Add RoSA sweep pod setup script

Add UCI engine generation, Parquet data pipeline, and Lc0/Stockfish containers

Add Lichess dataset extraction pipeline and update pod.sh

Add RoSA adapter with gradient-informed sparse masks (#3)

Merge branch 'review-fixes'

Fix terminal position diagnostics: report pad_prob for checkmate/stalemate, not legal_rate

Add sweep migration script (rsync between pods via local staging)

Fix sweep: don't pass --epochs to train.py, handle all-pruned results

Fix: create SQLite parent directory before Optuna study creation

Architecture sweep: GPU affinity, arch search space, train.py overrides

Fix outcome array dtype in ceiling computation, np.asarray in script

Fix bugs, performance issues, and doc errors from code review

Add Monte Carlo accuracy ceiling computation to engine

Add theoretical accuracy ceiling computation (E[1/N_legal] and outcome-conditioned)

Add Optuna hyperparameter sweep infrastructure

Consolidate all training logging through MetricsLogger

Log patience counter, best val loss/step in val records

Per-model early stopping: freeze converged variants individually

Monitor script: show step time, games/sec, ETA from synced metrics

Revert batch size default to 256

Fix monitor script to show per-variant metrics from pod

Fix SSH: generate host keys, use 'ip' field from runpodctl

Push metrics to HF at eval intervals, add dashboard HF sync

Remove hardcoded IP from monitor script, resolve SSH via runpodctl

Add early stopping patience to multi-model training

Remove .item() CUDA sync from hot path, batch size 512, run slugs

Add post-training evals, /dev/shm checkpoints, async HF push, and _orig_mod fix

Safetensors migration, checkpoint integrity, and multi-model training. (#1)

Add attributions and inline links for open source projects and academic publications

Add git version baking in Dockerfile, progress monitoring, and probe eval script

Add Docker BYOC deployment, discard-ply-limit ablation, dashboard, and deploy tooling

Move PAWN to open source repo.