Forces heavy internal computation (thinking tokens) before each response. Used for stress testing / A/B testing LLM backend KPIs.