Implement self-play training data generation

[kirillbobyrev]

Summary

Third in the sequence (after #325, #326). Turns the empty datagen stub into a working self-play data generator built on the unified Game — the data source a future policy/value network will train on.

Changes

• Self-play driven by MCTS. Each game is played by the search; at every move the search's visit distribution over the legal moves is recorded as the policy target, and the game's eventual outcome (from the side-to-move's perspective) is the value target. Opening plies are sampled proportionally to the visit counts for variety; later moves are greedy.
• Parallel + reproducible. Games are independent and generated in parallel with rayon; each game is seeded from seed ^ index, so a run is deterministic regardless of scheduling. An optional Syzygy tablebase adjudicates endgames.
• Output. Samples are written as <fen> | <value> | <uci>:<visits> .... The datagen binary exposes --games / --nodes / --exploration-plies / --max-plies / --seed / --tablebase / --output via clap.
• Search plumbing. Exposed the root visit distribution from the search as mcts::policy — the policy target — reusing the existing SearchResult.

Testing

• 73 lib tests (3 new datagen tests: well-formed output, reproducibility, and value-target orientation — a mate-in-one self-play game yields +1 for the winning side); full suite (137 total) green.
• Manual: cargo run --bin datagen -- --games 4 --nodes 80 --output … produces 240 well-formed samples.
• cargo clippy --all-targets --all-features zero warnings; cargo +nightly fmt --check clean.

Next in the sequence

UCI-completeness pass, then the transmute safety cleanup. (The network itself — consuming this data — remains the larger future effort.)

🤖 Generated with Claude Code

https://claude.ai/code/session_01GVrKvkMe4xDmkRAksacBzk

---

Generated by Claude Code