Live lists stop missing a just-created node
A view that watches a set of nodes — the chat token counter over a thread's usage, a notification bell, a folder listing — could miss a node created moments after the view opened, and never recover. Not show it late: never show it, for as long as that view stayed open, while opening the same list again showed the node immediately.
The give-away was how strange it looked. The node was written, correct, and readable: ask for it by path and it came back at once. Only the view that had been watching all along failed to see it.
What was happening
Every live list works the same way: it takes a snapshot, then follows the change notifications that follow. Until now it did that with two subscriptions to the change feed — one that buffered notifications while the snapshot was still being taken, and a second that took over once it was done, at which point the first was closed.
The change feed decides who a notification goes to before it delivers it, which it must, because subscribers come and go while a notification is on its way. So a write published in the instant before the second subscription existed was addressed only to the first — and by the time it arrived, the handoff had happened and the first subscription had been told to stop buffering. It discarded the notification. The second never received it, because it had not been on the list when the notification set out.
Nothing asks again after that. The list re-reads only when a notification arrives, so a lost notification means a lost row, permanently.
What changed
A live list now keeps one subscription for its whole life. It buffers, then switches to following — and the decision about which of the two a notification belongs to is made in the same indivisible step that performs the switch. There is no longer an in-between for a notification to fall into.
The same handoff existed in all four storage backends and was corrected in all of them.
What you will notice
Nothing, most of the time — the window was small. What it removes is a rare, confusing class of "the page is stale and reloading fixes it": a token counter stuck at zero after a reply that plainly used tokens, a folder that does not show a file that is certainly there, a list that quietly stopped keeping up. Token usage was always recorded correctly; only the reading of it could go missing.