Diagnosing Non-Markovian Observations in Reinforcement Learning via Prediction-Based Violation Scoring