Deflake the runMcpFlow 429 backoff test (#2548)
Wait for the backoff timer instead of racing it in the 429 spec
The rate-limit test advanced fake timers on a fixed 200-iteration budget
that a slow runner could exhaust before the flow ever registered its
backoff sleep, leaving the fake clock unadvanced and the test hanging to
its 30s timeout (flaked on CI run 33462543653). The backoff sleep is the
only client of the faked setTimeout, so wait for vi.getTimerCount() to go
nonzero — yielding real event-loop turns however long that takes — then
advance by exactly RATE_LIMIT_BACKOFF_MS[0]. Reproduced the hang by
delaying the scripted 429 behind 2000 event-loop turns: the old loop
times out, the hardened one passes in ~50ms.
