fix: improve env robustness, grader fairness, and inference compatibility

feat: implement Phase 1 core engine — models, DAG, agents, rewards, tasks