Multi-person tracking

MultiPersonTracker (core/tracker/multi_person_tracker.py) extends the same stage machinery to several people: the root stage's object detector runs every frame to propose candidates, candidates are matched to existing PersonTrackStates by Hungarian assignment on a cost blending bbox IoU and keypoint displacement (MultiPersonTrackingConfig: weights 0.5/0.5, max_age 10, min_hits 3), and matched detections are finalized with each track's own accumulated temporal state. Unmatched detections spawn new tracks; unmatched tracks age out.

FreeMoCap does not use this today, its configs cap detections to one crop per camera, with an in-source comment explaining that a fixed batch size prevents intermittent GPU OOMs from spurious detections.