Starting a data source no longer leaves a spare hub behind

When a hub with data started, each of its data sources opened its primary EntityStore stream — the one every read and write of that source flows through — and then, in the same breath, reduced that stream to its own full reference and threw the result away. The reduce was not free: it built a second SynchronizationStream with its own hosted sync/ hub, registered on the primary stream for disposal, that no caller could ever reach and that lived for as long as the node hub did. One spare hub per data source per partition, on every hub that carries data.

The spare hub is no longer created. Starting a data source opens its primary stream and nothing else. Reads, writes and the initialization gate are unchanged — they never used the discarded stream.

It was found through a test, not a heap: ReadPathStreamMintingTest counts the sync/ hubs five uncached reads leave behind and once read six. The sixth was this spare, minted on the hub's init turn a few milliseconds after the first data frame the test had been waiting on — so it landed inside the measurement on a loaded CI runner and outside it everywhere else. The test now takes its baseline after the host reports started, and a new test pins the population a start leaves behind at exactly one hub per source. The investigation is written up in The Read Path Minted a Hub Per Read, §8.