Same Simulation, Different Evidence: Where Retrieved Tool Output Costs LLM Agents Accuracy