Agency and Architectural Limits: Why Optimization-Based Systems Cannot Be Norm-Responsive — ThinkLLM