v2.1.21

• added npcpy.gen.decision module with DecisionSystem1, DecisionQuestion, DecisionRouter, and BatchDecisions, exported train_decision_model, load_decision_model, load_decision_examples, decision_choice, decision_score, decision_noul, decision_predict, and decision_route from npcpy and npcpy.gen, and added NPCPY_DECISION_MODEL environment variable for a default trained model (4968654)
• added load_decision_examples supporting json, jsonl, and csv training-data formats with automatic answer-type coercion (4968654)
• added tiered routing via DecisionRouter and decision_route with optional confidence-threshold fallback to the lowest tier (4968654)- added examples/example_system1_decisions.py demonstrating system1 decision model training, single/batch inference, module-level calls, and tiered routing (4968654)
• exposed ft.system1 classes and helpers (System1Config, System1Example, System1Predictor, choice, noul, score, predict_system1, train_system1, load_system1) in the server globals dictionary (addc6fe)- added /api/decision/predict, /api/decision/choice, /api/decision/noul, /api/decision/score, and /api/decision/route endpoints to the flask server, falling back to the NPCPY_DECISION_MODEL environment variable (addc6fe)
• added tests/test_gen_decision.py covering DecisionSystem1 training and inference, DecisionRouter, module-level decision helpers, json example loading, environment-based model loading, and batch decision (addc6fe)
• extended the npc compiler's _execute_step runtime environment with additional built-in symbols (8c4f139)