Add mode2_v1 schema — 73 pre-sliced clean segments from 27 source episodes (104.7 min, bad intervals excluded by construction). New ReactSegmentDataset + demo. No bad_frames.json filtering needed at the dataloader; data is constructively clean. mode1_v1 stays for backward compatibility. See segments.json + new README section.
