[P1][Image Production] Add image understanding, editing, compositing, generation, and deterministic layout pipeline

[Ankit6149]

Parent epic

#151

Product outcome

SignalFlow can take real user-supplied or captured images and produce the exact static visual needed by a ContentPiece without conflating reuse, deterministic layout, generative editing and new-image generation.

Canonical docs

• docs/MEDIA_INTELLIGENCE_AND_CREATIVE_PRODUCTION.md
• docs/CREATIVE_MEDIA_DOMAIN_CONTRACTS.md
• docs/INFERENCE_AND_PRIVACY_ARCHITECTURE.md

Capability classes

Treat these as distinct capabilities:

IMAGE_UNDERSTANDING
IMAGE_GENERATION
IMAGE_EDITING
IMAGE_COMPOSITING
BACKGROUND_REMOVAL
IMAGE_UPSCALE
IMAGE_RESTORATION
OCR
LAYOUT_DIRECTION
VISUAL_QUALITY_CRITIQUE

A provider route may support only a subset.

Production methods

Reuse

Use an existing final candidate unchanged or with allowed normalization.

Deterministic transform/composition

Prefer for:

• crop/resize;
• screenshot framing;
• browser/device frames;
• cards/shadows;
• before/after;
• collage;
• brand/text overlays;
• product UI composition;
• diagrams/layouts where exact typography matters.

Generative edit

Use for semantic edits such as complex object/background changes when explicitly allowed.

New generation

Use when no suitable real visual exists and generated imagery is appropriate/authentic.

Core contract

Implement ImageCompositionPlan with stable source versions, layers, text/layout, generative tasks, deterministic transforms, render targets and revision history.

Safety/truth rules

• do not generatively alter factual product screenshots while presenting them as exact evidence unless explicitly allowed and clearly represented;
• preserve originals;
• reference-only/evidence-only restrictions apply;
• face/identity-sensitive modifications require separate explicit permission;
• remote provider use must satisfy ProcessingPolicy/AssetUsePolicy;
• image generation is not the default for every post.

Example request

Combine these three screenshots into one clean launch graphic, keep the UI exact, use the logo, and don't use any AI-generated background.

Expected method: deterministic ImageCompositionPlan, not generic image generation.

Acceptance criteria

• Provider-neutral image task kinds route through [EPIC][AI Platform] Build provider-neutral inference fabric, privacy-aware routing, local intelligence, and user-owned AI modes #170/[P0][Inference] Add task-oriented provider capability registry, policy-aware routing, metering, and fallback #171.
• At least one deterministic composition path preserves exact source pixels where required.
• At least one bounded provider-backed image-edit route works through canonical contracts.
• Generated output is a new derived Asset with full lineage/provider provenance.
• User can choose/reference/edit/combine intent without changing original bytes.
• AssetUsePolicy and privacy are enforced before AI inspection/editing/public use.
• Exact composition revision is reviewable and approval-bound.
• Quality checks catch dimension/crop/text-overflow/basic visual failures.
• Failure/regeneration preserves prior approved/manual outputs.

Non-goals

• no Photoshop replacement;
• no unrestricted identity manipulation;
• no giant template marketplace;
• no requirement for every provider/model.

Definition of done

Close only when a real user can supply images, request a meaningful edit/composition, receive a traceable reviewable derived result, revise it surgically, and approve the exact static-media revision without the original or permissions being lost.