Vision input overview

Ability to accept image inputs (screenshots, photos, pasted clipboard images) as part of a user/agent turn. Required by interaction primitives that paste or attach visual content. Distinct from `text-streaming` or other text-only input modalities.