Building a distributed tokens-per-minute limiter for LLM APIs with one Redis Lua script

Token buckets, reserve-and-settle leases, crash-safe concurrency slots, and why a 429 isn't always a rate limit