Fix .gitignore to allow .npy in gradient_maps/ and upload all gradient numpy arrays

Upload v3 gradient .npy files (fix: re-upload missing numpy arrays)

Upload v3 gradient saliency plots (corrected desync + pre-sigmoid logit gradients)

Upload v3 gradient maps (corrected desync + pre-sigmoid logit gradients for LMIC-TSBN)

Add raw gradient saliency maps v3 (pre-softmax logit gradients, 1000 traces, corrected desync)

Add gradient saliency line/area plots (50 files): per-model overlays, arch comparison, desync progression

Add all gradient saliency heatmaps (65 files): arch comparison, per-model, desync progression, thesis summary

Add gradient visualization: fig9_quantitative_summary.png

Add gradient visualization: fig8_lmic_desync_robustness.png

Add gradient visualization: fig7_mlp_vs_cnn_attention.png

Add gradient visualization: fig6_tsbn_effect.png

Add gradient visualization: fig5_binary_encoding_effect.png

Add gradient visualization: fig4_hps_representation_competition.png

Add gradient visualization: fig3_lmic_tsbn_16byte_grid.png

Add gradient visualization: fig2_cnn_desync_degradation.png

Add gradient visualization: fig1_architecture_comparison_byte2.png

Add gradient visualization: fig10_seed_sensitivity.png

Add gradient saliency maps for all 139 completed runs

Add gradient saliency map generation and visualization tools

fix(snr): use masked intermediates (snr2, snr3) instead of unmasked snr1

fix: promote retry0.post_train_final_rank to top-level for dashboard display

fix: LMIC base class multi_bit support in build() and compile()

fix(agent): report full results on training completion without rank 0

Section 45.8: Single-run enforcement fix (no retries), trainer.py max(1,max_retries) patch

Section 43-45: dataset.py per-trace shift fix, orchestrator DELETE FK fix, queue rebuild, clean MLP/CNN resubmission

Update history ledger: config bug, upload path fix, V7b restoration, batch9

Update HISTORY_LEDGER: Section 43 - Cross-Desync Ablation Recovery & Worker Fixes (2026-05-02)

fix: handle return_code=None by waiting for process exit and trusting results.json

docs: Sections 39-42 — full rerun campaign, normalization bug fix, batch re-evaluation, decision log

feat: unified V1 model repo (ascad-v1-models)

Add per-byte accuracy/loss metrics to worker progress tracking and dashboard log viewer

Update HISTORY_LEDGER: Section 38 - Dashboard & Auto-Recovery Hardening

Fix dashboard: viewJobLogs uses correct element IDs, pause auto-refresh when modal open, fix newline escape in progress history

Fix dashboard: handle retry0/retry1 result structure, getBestAttempt helper

Harden orchestration: auto-recovery, job detail popover, completed job display

docs: Section 37 - infrastructure hardening + V7b attempt 1 results

feat: comprehensive W&B logging + retry=0 hardening

V7b SUCCESS: 16/16 rank 0 - multi-bit + DTP breakthrough

docs: update ledger with V7 post-mortem and V7b fix (Section 35)

fix(dtp): support multi_bit mode - look for binary_accuracy keys

ledger: Section 34 - V6 results, variant bug, V7 multi-bit

fix: agent.py variant mapping uses model_type as fallback

fix: indentation error in train_mtl.py --sigmoid-gates arg

feat: multi-bit labeling (Wu et al. TCHES 2024)

docs: Section 33 — V6 sigmoid gates experiment

feat: wire sigmoid gates through training pipeline (trainer, CLI, agent)

HISTORY_LEDGER: Section 32 - v5g retry 0 results (12/16 rank 0)

docs: Section 31 — DTP weight oscillation finding and EMA fix

feat: add DTP EMA smoothing to prevent weight oscillations from cosine annealing warm restarts

docs: Section 30 — v5e results (11/16), v5f submission (stronger DTP)