BP + ‖f‖² penalty: seed=42, lam=0.01, epochs=30
  ep 1: test_acc=0.3621
  ep 5: test_acc=0.4054
  ep 10: test_acc=0.4527
  ep 15: test_acc=0.4872
  ep 20: test_acc=0.5082
  ep 25: test_acc=0.5285
  ep 30: test_acc=0.5303

FINAL test acc: 0.5303
Compare to:
  BP-trainable (3-seed mean):   0.609
  Penalized DFA lam=1e-2:       0.363
  DFA-shallow:                  0.349

Margin vs DFA-shallow baseline: +18.13 pp
  → BP+penalty intermediate; partial capacity loss + residual mode 2
Saved results/bp_with_penalty/bp_pen_lam0.01_s42.json