feat(phase3b): d64.pt floor commitment — trained distinguisher + eval + regression PASS

Trained at depth 56 (Gohr pattern for 64-round attack peeling 8 rounds):
  - Wall clock:          62 min on RTX 5090
  - Config:              depth=5, width=512, 10M samples, 20 epochs, lr=2e-3
  - Δ:                   0x00000002 (best from Δ search, commit 7e1525f)
  - Final train loss:    0.506
  - Final val accuracy:  0.756

Evaluation on 1M held-out samples (seed=4242, disjoint from training):
  - Accuracy:            0.752
  - ROC-AUC:             0.828
  - TPR @ FPR=1%:        0.207
  - Confusion (TN/FP/FN/TP): 400997 / 99003 / 149119 / 350881

Regression test tests/test_neural_e2e_64r.py PASSED in 1.6 s:
  - 64-round hybrid attack, 8 neural target bits
  - Full 64-bit key recovered + cipher-verified

Phase 3b spec §10 success criterion 2 (floor commitment): met.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: cahlen

checkpoints/d64.pt

@@ -0,0 +1 @@
+{"rounds": 56, "delta": 2, "n_samples": 1000000, "accuracy": 0.751878, "roc_auc": 0.828261923916, "tpr_at_fpr_01": 0.20681799948215485, "confusion": [400997, 99003, 149119, 350881]}

docs/phase3b-results/eval_d64.json

@@ -0,0 +1 @@
+{"final_loss": 0.5058022391029697, "final_val_accuracy": 0.75595, "wall_time_s": 3732.0717719400127, "checkpoint": "checkpoints/d64.pt"}