Makes the stdio interaction test's subprocess flush its coverage data before printing the clean-exit line, so the transport's terminate escalation can no longer silently destroy the data file.

Motivation and Context

checks / test (3.10, locked, windows-latest) failed on #2838 with every test passing but the 100% gate tripping on tests\interaction\transports\_stdio_server.py at 52.17% — exactly the subprocess-only lines (failing job).

Root cause: the child only persisted its coverage data in coverage's atexit hook, which runs after the stdio-echo: clean exit print the test synchronizes on. The failing run's test took 12.40s ≈ 2.4s session + exactly the 10.0s PROCESS_TERMINATION_TIMEOUT the test monkeypatches — the child printed the line (test green), then didn't finish interpreter teardown + the sqlite save within the grace, and the Windows escalation (TerminateJobObject) killed it mid-save. A torn save leaves a valid-but-empty data file that coverage combine merges without a warning, which is why the failing run still combined the same 6 data files as passing runs.

This is the same race #2773 narrowed by raising the grace from 2s to 10s; the wider grace made it rarer, not impossible. Extracting this test's wall time from ~400 recent Windows CI job logs shows a smooth heavy tail — p50 3.3s, p99 ~7s, then 8.8s and 9.0s (passed, barely under the grace) and 12.4s (killed) — so any finite grace loses occasionally as long as the data write happens after the observable sync point.

The fix inverts the ordering instead: main() now stops and saves coverage before printing the clean-exit line. That makes the test's existing stderr assertion the durability barrier — once the parent has seen the line, the data is on disk and a late kill is harmless. A kill landing before the print now fails the stderr assertion loudly instead of tripping the coverage gate one step later.

Details that fall out of the ordering:

• Nothing measured can execute after the flush (it would be unrecordable by construction), so the trailing cov.save() and print(...) lines carry # pragma: lax no cover.
• cov.stop() is load-bearing twice: it ends tracing, and it leaves nothing for coverage's atexit re-save to flush — so a kill during interpreter teardown can't corrupt the already-written file (coverage opens it with sqlite journaling off; a torn rewrite would not roll back).
• The if cov is not None branch is # pragma: no branch: under coverage the instance always exists, and without coverage nothing is measured anyway.
• The test's monkeypatch comment is updated to describe the new invariant; the 10s grace stays (it now only bounds how long a pre-print straggler gets before a loud failure).

How Has This Been Tested?

• Reproduced the failure mode deterministically: with a 3s post-print stall injected into the child and a 1s grace, the pre-fix code passes the test while the child's data file goes missing (the CI signature on demand); the fixed code under the identical kill retains the complete data file.
• coverage run -m pytest tests/interaction/transports/test_stdio.py → both tests pass, _stdio_server.py reports 100.00% (statements + branches) from the combined parent+child data.
• Full ./scripts/test: 100.00% total, strict-no-cover clean.
• ~2,500 instrumented spawn/close cycles of this exact flow on windows-latest runners (and ~2,200 on Linux) while investigating: the child's post-print phase is interpreter teardown plus the coverage sqlite save, and the save is the phase that runner load stretches — which is what made the atexit-ordering the wrong side of the sync point.

Breaking Changes

None — test-support change only.

Types of changes

• Bug fix (non-breaking change which fixes an issue)

Checklist

• I have read the MCP Documentation
• My code follows the repository's style guidelines
• New and existing tests pass locally
• I have added appropriate error handling
• I have added or updated documentation as needed

Additional context

The import coverage at module top makes this test-support file depend on the dev environment even for the docstring's standalone python -m invocation; that's already true of the repo's dev venv everywhere this module is used, so it keeps imports-at-top rather than adding a guarded import plus more pragma machinery.

AI Disclaimer