fix(eval): add phase-aware hard and cascade action selection

fix(grpo): align hard and cascade training with final eval

Use constrained action scoring for model evaluation

Improve GRPO rewards and add model evaluation