test(relay): stabilize admission diagnostics under delayed lifecycle callbacks #292

Problem

After PR #291 merged, main CI 37892897777 failed in optional relay (macos-latest, python 3.13) and consequently in the release gate. ReceiverAdmission.test_saturation_counters_and_diagnostics_are_bounded_and_sanitized expected one warning but captured two.

The test starts a background Relay listener in setup. Server waitingRooms == 1 is not a barrier for completion of the client's websocket_open callback. Admission and unrelated lifecycle events use the same logger. The test enables diagnostics before joining that listener, so a delayed lifecycle callback can enter its warning-count assertion.

A local controlled scheduling reproduction delays a real owned WebSocket handshake callback until the test's asynchronous close. The unchanged test fails with 2 != 1; captured event kinds are receiver_admission and websocket_open. This proves a matching contamination path, not the identity of the original CI's second record, which the failure output does not include.

Proposed correction

• Cancel and join the background listener, with a bounded timeout, before this counter/log arithmetic test enables diagnostics.
• Retain the exact one-warning assertion, counter totals, sanitized event schema, absence of secret text, minute-boundary assertions and empty-resource checks.
• Add a deterministic delayed-observer regression that fails without the cancellation/join barrier.
• Leave product logging, admission limits, runtime version, packaging, operations and publication unchanged.

Acceptance

• Original controlled reproduction fails before correction and succeeds afterward.
• Delayed-observer regression also fails without the barrier; cancellation alone is insufficient.
• Focused repeated tests, the Relay suite and the current-head CI/release gate pass without weakening assertions or suppressing logging.
• No live agent messages, production changes or release publication.

Related: #291. This is a pre-publication v1.1.0 test-stability correction.