PASS: dynamics, episode guards, train-only normalization, final-label invariance, stopping rules, metrics, and artifact integrity.
80 episodes; 2400 actions; split: 48 train / 16 validation / 16 final-test.
Validation MSE affine: 0.00044645 (m/s^2)^2
Validation MSE knn-1: 0.00094054 (m/s^2)^2
Validation MSE knn-7: 0.00087992 (m/s^2)^2
Selected and frozen: affine
Final nominal action MSE: 0.00044231; first five actions: 0.00184831
nominal learned: success=16/16, boundary=0, timeout=0; mean successful time=2.225000 s
nominal expert: success=16/16, boundary=0, timeout=0; mean successful time=2.125000 s
  paired: both=16, learned-only=0, expert-only=0, neither=0
shifted learned: success=14/16, boundary=0, timeout=2; mean successful time=2.357143 s
shifted expert: success=15/16, boundary=0, timeout=1; mean successful time=2.440000 s
  paired: both=14, learned-only=0, expert-only=1, neither=1
Synthetic diagnostic batch; final labels never select the model. No hardware or untouched-benchmark claim.