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

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

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

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

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

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

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

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

feat(db_mtl): add log_loss flag, default False for homogeneous tasks

fix(db_mtl): remove erroneous 1/T averaging — match paper Eq. 2

fix: use Keras 3 API for metrics tracking in DB-MTL train_step

feat: implement DB-MTL gradient balancing callback (Lin et al., Neural Networks 2023)

feat: add XLA jit_compile support for LMIC/LMIC-TSBN

feat: disable sigmoid gates by default in LMIC-TSBN (Section 22)

fix(gradient_logger): use layer.name→var mapping for Keras 3 compatibility

fix(gradient_logger): support LMIC-TSBN layer naming in _classify_layer

feat: add gradient clipping (clipnorm) to LMIC/LMIC-TSBN — Phase 1 stability fix

feat: add per-layer gradient norm logging to W&B — logs shared conv, per-task BN, head, gate norms every epoch

fix(cosine): use optimizer.learning_rate.assign() for Keras 3 compat

feat: add Focal Loss + Cosine Annealing for LMIC v6

feat: LMIC-TSBN model + persistence fixes across restarts

feat: LMIC architecture - Localized Multi-Input CNN for 16-byte key recovery

feat: training speed optimizations — mixed precision, vectorized augmentation, cached eval predictions

v3: Add DTP + Spectral Decoupling, fix GradNorm OOM, fix _fail_job cancel