2026-05-06

Promote LatestModelCallback to a full training-state checkpoint (latest_model.pt now stores model + optimizer + scheduler + RNG + early-stopping state + best_value + config_hash)
Add Trainer.train(start_epoch=...) so a run can pick up partway through the epoch loop
Wire --resume flag into cli.py train and main.py; run() gains a resume parameter that loads latest_model.pt per seed and rehydrates optimizer/scheduler/RNG with config-hash verification
CheckpointManager no longer writes a redundant latest.pt; latest_model.pt is the single resume entry point. best.pt and periodic checkpoint_epoch_*.pt remain opt-in via checkpoint_config.enabled

2026-04-08

Add dual logging: logging: wandb (default) or logging: tui for agent-friendly terminal output
Add CSVLoggingCallback (always active): writes metrics.csv every epoch with dynamic column expansion
Add TUILoggingCallback: structured per-epoch terminal output replacing W&B
Add LatestModelCallback (always active): saves latest_model.pt every epoch
Add provenance tracking: env_snapshot.yaml + run_metadata.yaml per run
Add Rust TUI monitor (tools/monitor/): real-time loss curve visualization from metrics.csv
Add doctor CLI command: system environment health check
Add monitor CLI command: launch TUI monitor with auto-detection
Improve loss prediction: shifted exponential decay L(t) = a·exp(-b·t) + c
Complete migration skill documentation (M1–M6) with detailed code-level migration steps
Add migration enforcement: pre-push hook blocks push when source files change without migration doc updates

Add PFL (Predicted Final Loss) Pruner and its documentation
- Add PFL Pruner to optimize hyperparameter search
- Update default optimization template with PFL Pruner configuration
- Add Pruner documentation in README.md under Appendix section

Support TPE.GridSampler
- Add OptimizeConfig.grid_search_space
- Modify OptimizeConfig._create_sampler
- Caution: For GridSampler, all variables in search_space should be categorical