fix: normalize video audio for diarization

[FernandoAbishai]

Context

Phase 5B.6 real-media regression exposed that pyannote.audio 4.0.7 could not read a finite header duration from the frozen WebM container through torchcodec. Existing transcription already normalizes these video containers to temporary WAV audio.

Changes

• reuse extract_audio and cleanup_temp_audio for .mp4, .avi, .mov, .mkv, and .webm before diarization;
• preserve direct audio-path behavior;
• clean temporary audio on success, pipeline failure, normalization failure, and speaker-map failure;
• preserve overlap-based speaker mapping and token redaction;
• add subprocess-isolated CI-safe normalization and cleanup smokes.

Validation

• focused diarization smoke: 12/12 PASS
• backend smoke suite: 105/105 PASS
• optional-capability smoke: 8/8 PASS
• qualification harness: PASS
• python compileall: PASS
• git diff --check: PASS
• real Fixture B regression: PASS; 7 distinct speakers, 228 UNKNOWN words, temporary WAV existed during inference and was removed afterward

The real-media run is regression verification only and does not mark Phase 5B.6 qualification PASS.

Scope

No changes to requirements, pyannote model selection, transcription behavior, frontend/UI, project schema, qualification matrix, or release workflows.

Do not merge.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.