relay: the CLOSE-during-mint test synchronises on a 50 ms sleep and can fail under load #963

a_close_during_the_invite_mint_leaves_the_event_but_not_the_read (crates/nostream-relay/tests/handlers.rs:1135, added in #962) synchronises the CLOSE against the invite mint with a fixed sleep:

gate.add_permits(1);

...is preceded only by tokio::time::sleep(Duration::from_millis(50)) after ["CLOSE","invite"] is sent. Nothing confirms the server has processed the CLOSE — i.e. that unsubscribe::handle has run abort_read and set the in-flight read's cancellation — before the gate releases the mint.

Failure mode: a false failure, not a false pass. If a busy runner delays the CLOSE handler past the sleep, the mint resumes first, the handler reaches fetch_and_send with the cancellation not yet set, and the test then observes the stored EVENT and the EOSE it asserts must not arrive. The fix is correct and the test goes red anyway. The window is small (50 ms against a local in-memory store) and the test passed in CI on #962, so this is a flake risk rather than a live defect.

Suggested fix. Wait for a server-side signal instead of for elapsed time. The crate already has the passive half — Connection::read_in_flight(id) is observable — so the options are:

1. Poll fixture.connection(id).read_in_flight("invite") == false (the abort removes the entry) before gate.add_permits(1), with a bounded timeout that fails with a clear message rather than hanging.
2. Add an observable for "a read was aborted" on the test double, mirroring the existing gate_reads / gate_reads_per_item / gate_creates idiom in testing.rs, and await it.

Either way the release becomes ordered after the abort by construction, and the 50 ms sleep goes.

Residual risk worth stating in the fix: the CLOSE's handler runs in a spawned task (relay.rs spawns each client message independently), so a wait on server-side state is the only ordering that does not depend on scheduling. That is also the mechanism behind #958, which is a separate pre-existing defect and not this test's problem.

Found by the Greptile review on #962 (P2). Recorded rather than fixed there, per the rule that a bot finding which is not a showstopper becomes its own issue instead of blocking the merge that revealed it.