Audit and harden Claude Code, MCP, and AI-agent workflows: permissions, approval boundaries, failure recovery, verification, tests, and operator handoff.