v0.2.0

Fixed

• XGBoost 3.1+ multiclass base_score — XGBoost 3.1+ serializes per-class base scores as a comma-separated bracket string (e.g. '[-3.95E-2,1.97E-1,-1.57E-1]'). Timber previously discarded all but the first value and initialized all class accumulators to 0.0, producing systematic probability errors up to 7.5%. Now the full per-class vector is parsed and applied correctly.

Added

• per_class_base_scores: list[float] field on TreeEnsembleStage IR with full serialisation/deserialisation support
• _parse_base_score_list() helper in xgboost_parser.py for robust base_score parsing across XGBoost versions
• TIMBER_CLASS_BASE_SCORES[] static array emitted in model_data.c for multiclass models
• Comprehensive documentation overhaul — all four docs pages expanded and llms.txt added at repo root
• mkdocs.yml for docs site structure

Changed

• Version bumped from 0.1.0 → 0.2.0
• README completely rewritten: live terminal demo, compiler pipeline diagram, full CLI and API reference tables, runtime comparison feature matrix, roadmap with status indicators

---