`fields` is comma-separated with optional inline types, e.g.
"title, price:float, in_stock:bool". Field names are arbitrary.

A field that could not be extracted comes back as null with an entry in
`misses` explaining why -- that is a normal result, not a failure.

Cost control: `no_llm=True` uses only structured data and already-cached
selectors, so the call is free; `max_llm_calls` caps what an uncached page
may spend. `scroll` and `paginate` cost extra fetches, not extra tokens.

There is no `strict`: it exists on the CLI only to pick an exit code, and
`misses` already tells you what was not found.

Returns the entries that were dropped. Re-derivation is lazy by design: it
needs a page to look at, and it happens on the next scrape of one.
Config-pinned selectors live in a separate store and are never touched.

Call this only when you hit a real bug in reapfield during real work and
have verified a fix. It searches existing issues first; if `duplicate` is
true, stop. Otherwise show the user `body` and ask for authorization -- they
open `submit_url` themselves.

Never include credentials, private URLs, or personal data.