Validate audio stream timelines and add genuine Google transcript diagnostics

tools/speech-eval/test_contract.py

@@ -6,11 +6,49 @@ import tempfile
 import sys
 import os
 import wave
-from run import segments, run_owned, recognized_window, decode_command, decoded_frames
-from score import score
+from run import segments, run_owned, recognized_window, decode_command, decoded_frames, primary_audio_range
+from score import score, require_eligible
 
 
 class ContractTests(unittest.TestCase):
+    def test_scoring_rejects_incomplete_or_reduced_pool(self):
+        for report in [{}, {"complete": False, "quality_score_eligible": True},
+                       {"complete": True, "quality_score_eligible": False},
+                       {"complete": True, "quality_score_eligible": True, "excluded_windows": [{}]}]:
+            with self.assertRaises(ValueError):
+                require_eligible(report)
+        require_eligible({"complete": True, "quality_score_eligible": True, "excluded_windows": []})
+    def test_primary_audio_duration_and_nonzero_start_ignore_longer_video(self):
+        probe = {"streams": [{"index": 0, "codec_type": "video", "duration": "90"},
+                 {"index": 1, "codec_type": "audio", "start_time": "0.125", "duration": "31"}]}
+        self.assertEqual(primary_audio_range(probe), (125, 31125, 1))
+
+    @unittest.skipUnless(os.environ.get("SEARCHI_TEST_FFMPEG"), "Requires verified CPU FFmpeg")
+    def test_actual_video_longer_than_audio(self):
+        ffmpeg = Path(os.environ["SEARCHI_TEST_FFMPEG"])
+        ffprobe = ffmpeg.with_name("ffprobe.exe" if os.name == "nt" else "ffprobe")
+        with tempfile.TemporaryDirectory() as parent:
+            scratch = Path(parent)
+            source = scratch / "longer-video.mkv"
+            run_owned([str(ffmpeg), "-nostdin", "-hide_banner", "-v", "error", "-threads", "1",
+                       "-f", "lavfi", "-i", "color=c=black:s=16x16:r=1:d=33",
+                       "-itsoffset", "0.125", "-f", "lavfi", "-i", "anullsrc=r=48000:cl=mono:d=31", "-c:v", "ffv1",
+                       "-c:a", "pcm_s16le", "-threads", "1", "-output_ts_offset", "5", "-y", str(source)],
+                      scratch, "mux", [(source, 4 * 1024 * 1024)], 30)
+            run_owned([str(ffprobe), "-v", "error", "-show_entries", "stream=index,codec_type,duration,start_time:stream_tags=DURATION:format=duration", "-of", "json", str(source)], scratch, "probe", [], 30)
+            probe = json.loads((scratch / "probe.log").read_text())
+            audio = next(row for row in probe["streams"] if row["codec_type"] == "audio")
+            # Matroska stores duration in tags; test conversion is explicit, not a format fallback.
+            if "duration" not in audio:
+                h, m, sec = audio["tags"]["DURATION"].split(":")
+                audio["duration"] = int(h) * 3600 + int(m) * 60 + float(sec) - float(audio["start_time"])
+            start, end, _ = primary_audio_range(probe)
+            self.assertEqual((start, end), (5125, 36125))
+            self.assertGreater(float(probe["format"]["duration"]), end / 1000)
+            wav = scratch / "tail.wav"
+            run_owned(decode_command(ffmpeg, source, 35125, 1000, wav, audio["index"]), scratch, "tail", [(wav, 970000)], 30)
+            self.assertGreaterEqual(decoded_frames(wav, 1000), 15984)
+
     @unittest.skipUnless(os.environ.get("SEARCHI_TEST_FFMPEG"), "Requires verified CPU FFmpeg")
     def test_actual_ffmpeg_full_window_and_partial_tail_sample_coverage(self):
         ffmpeg = Path(os.environ["SEARCHI_TEST_FFMPEG"])