Simulate the three buckets and the cache cliff

The simulator is a provider, not a client: it decides admit or 429 for each request, and you watch the result. Everything is pure Python with no network and no model, so ten simulated minutes run in a fraction of a second. The decision that matters most is when each bucket is charged: RPM and ITPM at admission, OTPM only as output tokens are produced. That is what the docs describe, and it's why an admitted request can still push OTPM into debt and cause the next request's 429.