## Rate limiting

Requests are rate-limited per token. When a token exceeds the limit,
requests return `429 Too Many Requests`. Implement exponential backoff
on `429` and `5xx` responses, and avoid tight polling loops.