The guided thinking within the tags works better and doesn't run the
risk of hitting request rate limits; the DR comment poster needs to use
SVG coordinates, not JSON.
We're no longer relying on the LLM to provide viewboxes for comments, so
we can save a significant amount of tokens by stripping all positional
data and elements on the page that are purely positional (e.g. wires, no
connects and junctions)
