Autocomplete waits for every source before it answers

Typing @ gathers suggestions from several places at once: what is nearby in the node you are editing, what the search index knows, what other partitions hold. Some of those answer in a millisecond; the ones that reach across partitions take longer.

The list used to be sent out as soon as it had stopped changing for a moment. That reads as "everything has answered", and most of the time it was — but a moment of quiet is not the same thing as being finished. On a busy portal the slower sources simply arrived after the pause, and the list you saw was the one assembled without them. Nothing reported this: the suggestions carried a flag meaning this is the complete list, and it was set whether or not it was true. The symptom people met was a name that was there yesterday and missing today, with no way to tell which of the two was right.

Suggestions now go out when every source has finished, which each of them reports for itself. Waiting for the slowest costs nothing when they are all fast, and when one of them is slow the list is simply a little later — and correct.

A source that never finishes at all can no longer hold your suggestions hostage either: after two seconds the best list so far is sent anyway, and it is now honestly marked as incomplete, with the misbehaving source named in the portal's log so it can be fixed rather than absorbed.