Documents how the per-frame face/lip tracks are produced:
- inputs and the two frame-rate conventions (06b native fps vs TalkNet 25 fps)
- per-frame active-speaker choice with SWITCH_MARGIN=0.30 hysteresis, IOU_MIN=0.05,
and the deliberate absence of a top-score fallback (no match_seq==2)
- field reference incl. track_seq / n_visible_tracks / speaker_switches
- alignment with camera_extrinsics_kf_inds, and consumption guidance
(gate on match_seq==1; filter speaker_switches==0 for single-identity clips)
- what the old clip-level best_track rule got wrong and who should re-download
Also corrects pipeline-summary step 9, which still described the retired
'TalkNet best-speaker track' behaviour.
