Durable streams are mesh nodes

Cross-silo delivery on the platform still had one in-memory Orleans stream underneath it, and every so often a request would wait out a full minute because that stream's queue or registration had stalled inside the cluster. The pending question was whether to buy a durable stream provider to harden it.

The answer, now recorded as an architecture page, is that no provider is needed: each kind of traffic the stream carries already has a durable home in the mesh itself. Synchronized data recovers from the node's version chain; a request to a hub that is not live should be told so in milliseconds rather than queued; cross-silo cache invalidation belongs on the database's own change feed, which every pod already listens to; and work that must survive a restart already uses the _Inbox node pattern. The page names the order in which the remaining slices land and the log line that decides when the stream can go.

Alongside it, four code comments that still claimed the database change listener "is never started" have been corrected — it has been running on every pod since the fix for the August outage, and a test guards it.