Startup adopts the refs it finds so a daemon that lost its state file does not rerun the whole record. That told it a ref was handled even when the run handling it had been killed, leaving the check record on its first write and reading `running` for good, with nothing left to finish it. The runner's own check records outlive its state file, so startup reads them. A ref whose latest check is still `running` stays unadopted and runs again on the next event for that repository. Observed on the deployment: replacing the host at 06:12 killed publish-wizard 33 seconds into its run, and the check sat `running` until it was asked for by hand. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>