Test Plan

Tests are written before (or together with) the code they cover. Every optimized path must be proven equal to a naive reference.

API tests (tests/test_api.py)

• model can be initialized; __init__ only stores parameters (no validation)
• get_params returns constructor parameters; set_params round-trips
• (Phase 3) fit returns self; predict shapes; predict_proba rows sum to 1; NotFittedError before fit

Correctness tests

• FM naive O(n^2 k) prediction equals optimized O(nk) prediction (tests/test_fm_correctness.py)
• FFM prediction matches a hand-computed tiny example and the naive loop (tests/test_ffm_correctness.py)
• Dense input and CSR input produce identical predictions for FM and FFM (tests/test_sparse_dense_equivalence.py)
• Losses match hand-computed / scipy values; label smoothing formula; numerical stability at extreme logits (tests/test_losses.py)
• (Phase 2) logistic/softmax loss decreases during training on synthetic data
• (Phase 2) Rust predictions equal Python reference predictions

Reproducibility tests (Phase 2+)

• same random_state gives same predictions
• different random_state can produce different initialization

Serialization tests (Phase 4)

• save_model/load_model preserves predictions
• pickle roundtrip preserves predictions

Edge cases

• all-zero rows (prediction = w0 + 0)
• single nonzero feature (pairwise term = 0 for FM and FFM)
• empty interaction sets, k=1, single field
• (Phase 2+) unseen categories in encoder, class imbalance, sample_weight