Text this: Semantic Reward Collapse and the Preservation of Epistemic Integrity in Adaptive AI Systems