2026.10.63 — 2026-10-28

Fixed

• Keep a cluster's notification stream alive when its credentials cannot be read (d7b6f856)

  A failure to read a cluster's broker credentials from Rahdar's database, for example while a replica hands the cluster over or the database restarts, no longer ends that cluster's broker notification stream for good. The broker client reads the credentials on its own threads and backs off and retries on a broker error, but any other exception ended the thread, so the cluster's live notifications stopped until a restart. The lookup failure is now reported to it as the broker being unavailable.

• Converge a message registration before register and unregister return (41c552da)

  A plugin's PluginMessaging.register now returns with its registration in place on the broker, and unregister with its taps gone, even while another pass over the cluster is running, on this replica or another. Until now such a change was left to the next scheduled pass (ten seconds by default), so a tap installed that late missed the messages sent in between, and a plugin registering several registrations in a row kept most of them waiting.

  An unregister that raced a running pass also made that pass fail on a foreign key ("insert or update on table plugin_message_registration_node") and give up, so every other registration on the cluster waited for the next one. The change and its pass now run together under the cluster's lock; a call waits up to one reconcile interval for a pass already running, then stores its change anyway and logs a warning (ADR-0213).