`src/stream/rate_limit.rs` — the token bucket

The proactive rate limiter: wait before firing into a throttled endpoint, instead of only reacting to 429s.

Key items

• TokenBucket::new(capacity) — burst = capacity; continuous refill at capacity/60 tokens per second; lazy refill on access (no background task).
• take() / take_at(instant) — consume one token → Ok(()) or Err(RateLimitError::Wait(duration)) (time until a token refills; Duration::MAX for zero-refill buckets) or Err(Poisoned).
• available() / available_at() — non-consuming queries.
• RateLimiter::new(requests_per_minute) — 0 disables; one bucket per base URL (lazily created, shared).

Behavior notes

• Provider identity is the base URL: OpenAI and Ollama get independent buckets; two clients on one endpoint share.
• The handler’s gate_on_rate_limit sleeps in slices bounded by rate_limit_max_wait (30s default) and then proceeds anyway — better to risk a 429 than hang.
• Reactive throttling (honoring 429 responses) lives in the handler’s retry ladder; this module is the proactive complement.

Deep dive: Stream events.