The Telephone Game: Evaluating Semantic Drift in Unified Models