[AlxcBTW]

The flake: a_starved_pool_answers_429_with_retry_after_within_the_acquire_timeout failed on Windows with left: 429, right: 200 from Reply::ok(200), not from the assertion about the starved pool. The test builds a cloud with a two-connection pool and a 500 ms acquire timeout and then signs up through the API. On a slow runner the cloud's own background work (sweep, heartbeat) holds a connection while the sign-up needs one, so the sign-up itself was answered 429. The earlier fix (#103) only made the check after the release patient.

Now every test of this file does its setup patiently (until_served asks again while the answer is 429, sign_up uses it) and hold takes the whole pool, waiting for the background work to give connections back. From then on the pool is empty by the test's own doing. All assertions are unchanged (429 with rate_limited, Retry-After of 1 to 2 s, timing bounds, 503 for the request timeout, the shed and health tests).

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}