test: acp-goal-plan-child-process times out on main @ 0fd754083 (AskUserQuestion never resolves) #5777

What happened

The test job on main is red at 0fd754083; its parent ca1ac87ca was green. The failing file is the runtime-host suite's acp-goal-plan-child-process.test.js, subtest "submits and executes a Plan through the model, ACP, and real Host" (test at line 481):

Error: ACP client operation timed out after 15000ms; exitCode=null signal=null

The child-process stdout in the failure dump shows the session reaching end_turn on the plan-approval turn and _maka/plan/changed carrying an activeExecutionId, then a tool_call for AskUserQuestion ("Proceed with the fixture?") entering in_progress — and nothing answering it. The file then runs to the 900s harness timeout ([cli] timed out after 900000ms), so the later subtests in the file are lost with it.

How to reproduce

Run the runtime-host test job on main. Observed on the shared CI runners at 2026-09-28 02:16-02:40 UTC across three runs on two unrelated PR heads plus main itself, all with the same failure signature; the runs that checked out ca1ac87ca passed.

Environment

• Maka commit: 0fd754083 (red) vs ca1ac87ca (green)
• Surface: Runtime Host test suite (ACP Goal/Plan real Host routes), ubuntu CI runner
• Node.js: 24 (CI)

Context

The correlation point is #5434 (squash-merged as 0fd754083): its own PR-head test job was green, but the affected-test scope on main includes this ACP file, and the red starts exactly at that merge. I have not established causation — the failing step is a 15-second no-response on an interactive tool call, so a timing-sensitive flake that happened to surface at the merge is also possible. Flagging the correlation because the red gate now blocks unrelated pull requests: #5265 and #5744 both fail at their current heads on this exact test while their own changes are untouched by it.

_{Filed on behalf of @ggbdpq by GLM-5.3-Flash (ZCode).}