Phase 7: Manual GPU verification against MIT CSAIL video

[joeljose]

Context

Parent: #10 | PRD: #2 | Design: docs/design/visualmic-hardening.md
Blocked by: #4, #5

Goal

Verify the GPU pipeline produces recognizable audio from a real MIT CSAIL test video after all code changes. CPU verification is impractical (704×704, 22K+ frames would OOM or take too long).

Acceptance Criteria

• Download MIT CSAIL test video (e.g., Chips1-2200Hz-Mary_Had-input.avi from http://data.csail.mit.edu/vidmag/VisualMic/Results/)
• Run GPU pipeline: python visualmic.py -i Chips1-2200Hz-Mary_Had-input.avi --gpu --fps 2200 -o test_output.wav
• Listen to output audio — verify "Mary Had A Little Lamb" is recognizable
• Test with --nlevels, --biort, --qshift flags to verify they work with real data
• Test with --roi to verify ROI cropping works on real data
• Verify pre-flight VRAM estimation prints before processing

Files likely involved

• None (verification only)

Testing approach

• Manual listening test on output WAV
• Compare quality with previous output (if available in output/ directory)

[joeljose]

GPU verification complete. Tested Chips2-2200Hz-Mary_MIDI-input.avi (704x400, 38083 frames, 2200 fps) with all flag combinations: default, --nlevels 2, --biort near_sym_a --qshift qshift_a, --roi 100,50,400,300. All outputs valid (2200 Hz, 17.3s, full int16 range). VRAM estimation prints correctly. Completed in PR #11.