Add subject-aware local adjustment and mask workflows

[cgasgarth]

Goal

Enable subject-aware local adjustments so the agent can target faces, skin, skies, products, and other high-value regions instead of relying only on global edits.

Current state

• We now have lightweight image-analysis signals and coarse region summaries (for example sky/skin candidates) from the richer image-state work.
• We do not yet have a mask-generation pipeline, mask transport format, or a darktable-side application path for subject-targeted local edits.

Scope

• Define how the agent requests or receives masks for common subjects and regions.
• Connect those mask artifacts or region selections to darktable modules such as tone equalizer and retouch-oriented workflows.
• Keep the architecture compatible with external vision helpers instead of forcing all perception into the planner.
• Document safe fallbacks when masks are unavailable, low-confidence, or ambiguous.

Acceptance Criteria

• The roadmap supports subject-targeted local edits without breaking current protocol guarantees.
• Mask generation and mask application responsibilities are clearly separated.
• Local-adjustment workflows can be tested independently from global edit planning.
• Failure modes are explicit when subject detection or masking is unavailable.