fix(test): poll the subtree re-resolution instead of sleeping once and hoping

[wshallwshall]

test_subtree_re_resolution_picks_up_a_late_spawned_child failed on windows-2025 with AssertionError: assert 8204 in [] — twice in the same job (pytest-rerunfailures already retried), while windows-2022 and ubuntu passed the identical commit.

The empty list is the tell

It isn't "the walk ran and missed the child". It's "no walk succeeded at all".

On Windows _descendants_windows shells out to Get-CimInstance Win32_Process with a 5 s timeout. Under load that times out and it returns None — deliberately, to signal ERRORED rather than "no descendants". _resolve_pids then does exactly what its docstring promises: returns root-only for that tick, does not cache, and leaves _pids unresolved "so the next tick retries".

The test gave it one tick. So it asserted a stricter property than production requires — that a single enumeration always succeeds — on a probe explicitly written to tolerate one that doesn't. time.sleep(1.0) then one sample_proc() is the same sleep-and-hope shape df607719 removed from the /ws/stats tests hours earlier.

The fix

Poll sample_proc() until the child appears or a deadline expires — which is what the production caller does, tick after tick.

The deadline is derived, not hardcoded: max(30.0, 6 * _PROBE_TIMEOUT_S), imported from the probe. Raising the probe's per-walk timeout therefore can't silently leave this deadline too short to fit even one attempt.

The bigger fix is the diagnosis

The old assertion collapsed two unrelated failures into one useless message. They now report separately, and they mean opposite things:

• "the process-table walk never succeeded in 30s (N attempts)" — the probe couldn't measure at all. An environment problem; the test could not assess re-resolution either way.
• "the walk SUCCEEDED but did not include the late-spawned child <pid>" — the A3 regression this test exists to catch: a subtree resolved once pins the sampler to an idle parent, so a sharded engine's serve --shard workers are never counted.

Verification

Each regression injected into probe.py, checking both the failure and its message:

injected	caught	message
enumeration always errors (None)	✅	"never succeeded"
enumeration succeeds, no descendants	✅	"did not include the late-spawned child"

probe.py verified byte-identical to HEAD afterwards — no production code changed.

Stable rather than merely passing: 5 consecutive runs, ~2.1 s each — it exits as soon as the child appears, so it's no slower than the fixed 1 s sleep it replaces. Full module 12 passed; with test_connscale_smoke, 14.

The 24 mypy errors in this file are pre-existing (main's unmodified copy reports the same 24) and outside CI's mypy scope, which covers messagefoundry + messagefoundry_webconsole, not tests/.