Feature: Expose JSON responses for headless operations

Right now responses are formatted for manual interaction. It would be nice for automating if the JSON responses from the LLM are optionally exposed.

Expected behavior:
- In: `--query="what is this binary" --output=JSON`
- Out: JSON which has skipped the `pydantic` formatting step

Start at the CLI handling for `--query` and `--output=JSON`, then trace where LLM responses pass through the pydantic formatting step. Done means the example command can optionally emit the LLM's JSON response without that formatting step.

Written by the indexing model from the issue text.