fix(tests): drain blocker completions in the async-delegation capacity test

[Soju06]

Problem

tests/tools/test_async_delegation.py::test_crashed_runner_produces_error_completion fails intermittently in CI (and locally) with:

AssertionError: assert 'completed' == 'error'

test_dispatch_rejected_at_capacity releases its two blocker workers with ev.set() and returns immediately. The workers' completed events land on the shared process_registry.completion_queue asynchronously — often after the next test's autouse start-of-test drain has already run on a still-empty queue. The crashed-runner test's _drain_one() then picks up a leaked completed event from the previous test instead of its own crash's error event.

Reproduction (running just the two tests in sequence): 6/15 runs fail on main, 0/15 with this fix.

This flake currently trips Run tests slices on unrelated PRs (e.g. it failed three different PR runs today on three different slices).

Change

Test-only: the capacity test consumes both blockers' completion events (_drain_one() × 2) before returning, so nothing leaks past the next test's drain. Assertions unchanged.

Tests

• Full module: 27 passed (×3 consecutive runs).
• Targeted pair stress: 15/15 clean with the fix vs 9/15 on main.

🤖 Generated with Claude Code

[Soju06]

Noting #64431 removed the flaky test outright — this PR is the alternative that keeps the coverage: the flake was the capacity test leaking its blockers' late completion events into the next test's drain, not a defect in the crashed-runner test itself (0/15 failures with the drain vs 6/15 on the old main). Happy to rebase this to restore the removed test alongside the drain fix if you'd rather keep the assertion; equally fine closing this if the deletion is the intended end state.

[teknium1]

Thanks for tracing the completion-queue race and documenting the alternative to #64431. This is an automated hermes-sweeper review; current main has superseded this patch.

• f5c2ea49a4716ccd377c99511ee1bb9be1b56275 added module-level teardown at tests/tools/test_async_delegation.py:28-37, which waits for async workers to finish before draining the shared completion queue. This covers the late-completion leak for the capacity test and other tests in the module.
• The wait covers the enqueue window: active_count() includes finalizing records at tools/async_delegation.py:410-413, and completion events are enqueued before the record becomes terminal at tools/async_delegation.py:815-826.
• c084085a3e520178fb3aa27c6ba7a411df8c79a0 (test: remove flaky test_crashed_runner_produces_error_completion #64431) removed test_crashed_runner_produces_error_completion, the test this PR was intended to preserve. As noted in the discussion, restoring deterministic coverage would be a separate follow-up if desired.

Closing as implemented on main.