withdrawBridge re-asked every unsettled answer, so on the never-started
path a helper that errors (a hung or blocked interpreter) cost 3 x the 20 s
MARKER_HELPER_TIMEOUT_MS with the user waiting. Only `busy` (the sidecar
lock, usually brief) is re-asked now; every other answer stands at once
and the existing settled / quit decision is unchanged. The existing test's
"asked three times" for `error` encoded that cost and now expects one ask
(review K132354 P3 withdraw-retries-error).
c3d19d497a