4.2.0

• Enhancement
  • Support numpy>=1.26,<3 (adds numpy 2.x support).
  • Bump scikit-learn range to >=1.5,<1.8 (numpy 2.x support starts in 1.5; this
    raises the minimum supported version, which is a consumer-visible change).
  • Bump shap range to >=0.43,<0.49.
  • Bump lightgbm range to >=4,<5.
• Bug Fix
  • xgb_classification_learner / lgbm_classification_learner: handle SHAP >=0.45
    returning a single 3D ndarray for multiclass TreeExplainer (instead of the legacy
    list of per-class 2D arrays).
  • lgbm_classification_learner: handle SHAP >=0.45 returning a single 2D ndarray
    with a scalar expected_value for binary TreeExplainer (instead of the legacy
    [neg_class, pos_class] pair).
  • lgbm_classification_learner: pass feature_name and categorical_feature to
    lightgbm.Dataset instead of lightgbm.train (required by lightgbm >=4). Note:
    a user-supplied feature_name is now honored at the Dataset level — previously the
    Dataset always used string-cast features regardless of the feature_name argument.
  • cumulative_effect_curve (and the gain / relative-gain / effect_curves functions
    that build on it): use a stable sort when ordering rows by prediction score. Previous
    default (kind="quicksort") produced non-deterministic tie-breaking across numpy
    versions, which changed the rows included in each cumulative prefix when prediction
    had ties.