The signal is the ceiling: Measurement limits of LLM-predicted experience ratings from open-ended survey text