A page no longer goes dark on the first view after a restart

Each portal replica announces which addresses it serves, so its peers can reach it directly. That announcement is republished whenever the cluster's shape moves — a replica starting, stopping or being rescheduled — because moving the cluster is exactly what can lose it.

Two narrow windows let one of those republish signals go missing. Both fell inside the moments right after a replica starts, which is when the cluster's shape is most likely to move, and in both the signal was dropped rather than delayed: nothing retried it and nothing recorded it. The replica then kept an announcement that no longer matched where its peers thought it lived, and nothing would repair that until the cluster happened to move again.

What you could see when it happened: a read that should be instant instead ran to its time limit and came back as an error, so an image or a document was blank on first view and fine on a refresh — because the refresh took a different route. It looked like a slow network and was not.

Both windows are closed, and the announcement is now made on every cluster change including the ones that arrive while an announcement is already in progress — still exactly once per change, so nothing announces more often than before.