fix(mcp): stop leaking an unawaited watcher coroutine per stdio tool call (salvage #96044)

[kshitijk4poor]

Summary

The stdio fast-fail gate no longer leaks an unawaited coroutine on every MCP stdio tool call: the _watch_ok probe inspected awaitability by calling the child watcher (inspect.isawaitable(_watch_children())), creating a fresh never-awaited coroutine per call ("coroutine was never awaited" RuntimeWarning spam + GC churn). It now inspects the function without invoking it.

Salvage of the unique hunk from #96044 by @loulanyue (authorship preserved). The other half of that PR — the _stdio_children_dead polarity fix — was already on main via #94339.

Changes

• tools/mcp_tool.py (@loulanyue): probe _watch_children without calling it. Our follow-up tightens the replacement to inspect.iscoroutinefunction(_watch_children) alone — the originally proposed or callable(...) arm would have flipped MagicMock-stubbed sessions into the fast-fail race (callable(MagicMock) is True), breaking the plain-await path the surrounding comment routes stubs to.
• 2 regression tests: source pin that the probe no longer calls the watcher (behavioral test infeasible — the leak only manifests as a GC-time RuntimeWarning), plus a semantic contract test (real async def / AsyncMock → race path, MagicMock → plain await).

Validation

tests/tools/test_mcp_stdio_children_dead.py + test_mcp_tool.py	109 passed
Mutation check (mcp_tool.py reverted to main)	pin test fails → restored passes
Probe semantics (live)	iscoroutinefunction: real async def True, AsyncMock True, MagicMock False
Behavior-delta audit	only disagreement case (sync def returning awaitable) unreachable in-repo, degrades to plain await (safe)
Sibling sweep	25 other isawaitable() sites all probe already-produced values — this was the only call-to-probe
ruff	clean

Credit

Coroutine-leak find and fix by @loulanyue in #96044. Closes #96044 (its polarity half was superseded by #94339).

First report of the leak: @glupta in #95938 (also reported independently by @ilgo in #96030). @liuhao1024 shipped an equivalent earlier fix in #95939 (create-once-and-reuse shape) — see that PR for the credit note on why this shape was preferred.

Closes #95938. Closes #96030. Refs #96016.

[Enough1122]

AI code review — automated review for reference; please use your judgment.

fix(mcp): stop leaking an unawaited watcher coroutine per stdio tool call — correct leak fix

• Fix: tools/mcp_tool.py:6174 changes inspect.isawaitable(_watch_children()) → inspect.iscoroutinefunction(_watch_children). The old probe called the async watcher to test awaitability, creating a coroutine that was never awaited — RuntimeWarning: coroutine was never awaited under -W error and GC churn per fix(mcp): correct _stdio_children_dead logic and prevent coroutine leak (#96016, #96030) #96044.
• Semantics preserved: test_mcp_stdio_children_dead.py:11-40 proves iscoroutinefunction splits the world identically for the fast-fail race: real async def and AsyncMock → True, plain MagicMock (used in unit stubs) → False, so mocked watchers still take the plain-await branch without the leak.
• Guard: source-level test asserts isawaitable(_watch_children()) no longer present and iscoroutinefunction(_watch_children) present — durable.

Non-blocking — minimal, precise.

[teknium1]

Landed on main via #100920 (rebase-merge, your commits preserved as-authored: fix by @loulanyue, tests by @kshitijk4poor — main fe3e5dc). The branch here had gone stale against main so it was re-based in a fresh PR rather than pushed to. Thanks for carrying the salvage.