Keep a credible way back.

Recovery begins with ownership and durable evidence.

Distinguish failure phases

Before cutover, a failed candidate leaves the established owner and route intact. After handoff, rollback must respect current fenced ownership and completed effects. Never restart an old owner blindly while the candidate can still execute.

Recover durable acceptance

Queues and schedules dispatch from bounded memory. Recovery journals preserve accepted work and critical transitions across restarts. Durable recovery does not make an uncertain external effect safe to resend.

Verify restore, not just backup

Backups must preserve the database, required encrypted credential bindings and compatible release metadata. Test restores in isolated resources. A healthy API does not prove workspace custody has been unlocked, and a single physical server cannot tolerate whole-host loss without interruption.