Grounded image generation detection and evidence gating

[paullizer]

Summary

Implement Phase 1 of the grounded generation orchestration roadmap: detect grounded image generation requests and gate image proposal behavior until evidence gathering has been attempted.

User Value

Users can ask for image outputs grounded in selected images, workspace content, public web research, agents, actions, or enterprise sources without the assistant producing an unsupported image proposal too early.

Proposed Behavior

• Detect image generation intent combined with grounding or evidence requirements.
• Return structured detection state, including requested evidence sources and missing source warnings.
• Inject system guidance requiring evidence collection before simpleimage proposal generation.
• Preserve normal behavior for generic image prompts and non-generation image questions.

Acceptance Criteria

• Grounded image requests such as M365, LinkedIn/public profile, SQL, CRM, workspace document, selected agent/action, or selected image plus role/collaborator/value prompts are detected.
• Simple image prompts without grounding are not marked as requiring evidence.
• Selected image analysis or image summarization requests are not misclassified as image generation.
• Chat/image proposal guidance receives the grounded request state and instructs the model to gather evidence before proposing generation.
• Focused functional tests cover positive and negative detection cases.

Notes

Roadmap reference: priv-simplechat/01_PHASE_1_GROUNDED_IMAGE_REQUEST_DETECTION.md. This issue tracks the first implementation slice toward the broader grounded generation orchestration capability.
Priority: P0.
Inferred size: M.