fusion

Rust engine for TETR.IO replay coaching: move generation, search, eval, S2 attack, native and WASM builds. Python training tools live under training/. Movegen uses cobra-movegen's smear-board reachability approach.

Setup

Rust via rustup, Python via uv.

CLOUD_EXEC_SKIP=1 sh scripts/test.sh
CLOUD_EXEC_SKIP=1 cargo clippy -- -D warnings
cd training && uv sync && uv run pytest tests

Copy .env.example to .env only for replay collection, Modal training, or the label generator; tests and clippy run without it.

Large artifacts (replay corpora, training bins, label sidecars, ONNX models) are not tracked; keep a model's .metadata.json next to it. See training/TRAINING.md for the training pipeline.

The test command verifies and caches checksum-pinned opener inputs and generates the shared recommendation vectors before running the library suite. It needs Python 3.11+, curl and either sha256sum or shasum on first use; intact cached opener inputs work offline. The inputs stay ignored under fixtures/openers/ and tests/fixtures/recommend-v1.json, with no sibling application checkout or storage credentials required. Run sh scripts/test.sh --test recommend_contract to include the native recommendation boundary tests. Before running Cargo test commands directly, run sh scripts/hydrate-opener-tests.sh, create tests/fixtures/, and run python3 scripts/generate-recommend-vectors.py tests/fixtures/recommend-v1.json. Generated reports stay under ignored evidence-out/. The ignored historical census comparison requires its separately retained baseline report; normal compiler tests do not read that report.

The opener runtime accepts catalog bytes from its caller. Dataset downloads are test setup, not engine runtime behavior.