v2.2.0 WoE Stability: Class Count Constraint Enforcement

[2.2.0] - 2026-02-18

Added

• min_negatives parameter in BinningConstraints for controlling minimum negative samples (y=0) per bin, ensuring stable WoE calculations
• positives and negatives properties on Block class for easier access to class counts in binary classification
• Feasibility warnings during resolve() when min_positives or min_negatives constraints cannot be mathematically satisfied with the given min_bins
• _enforce_min_class_counts() function for unified enforcement of both min_positives and min_negatives constraints in a single pass
• New tests for class count constraint enforcement and feasibility warnings

Fixed

• min_positives constraint is now actively enforced - Previously it was only a soft penalty (1.4x score multiplier) in the merge scorer; now bins violating min_positives are actively merged until the constraint is satisfied or min_bins floor is reached
• _validate_merge_result() now properly checks and warns about min_positives and min_negatives violations when constraints cannot be fully satisfied

Changed

• merge_adjacent() now includes Phase 3 for class count enforcement after min_samples enforcement
• _validate_merge_result() now accepts is_binary_y parameter to enable class count validation
• Block.as_dict() now includes positives and negatives keys in the exported dictionary
• MergeScorer._apply_penalties() now applies bonuses for merging bins with insufficient negatives (matching existing positives behavior)
• BinningConstraints.__repr__() now includes min_negatives in the string representation
• BinningConstraints.copy() now properly copies min_negatives parameter