Live evidence showed a healthy candidate connector and a restarting router:
legacy router environment omitted HN_CONNECTOR_PUBLIC_KEY, so the new routing
verifier tried runtime/pki/connector.pub instead of the existing /trust mount.
Resolve the public key from the existing read-only trust bind, set its path
in new overlays, and actually load it during isolated runtime checks. Add
repair-router for a known healthy candidate connector: preserve credential
keys and live configuration, back up overlay/plan, add only the router path,
update integrity metadata, probe, and recreate router alone with --no-deps.
Retries do not regenerate keys or overwrite repair backups.
Add read-only state including stopped/restarting containers and fixed startup
error classes, including stderr, without printing raw logs or credentials.
Identify command stage/exit code and activation progress. Permit rollback
of known stopped candidate targets and old overlays without requiring the
new router key; unknown images still fail closed.
Validation: full suite 262 passed, 3 skipped (real PostgreSQL unavailable),
30 existing warnings; final tooling suite 50 passed after the added CLI
rollback regression. Python 3.7 syntax and staged diff checks pass.
No live repair was executed by this agent; operator must run repair-router
and verify health and the real client handshake on ECS.