implement-multimodal-observation

name	implement-multimodal-observation
description	Implement typed image/state observations, deterministic chess rendering, multimodal provider lowering, redundancy or conflict conditions, and modality provenance.

Implement multimodal observation

Workflow

1. Define canonical state independently of its visual rendering.
2. Render image into immutable CAS bytes with MIME, dimensions, theme, orientation, coordinates, renderer version and source-state hash.
3. Build typed ImageArtifactPart; never rely on mutable remote URL by default.
4. Declare observation profile and modality authority policy.
5. Add provider capability preflight for image input.
6. Preserve canonical request and provider-lowered form.
7. For redundancy, confirm image and symbolic state hashes map to the same state.
8. For conflict experiments, generate deliberate delta and record conflict manifest; never let accidental mismatch pass.
9. Test themes, orientation and resolution as controlled variables.
10. Separate perception/state reconstruction from move quality in metrics.

Required controls

• image only;
• symbolic only;
• consistent image + symbolic;
• one-piece conflict;
• explicit textual-authoritative vs image-authoritative instruction;
• no-image provider must fail preflight, not silently OCR/describe.