A shutdown releases every shared subscription it owns

A number of services share one background subscription among every caller that asks for the same thing — the compile pipeline's activity record, a content collection's initial load, a repository's canonical name, the user directory index, the in-flight activity and routing counters, and the node cache's synced queries. Sharing is right: the work runs once and every reader sees the same result. What was missing was an owner. The mechanism that opened the shared subscription kept its handle to itself, so when the service shut down nothing could close it, and when the subscription was still queued for a background thread at that moment, it started after the shutdown and tried to build itself against services that no longer existed. On the build servers that surfaced as a burst of "already disposed" errors after a test had reported a clean teardown.

Every such subscription now belongs to the service that opened it and is released as part of that service's own shutdown: one that is running is unsubscribed, one that is still queued is cancelled before it starts, and one that has already finished its work no longer counts as open. A caller who asks after the shutdown is answered immediately with an error naming the service, rather than being handed a stale result and then left waiting for updates that will never come. A guard in the test suite keeps every future shared subscription on the same rule.