Wrong Prediction, Right Answer: Recovering Evidence from Collapsed LLM Sequence Scores