Similar Items: Semantic Reward Collapse and the Preservation of Epistemic Integrity in Adaptive AI Systems