Recover truncated video decodes via an ffmpeg fallback

[JoshW-dev]

What

extract_visual lost the tail of any clip that OpenCV stopped decoding partway through. Some variable-rate AVIs make cv2.VideoCapture give up early and report no error, so features covered only the front of the clip and every later time bin came out NaN, while the audio features (decoded separately by ffmpeg) looked fine.

Found while extracting Dr. Lin's shared movie clips: 10a4.avi declares 614 frames, OpenCV decoded 330, and the back half of the clip was blank.

Fix

• Compare OpenCV's decoded frame count against the container's declared count.
• On a short read (< 98%), retry with ffmpeg, which reads these files whole.
• Keep whichever backend got further; raise a RuntimeWarning instead of silently emitting NaN.
• Decode loop split into _iter_frames_cv2 / _iter_frames_ffmpeg / _visual_rows so both backends share one feature path.

Verification

• 10a4.avi now recovers all 614 frames straight from the raw AVI, zero NaNs. Cross-checked against an independent clean transcode: luminance r=0.99, faces r=0.99, audio r=1.00.
• Full 165-clip corpus re-extracted from the raw AVIs: 0 NaN cells, uniform schema, fallback fired on exactly the one broken clip (no false positives).
• New regression test truncates the OpenCV path and asserts full recovery + the warning; verified it fails without the fix. Green under both pytest and the bare python tests/test_pipeline.py runner (16/16).