fix(runtime): stop retrying dead directory members

Merged
RE
3mo ago
+40-6

Directory recovery could repeatedly send RecoverRegisteredActivations requests to a silo which the messaging layer already knew was dead. The recovery loop consulted a lagging directory membership view and retried failures indefinitely at a fixed 100 ms interval, producing sustained message floods.

Use the latest cluster membership snapshot before recovery and snapshot-transfer calls so unavailable targets are abandoned immediately. Transient failures retain retries, but now use randomized exponential backoff capped at five seconds. Regular directory calls also refresh and reroute immediately when their current owner is known unavailable.

Microsoft Reviewers: Open in CodeFlow