Evidence-first, read-only audits that always use one gpt-5.6-sol xhigh analyst and return a self-contained fix plan suitable for a lower-capability execution model. Use when the user asks to audit code, a diff, a branch, an artifact, or a runtime flow.
