Safety gates
- Offline replay: deterministic tests, golden reports, bounded sensitivity.
- Observer: collect 1 Hz storage and management signals without proposals.
- Shadow: journal bounded proposals and tune alerts without mutation.
- Controlled load: use one non-critical disk with a reviewed, storage-domain-specific static rollback limit.
- Fault injection: stale telemetry, restart, lease conflict, drift, apply/read-back mismatch, and rollback.
- Canary approval: bind exact resource, immutable policy, envelope, expiry, prior effective state, and rollback.
- Review: expand only after SLO and management-plane evidence passes.
Production service installation, configuration changes, or actuation are human checkpoints. GitHub Actions never deploy directly to a PVE host.