A null doc presence replays ops just to advance its version #747

Summary

Every p: null doc presence broadcast reads the op log, only to advance the presence's v.

Agent._broadcastPresence handles a null presence by calling _unsubscribeDocVersion first, which clears latestDocVersions[c][d]. The up-to-date check that follows compares against that cleared value, so a null doc presence never counts as up to date. It always goes through Backend.transformPresenceToLatestVersion, which runs getOps from the presence's last v to the latest version, and then op middleware on each op.

If a cursor sat idle on a busy doc, that read can be long. A mass disconnect, such as a deploy or a load balancer drain, turns every open doc presence into one of these reads at the same time.

The presence still needs a current v. A null presence with an old v makes the receiving client fetch the doc and wait for ops to catch up (RemoteDocPresence), so skipping the transform outright isn't enough.

Suggested fix

Give a null presence the doc's current version from a snapshot or version read, rather than replaying ops to transform a value that is null anyway.