feat(parser/generator): body images as objects — same blob + anchored-interior model as tables/text boxes (#300 follow-up)

[thewrz]

Why

#300 establishes the body object model (verbatim-XML blob + document-order slot + anchored interior text) for tables and text boxes. Body images are the third editor-facing content class that currently drops silently from the body path (H/F images are handled — ADR-069/#487 — but a picture sitting in the spec body is not).

What (when picked up)

Extend the #300 object model to kind: 'image':

• Capture body w:drawing (inline + anchored/floating) and VML w:pict images as object nodes in document order (floating → host-paragraph slot + meta.floating, same KNOWN AMBIGUITY posture as feat(parser/generator): represent + round-trip visible tables in specs (AST model, inference exclusion, render) #300 decision 3).
• Unlike tables/boxes, the blob alone is insufficient: the subtree references word/media/* via r:embed — the object must also carry the media bytes + relationship (reuse feat(parser): extract embedded DOCX header/footer logos into structured image fields #487's H/F ingress machinery: sniffImageMediaType, EMU passthrough, 5MB cap posture from ADR-069/070).
• Generator re-emits via the blob + re-registered media part/relationship.
• No interior text (images have no objectText children) — editability locked wholesale.
• Markdown: labeled placeholder (**[IMAGE: name/alt, W×H]**).
• Out-of-scope drawables keep warning per feat(parser/generator): represent + round-trip visible tables in specs (AST model, inference exclusion, render) #300 decision 10 until this lands.

Sequencing

Blocked by #300 WS1/WS2 (the object model must exist). Media-part plumbing is the only genuinely new piece.

Acceptance

• Body DOCX with an inline photo and a floating logo round-trips with both images present, bytes/dimensions intact.
• Floating image anchor geometry preserved verbatim; host-paragraph slotting pinned as KNOWN AMBIGUITY.
• Dropped-image parse warnings from feat(parser/generator): represent + round-trip visible tables in specs (AST model, inference exclusion, render) #300 are replaced by real capture.