Tests

Two independent test paths.

Python suite — test_rf_macros.py

Checks every fit, predict, evaluate, out-of-bag and tuning macro against an equivalent scikit-learn / numpy reference on the same fixed-seed data, so a failure means the macros disagree with a trusted implementation (not merely that a recorded number drifted). The strongest checks are the single-tree CART equivalences: with n_trees:=1, sample_frac:=1.0, replace_sample:=false, mtry:=d duckRF grows one deterministic CART that matches sklearn's DecisionTree{Regressor,Classifier} (max_features=None) to ~1e-9 on tie-free data — predictions, predict_proba, node impurities (gini / entropy-in-bits / variance) and MDI rf_importance all agree exactly. Where equally-good tied splits exist, sklearn and duckRF may break the tie differently (both valid), so those configurations assert the weaker adversarial property that duckRF's realized objective is no worse than sklearn's.

Also covered: soft-voting forest accuracy/R² within tolerance of sklearn's RandomForest on a holdout; every *_evaluate metric vs the sklearn metric function (accuracy / log-loss / brier / AUC / RMSE / MAE / R²); categorical subset splits vs a brute-force optimum (exact for regression and binary classification, "at least as good as the best singleton" for the K>2 heuristic), including numeric-looking levels that stay categorical, ENUM columns, and the headline categorical-outcome / categorical-features use case; out-of-bag membership re-derived in numpy and OOB score near a true holdout; determinism under PRAGMA threads=1 (bit-identical model, different seed differs) and the md5 bootstrap-draw replay; weights_col and class_weight:='balanced'; rf_cv / rf_cv_depth; and the full NULL / empty / degenerate / reserved-name / type / guard error contract.

python -m venv .venv
.venv/bin/python -m pip install -r tests/requirements.txt   # macOS/Linux
.venv/bin/python -m pytest tests/ -q

SQL smoke test — smoke.sql

No Python required — just the DuckDB CLI. Fits on deterministic inline data and aborts (non-zero exit) on the first failed check. Runs under PRAGMA threads=1 so it is bit-for-bit reproducible. Run from the repo root:

duckdb < tests/smoke.sql