A year of eight-hour certificates rather than an authorized_keys file, and the operational reality of it.
what worked:
onboarding a third person: one signing command
offboarding: nothing to do. the certificate expired.
the audit trail: every signature logged with a
principal and a validity window
what did not:
a certificate expiring mid-incident, twice. the
renewal is a command and nobody wants to run one
at 03:00.
→ a 24-hour validity for the on-call principal,
which is a deliberate weakening and is written down.
and the thing nobody thought about: the CA key's own
backup, which lived in the same place as everything else.
The mid-incident expiry is the failure mode that theory does not predict, and the answer is a longer window for a specific principal rather than a longer window for everybody. The CA key being backed up alongside everything it protects is the more serious finding and was fixed by the offsite work in December.