read-figure

Give text-only models a fixed three-section figure-reading handoff: a vision model describes a physics figure (structure / restatement / confidence), the tool returns it as text, and uncertain marks must be listed instead of guessed.