A cold read no longer says a node is missing for the rest of the day

A portal keeps one live query per question it has been asked, and replays that query's first answer to everyone who asks again — which is what makes reading the same thing twice cheap. The problem was what counted as an answer. While a portal is still warming up, a storage backend can finish a query without saying anything at all; the platform treats that as "nothing matched" so the page renders instead of hanging, and that fabricated empty then became the cached answer. Nothing ever corrected it, because the event that refreshes a query is a change to a matching node, and re-writing a node that is already correct changes nothing.

The visible effect was a restarted portal insisting, for as long as it ran, that records it had itself stored were missing: access grants read as absent, settings read as unset, a reconcile reporting that its own writes had not been saved.

A query result now names the backends that never answered, the platform keeps only answers somebody actually gave, and the next read asks again. An answer everybody gave is cached exactly as before.