Safety Alignment Inverts the Tension–Response Curve in LLM Agents for Civil Violence Simulation