llm-fetch — fetch a web page as Markdown.

  POST /api/auth    {"code": "123456"}            -> {"token": "...", ...}
  POST /api/read    {"url": "https://..."}        -> {"markdown": "...", ...}
  GET  /api/session                                -> remaining quota
  POST /api/logout                                 -> drop the session

Authentication (required):
  Ask the operator for the current 6-digit TOTP code, POST it to /api/auth, then
  send `Authorization: Bearer <token>` on every later call. The session lasts
  300 s and is extended by each request, up to 7200 s total.
  A code can be traded for a session exactly once.

/api/read body:
  url        required, http(s), public hosts only
  need       "text" (default) or "live" — "live" skips straight to a real browser
  max_stage  1 plain, 2 browser-TLS, 3 extraction API, 4 headless Chromium, 5 archive
  max_chars  truncate the Markdown (default 40000, 0 = no limit)

The response reports which stage produced the text and what the earlier ones
failed with, so a caller can tell "the page is empty" from "the page blocked us".

Errors: 401 no/expired session, 403 URL rejected by the SSRF guard,
429 quota exhausted or too many failed codes, 502 every stage failed.

This page is the whole protocol. /api/docs and /openapi.json are LAN-only.