STOCKTAKE: Measuring the Gap Between Perception and Action in LLM Agents with a Fair Oracle