Retrain classifier on per-node recordings (target >80% all classes)

[taylorjdawson]

After re-recording with per-node server, retrain using 62-dim features. Evaluate accuracy improvement vs old 15-dim. Add room-aware classes if room fingerprinting is ready.