Wire up validate_mental_model_refresh hook in the HTTP routes for both create and refresh mental model endpoints, allowing extensions to reject operations (e.g. insufficient credits) before queuing async LLM work.