After a replicated `deploy_component{restart:true}`, the component's **MQTT topic resource is registered per-worker

From dispatch finding (task `qa-wave-2026072914`):

- F-255 — After a replicated `deploy_component{restart:true}`, the component's **MQTT topic resource is registered per-worker and diverges**: on the publisher's node `[http/1]` initialized the table at `.400` but still had no MQTT resource at `.541`, rejecting 7 consecutive QoS1 publishes over 5 s (`DurableSubscriptionsSession.ts:417`), while `[http/2]` on the same node initialized it 0.65 s later. The client sees only reason `0x80` "Unspecified error", then an **unannounced socket close whose in-flight publish never settles** — the exact mechanism that hung two prior waves. A readiness probe passed on both nodes at +84 ms because it landed on a worker that had the resource, so connection-to-worker pinning makes any health check blind to this. MQTT-surface twin of harper#1951. Not established: whether the un-registered worker ever self-heals (window only ~10 s) — measure that first, since "eventually consistent" and "broken until restart" are different bugs.