[myobie]

This fixes a test-only race behind one failure of registry::keyed::tests::a_record_naming_no_pid_is_probed_but_never_turned_into_a_loss. It failed in #210's Nix job and passed on rerun. The run got Found(Running, Unknown) where Indeterminate(SocketProbe) was expected (keyed.rs:1375).

Cause. The test drops its UnixListener and expects the next probe to be refused. A dropped listener can still accept connections:

• pty-core's unit tests run on parallel threads, and some of them spawn processes.
• A child that another thread is spawning holds a copy of every descriptor, from its fork until its exec closes them.
• A probe that connects in that window is accepted, and the observer correctly reads the session as running.

registry_support::stale_socket documents the same race and already avoids it for the integration tests. It failed 44 runs in 1000 on a Mac on 2026-09-28.

Fix. The test now replaces the dropped listener with a socket made the same way as stale_socket: bound and never listened on. Such a socket refuses every connect, whichever process holds it. The helper also checks that it refuses. The observer is unchanged.

Measured on Linux, at main d997dd3 (its pty-core is identical to this branch's base):

• Deterministic control. Before the drop, I made another thread spawn true with a pre_exec that signals and then sleeps 300 ms, so its child holds the copies while the test probes. With that injection, main's test failed 5/5 with the CI symptom: Found(... status: Running, identity: Unknown) at the refusal assertion. This branch's test passed 5/5 under the same injection.

• Natural rate. Main failed 0/2000 with the test alone, and 0/200 with the whole pty-core lib binary unloaded. Pinned to one core with two spinners, main and this branch ran side by side: the keyed test failed 0/60 on each.

  So on this host the window is rare. CI's build sandbox can be much slower to exec.

• This branch: pty-core lib, 205 passed.

The one-core runs also caught an unrelated timing flake: operation::record::tests::an_observed_exit_is_gone_once_the_zombie_shows_within_the_cap failed 1/60 on main. I'm reporting it separately; it isn't touched here.

[myobie]

Independent read GO at exact head 636bdb8ebddd0a64b07d232838bd4a2358264510. All 9 checks are SUCCESS on this head and the PR is CLEAN.

Read:

• Test-only: one file (crates/pty-core/src/registry/keyed.rs, +38), no product code, no wire-format change, no tier-1 file (events.jsonl, <id>.json, docs/disk-layout.md).
• The flaky test is a_record_naming_no_pid_is_probed_but_never_turned_into_a_loss. It dropped a UnixListener and expected the socket to refuse connections. A child that another test thread is spawning can hold a copy of the listener from its fork until its exec, so the dropped socket can still accept, and the observer then correctly reports the session as running. That is a race in the test, not in the product.
• The fix removes the socket file and binds a socket that is never listened on, built the same way as the existing integration-test helper, and asserts the connect is refused before the observer runs. The test's intent and assertions are unchanged.
• Per the author, a held listener copy injected into the real test made the old version fail 5 of 5 times and the fixed one pass 5 of 5; the race did not show at its natural rate in 2000 runs, which fits CI's slower exec.

Risks (low): the helper adds an unsafe block of plain socket calls to test code (the same technique as the integration helper), which CI compiled and ran.