_prob_threshold": -1.0,
+        "no_speech_threshold": 0.84,
+        "repetition_penalty": 1.3,
+        "no_repeat_ngram_size": 3,
+        "chunk_length": None,                      # ← THE TEST: remove chunk_length
+    },
 }