feat(image-gen): add multi-provider fallback chain + north-star onboa…

[Clover-ASH]

…rding + body-fat ensemble

This PR consolidates the image-generation pipeline and the supporting body-fat assessment upgrades that drive it:

• backend/src/image-gen/image-gen.module.ts

  • Add L0 (image2) -> L1 (nano-banana-2) ordered fallback chain
  • Persist resultImageUrl on success (fixes stage0 preview lookup)
  • Per-provider proxyUrl override (kept IMAGE_GEN_PROXY_URL for back-compat)
  • Image-edit retry: drop the optional reference image and retry once
  • Tag generations with provider label (L0:image2 / L1:nano-banana-2)

• backend/src/evolution-stage/evolution-stage.service.ts

  • New generateNorthStar(userId, startImageUrl): onboarding north-star pipeline (vision body-fat estimate -> persist onboarding record -> initialize 7-stage deceleration curve -> generate stage6 image)
  • Updated STAGE_TITLES/STAGE_PROGRESS to the deceleration curve from docs/evolution-image-strategy.md sec 2.1
  • generateNextStagePreview now passes the start photo as a reference
  • Records multi-model body-fat detail (breakdown / kept count / spread)

• backend/src/ai/

  • Split body-fat estimation into a multi-model ensemble (body-fat-ensemble.service.ts + body-fat-aggregator.ts)
  • ai.service.ts.estimateBodyFatFromImage now returns { value, aggregate } instead of a raw number — required by the evolution-stage updates above
  • Adds aggregator unit spec (body-fat-aggregator.spec.ts)
  • ai.module.ts wires in the ensemble service

• backend/prisma/schema.prisma

  • EvolutionAssessment += { multiModelDetail, modelCount, modelSpread } so the ensemble breakdown is persisted alongside each assessment

• docs/evolution-image-strategy.md (NEW)

  • Single source of truth for the body-fat -> image-generation system
  • Frozen prompt templates (A/B/D), variable dictionary, fallback chain, and a file:line code-integration map for implementers