feat: add Cloudflare Workers AI fallback for kimi models

feat: non-streaming auto-continue, higher timeout, default max_tokens

feat: auto-continue + keepalive for streaming

feat: API key stats, error capture, key states tracking

feat: proper SSE streaming via pipe for stream:true

feat: use native https module instead of fetch for upstream

fix: add Content-Length header explicitly

refactor: read body in proxy, return text directly

fix: use text() instead of arrayBuffer() for non-streaming

fix: handle non-streaming responses with arrayBuffer

feat: impersonate Kimchi CLI headers (User-Agent, X-Stainless-*))

fix: key rotation - zero delay on 402, 55 retries, handler timeout 55s

perf: reduce timeouts to fit Vercel 60s limit

feat: add 402 to retryable statuses for credit exhaustion rotation

feat: key rotation on retry - rotate key per attempt

fix: undefined attempts var in proxy.js

init: kimchi proxy with multi-key rotation