Fine-tuning ~ saving them tokens

- [x] Don't send the full-doctype list in context, support search or nudge filtering.
- [x] `doctype show ` should only show fields that make sense. A lot of fields have no relevance to "assistant" behaviour... all wasted tokens! Blacklist unnecessary fields (so new stuff can stil keep coming up on its own). Reveal minimal fields, ask again for more. Add some sparse schema. Flags for just getting fields, or dumping it out and processing.
- [ ] large result sets should be stored off-context and LLM should be nudged to inspect that instead of loading it all in context.
- [ ] Some kind of `--agent` mode for different behaviour instead of standard behaviour. We don't want scripts to break.
- [x] compact JSON
- [x] Revisit `--raw`.
- [ ] Give summary first? Count, pagination info etc.

Analysis from pi session 019fa7f6-550b-7e30-b6d7-0586c365e676

Start by separating the unchecked requests around large result sets, --agent behavior, and summary output from the completed items. Review the existing doctype show and --raw CLI behavior, then define a focused scope with clear token-saving and compatibility checks before implementation.

Written by the indexing model from the issue text.