fix(rivetkit-core): close attached connections at the sleep grace deadline

[eersnington]

#5860 runs the sleep hooks, including connection disconnects, after in-flight work drains. If that work outlives the grace period, connections can still be attached when the grace deadline fires.

• At the grace deadline, the actor closes every connection that grace closes: all of them on destroy, and the non-hibernatable ones on sleep. It closes the transports directly, because hooks dispatched at the deadline are never awaited.
• emit_grace_events uses the same closes_on_grace filter, so the hibernation rule lives in one place.
• Adds a drain test: an action in flight when the runner drains still returns its result to the client.

// rivetkit-rust/packages/rivetkit-core/src/actor/task.rs, on_sleep_grace_deadline
self.ctx
	.disconnect_conns(|conn| closes_on_grace(grace.reason, conn))
	.await

---

This is part 6 of 6 in a stack:

•  6  fix(rivetkit-core): close attached connections at the sleep grace deadline #5871 👈
•  5  fix(rivetkit): run onSleep after in-flight work and refuse dispatch once it starts #5860
•  4  fix(rivetkit): refuse actions after onSleep with actor.restarting #5876
•  3  fix(rivetkit-core): close drained connections with ws.going_away #5867
•  2  fix(pegboard-gateway): send the client close without waiting on the envoy #5874
•  1  fix(rivetkit): retry connect after guard actor_ready_timeout #5866

[claude]

Review of #5871

The change is small and targeted, and pulling the hibernation rule into closes_on_grace means emit_grace_events and the deadline path can't drift apart. I read the diff only and did not build or run the tests.

Possible issues

• Deadline path is not covered by the new test. work sleeps 300ms, which is almost certainly well inside the grace period. The test therefore exercises the drain-then-hooks path from fix(rivetkit): run onSleep after in-flight work and refuse dispatch once it starts #5860, not the new on_sleep_grace_deadline close. Consider a Rust test in rivetkit-core/tests/ or a fixture action that outlives the grace period. It should assert that non-hibernatable connections are closed and hibernatable ones are kept on sleep, and that all are closed on destroy.
• disconnect_conns(...).await at the deadline. The comment says hooks dispatched at the deadline are never awaited, so this path is already time-critical. Please confirm disconnect_conns cannot block on a slow transport close or user onDisconnect. If it can, bound it with a timeout or run it concurrently with the run_handle abort. A failure is only logged with warn!, which is fine, but a hang would stall teardown.
• Double cleanup in the new test. startDrainingRuntime registers runtime.cleanup() in onTestFinished, and the test also calls await runtime.cleanup() inline. Confirm cleanup is idempotent.
• Weak assertion. The test checks only that the action returns "done". It does not check that the connection was then closed with ws.going_away, which the first test does check.

Nits

• In actor-conn-drain.test.ts, the reflow of output.split(...) looks like an unrelated formatting change.
• In task.rs, the warn! message follows the logging conventions. It could include the likely cause, per the CLAUDE.md guidance on warning logs.
• The doc line in sleep-sequence.md is accurate and concise.

Overall this is a reasonable fix. The main gap is test coverage of the deadline close itself.

🤖 Generated with Claude Code