Policy-Invariant Reward Shaping from LLM Feedback: A Framework for Hybrid RL Agents