Multi-agent dry run monitoring, governance quality review, and failure analysis. Current mode: dry_run by default — no execution occurs until switched to execute_approved.