fix: preserve upload recovery and harden speech indexing

tests/test_audio_index.py

@@ -16,9 +16,32 @@ from imagefind.jobs import JobRetry
 from imagefind.main import create_app
 from imagefind.media import MediaInput
 from imagefind.speech import SPEECH_INDEX_REVISION, SpeechService, SpeechStageError
+from imagefind.speech_quality import aggregate_transcript_quality
 from imagefind.text import fts_query
 
 
+def test_ending_hallucination_is_risky_twice_or_once_in_weak_speech():
+    repeated = aggregate_transcript_quality([
+        {"text": "拜拜"},
+        {"text": "正常对话"},
+        {"text": "拜拜"},
+    ])
+    assert "whole_ending_hallucination" in repeated.flags
+    assert ("拜拜", 2) in repeated.repeated_phrases
+
+    weak = aggregate_transcript_quality([
+        {"text": "谢谢大家收看", "speech_ratio": 0.1},
+        {"text": "正常对话", "speech_ratio": 0.8},
+    ])
+    assert "whole_ending_hallucination" in weak.flags
+
+    single = aggregate_transcript_quality([
+        {"text": "拜拜", "speech_ratio": 0.8},
+        {"text": "正常对话", "speech_ratio": 0.8},
+    ])
+    assert "whole_ending_hallucination" not in single.flags
+
+
 def _app(tmp_path: Path):
     settings = Settings(data_dir=tmp_path / "data", embedding_backend="hash", upload_reserve_gb=0)
     settings.prepare()