Run a metric-gated refactoring loop that curbs fake architecture debt by requiring a measured signal (radon cyclomatic complexity + grimp/networkx Louvain community detection) AND a named semantic problem before any change — never the model's taste. Use when refactoring autonomously, cutting cyclomatic complexity, untangling module boundaries, or commissioning a /goal-style overnight refactor campaign. Front-loads all decisions in one grill, then GENERATES a ready-to-paste /goal prompt that runs the loop until the metrics are clean. Python projects only (radon/grimp are Python tools).
