Changes
- Fix classifier and regressor coordinate-descent initialization and feasible split selection.
- Correct categorical missing-value routing and numerical fallback handling.
- Select outer splits using weighted, regularized best-first growth and enforce strict leaf budgets, including multiway splits.
- Screen bivariate candidates using unregularized impurity gain.
- Warn once per fit when MAE coordinate descent is disabled; opt in with
SGTLEARN_MAE_CD=1. - Re-execute all 10 tutorial notebooks and refresh their rendered outputs.
Compatibility notes
Outer impurity improvement now uses total sample weight and the mean across target outputs. min_impurity_decrease, branching_penalty, and pairwise_penalty subtract constant costs in these units; existing positive penalties may need retuning. Inner CART and TAO retain their separate settings.
Binary wheels and the source distribution are published to PyPI by the tag-triggered CI workflow.