fix(audio): detect transcript-wide hallucinations

tests/test_audio_index.py

@@ -120,6 +120,42 @@ def test_audio_indexer_writes_timed_fts_entries_and_reconciles(tmp_path: Path, m
     assert json.loads(job["payload_json"]) == {"video_id": "video"}
 
 
+def test_repeated_hallucinations_are_saved_but_not_searchable(tmp_path: Path, monkeypatch):
+    app = _app(tmp_path)
+    service = app.state.services
+    monkeypatch.setattr(service.media, "input_for", lambda _video: MediaInput("movie.mp4"))
+    monkeypatch.setattr(
+        service.media,
+        "probe",
+        lambda _media: {"raw": {"streams": [{"codec_type": "audio"}]}, "duration_ms": 30_000},
+    )
+    monkeypatch.setattr(
+        service.speech,
+        "transcribe",
+        lambda *_args, **_kwargs: [
+            {"text": "拜拜", "start_ms": index * 1000, "end_ms": index * 1000 + 800}
+            for index in range(8)
+        ],
+    )
+    job_id = service.jobs.enqueue("transcribe_audio", {"video_id": "video"}, dedupe_key="audio:video")
+    service.audio_indexer.index(job_id, "video", quality_repair=True)
+    with service.db.read() as conn:
+        entries = conn.execute(
+            "SELECT count(*) FROM text_entries WHERE video_id='video' AND kind='audio'"
+        ).fetchone()[0]
+        searchable = conn.execute(
+            "SELECT count(*) FROM text_fts WHERE entry_id IN "
+            "(SELECT id FROM text_entries WHERE video_id='video' AND kind='audio')"
+        ).fetchone()[0]
+        video = conn.execute(
+            "SELECT audio_quality_score,audio_quality_flags_json FROM videos WHERE id='video'"
+        ).fetchone()
+    assert entries == 8
+    assert searchable == 0
+    assert video["audio_quality_score"] < 0.7
+    assert "whole_ending_hallucination" in json.loads(video["audio_quality_flags_json"])
+
+
 def test_audio_indexer_retries_medium_memory_pressure(tmp_path: Path, monkeypatch):
     app = _app(tmp_path)
     service = app.state.services