Shapes endpoint rate-limit descriptors into Bourse.RateLimiter checks and updates rate-limit state from response headers.

Authored buckets execute as token buckets: max_size is capacity and refill_per_sec is the drain rate. A request is admitted when the bucket holds its cost. There is no fixed 60s window and no skip-record exemption.

Endpoints whose authored cost outruns the wait bound

Because no cost is exempt, an endpoint accrues cost / refill_per_sec seconds before it is admitted. Where that accrual exceeds the per-call wait budget (default Bourse.Defaults.rate_limit_max_wait_ms/0) the call returns {:error, %Bourse.Error{type: :rate_limit_exceeded}} naming the venue and the required wait, instead of a skip-record pass-through that let the request go out unlimited and collect the venue's own 429.

Callers that accept a longer wait pass :rate_limit_max_wait_ms on the request (or :max_wait_ms on maybe_rate_limit/4). The budget is per call, so concurrent consumers do not race on process-wide config. Exceeding the budget never bypasses accounting. Long-period endpoints (okx's published one-per-month POST asset/monthly-statement, binance-family POST papi/margin/repay-debt) are refused immediately at the default bound; the error's retry_after is the accrual, not a multi-hour sleep.