Fix evaluation artifacts and align notebook outputs

Fix evaluation to compare heuristic vs local trained model

Add training-first RL architecture with tracking