[sproctor]

Builds on #1433, now merged into master; master is merged in here.

What changed

1. A message with ref 0 could be taken for a heartbeat reply. heartbeatRef used 0 to mean "no heartbeat pending", but refs start at 0 on every socket. While no heartbeat was pending, any message carrying ref 0 matched and was dropped as a heartbeat reply. The server's phx_close or phx_error for a socket's first join is one: it carries that join's join_ref, 0, as its ref. And a heartbeat that itself got ref 0 (the first message on a socket with no channels) counted as not pending, so a missing reply to it went unnoticed for an extra interval. The sentinel is now -1 (NO_HEARTBEAT).
2. Only messages on the phoenix topic are taken for heartbeat replies. Heartbeats are sent on that topic, so a channel message can no longer match the pending heartbeat, whatever its ref.
3. sendHeartbeat() sends the ref it made. It read the ref back from heartbeatRef to send it, and a reset on connect in between could change it.

#1423 (merged into master after this was opened) resets the heartbeat state when a socket connects. It stored 0, which with this change is a real ref: every new socket would start with heartbeat 0 pending and time out on its first tick. The merge makes it store NO_HEARTBEAT. This branch's own reset in disconnect() is dropped, since every socket goes through connect().

Why it builds on #1433

Swallowing the first join's close hid a race: when a channel rejoins (enabling presence on a subscribed channel, or after a channel error), the server's close for the join it replaced can arrive afterwards, and RCloseEvent removes the channel. #1433 drops messages for an earlier join_ref, as Phoenix's client does. Without it, this fix alone would let a rejoin lose the socket's first channel, which master protected only by accident.

Testing

• RealtimeHeartbeatTest, on the test clock:
  • a channel close with ref 0 reaches its channel;
  • an unanswered first heartbeat (ref 0) times out on the next tick;
  • fix(realtime): reset the heartbeat ref when a new socket connects #1423's lost-socket test: the next socket isn't torn down for the lost one's unanswered heartbeat. It now uses ScriptedWebsocket (with a new lose()) instead of its own socket, and also answers the new socket's heartbeat and checks that the next one goes out on the same socket.
• With the connect reset storing 0, the two heartbeat tests fail. With a topic other than phoenix matched, the lost-socket test fails. The ref-0 close test fails on master, where both the topic check and the sentinel are missing.
• fix(realtime): handle channel errors and rejoins like realtime-js #1433's rejoin tests now send the server's real ref (the join_ref) for its errors and closes, which they couldn't while ref 0 was swallowed.
• ./gradlew :realtime-kt:jvmTest :realtime-kt:detekt pass.

🤖 Generated with Claude Code

[jan-tennert]

@sproctor Might need to resolve the conflict

[sproctor]

Thanks. Yeah, this conflicted with #1423. I'll clean it up.

[jan-tennert]

Thanks, I think the tests broke too

[sproctor]

Thanks. It was a flaky test. It passed here.