[AlxcBTW]

Both seen on cloud#111 (first attempt of ci run 40618693556).

postgres: played_with::a_pair_is_remembered_only_after_the_minimum_time (left 0, right 1 at line 156). sweep_played_with is throttled by a node-wide atomic: the first caller to compare-exchange it runs record, every other caller returns at once. The test resets the atomic to 0 and calls sweep, but the cloud's own background sweep loop can win the exchange in between; the test's call then returns early while the background record is still writing, and the list is read empty. The test now repeats reset + sweep until the pair shows (10 s limit).

windows: three e2e tests, Email ... failed: rate_limited "Too many requests" at sign-up. The limiter is per cloud state, so tests cannot share one. The same wording is what a saturated store answers (a pool or SQLite write-turn wait past database_acquire_timeout_ms, 5 s, is mapped to 429 "rate_limited" so clients retry). Six e2e tests with relays run in parallel on the Windows runner; the first calls wait past 5 s. The e2e clouds now use a 60 s acquire timeout. Production limits and defaults are unchanged. (The warn log of the busy store is not captured in CI, so the cause is inferred from the code path, not seen directly.)

---

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