Vision fallback (Set-of-Marks) for canvas / poor-markup pages

`look()` is a11y-tree-first, which is right ~90% of the time but returns near-empty on canvas-heavy or poorly-marked-up pages.

**Task:** when the a11y tree is sparse, fall back to a screenshot with numbered Set-of-Marks overlays so elements are still addressable by stable integer ref (never raw coordinates). `look({ mode: "screenshot" })` already exists as the seam.