track(design): run the KAMI image render + auto-review pipeline (nolte-media) — generate assets from #593 prompts

[nolte]

Context

Follow-on to #593 (KAMI prompt docs, resolved by #598). Branch feat/kami-media-pipeline adds a machine-driven generate + Claude-vision conformance-review loop targeting the nolte-media plugin. The pipeline is built and dry-run-validated but not yet executed (no Cloudflare credentials in the authoring env), and the assets are not yet rendered/integrated.

This issue tracks the remaining work so it is not forgotten.

Delivered on feat/kami-media-pipeline (commit 485750762)

• spec/design/_generation-manifest.yaml — machine SSOT, 79 render jobs (Batch 1: 20 inline FLUX prompts for the 8 new docs(design): derive KAMI image-generation prompts from the illustration audit (#591) #593 docs; Batch 2: 59 series jobs — features G-06, tank G-05, nav-icons G-04 — via doc:+motif_heading: references).
• scripts/kami/render.py — driver over nolte-media image-generate (FLUX.1-schnell via Cloudflare), per-job state machine, fresh-seed regen, flux_normalize.
• .claude/agents/kami-image-reviewer.md — vision reviewer vs spec/design/KAMI-CHARACTER-REFERENCE.md (§3/§3.2/§4.2/§5/§6/§8).
• scripts/kami/README.md runbook; task kami:status|generate|worklist.

Model note (nolte-media)

Default provider cloudflare = FLUX.1-schnell (free, ignores width/height → always ~1024² square). gemini-2.5-flash-image requires billing + SynthID watermark. FLUX has no negative-prompt channel → the docs' Avoid: clauses are stripped by flux_normalize.

Open tasks

• Open a PR for feat/kami-media-pipeline → develop.
• Provide CLOUDFLARE_API_TOKEN + CLOUDFLARE_ACCOUNT_ID (Workers AI scope) in the run env.
• Run the generate → vision-review → regen loop per scripts/kami/README.md (batch by priority: G-01/G-02 app-icons + G-03 onboarding first).
• Add per-icon dark nav-icon jobs (currently light-only; same mechanic, add variant: dark entries).
• Downstream per approved PNG (out of pipeline scope): crop to target size → nolte-media:png-to-transparent-svg (transparent motifs; app-icons stay opaque) → place under each job's out: path.
• Wire consumers: barrel assets/brand/illustrations/index.ts, Sidebar.tsx (nav-icons), public/manifest.json + public/icons/ (fixes the broken PWA icon 404), KAMI_PHASE_IMAGES (post-harvest), tank fill-level mapping, EmptyState.tsx default.
• Verify Light + Dark per motif; reconcile spec/design/_index.md status column to Prompt ✅ / Asset ✅.

References

• Branch: feat/kami-media-pipeline (worktree, not yet pushed)
• Prompts source: docs(design): derive KAMI image-generation prompts from the illustration audit (#591) #593 / docs(design): derive KAMI image-generation prompts from the audit (#593) #598, spec/design/*.md, spec/design/_index.md
• Norm: spec/design/KAMI-CHARACTER-REFERENCE.md
• Audit: spec/analysis/kami-illustration-audit-2026-07.md (G-01…G-17)

[nolte]

Orchestration status — KAMI illustrations aligned to the face-on-pot brand mark

Classification: docs / design. Draft PR: #686 (branch feat/kami-assets-618).

Root cause fixed

The generated KAMI illustrations put the face on the plant instead of the pot — off-model versus the shipped brand/logo@2x.png / icon@2x.png. The character reference §2 anatomy had drifted (a separate "head area" between the leaves and the pot). Corrected: face + arms sit on the pot body, leaves are plain foliage.

Delivered in #686

• Reference §2 anatomy fix + tightened kami-image-reviewer K1 (reject face-on-plant).
• Shared defaults.anatomy_anchor prepended to every render prompt by render.py — one data edit steers the whole set (face-on-pot, muted-terracotta striped pot, no blush). Plus a MAX_PROMPT_CHARS guard (FLUX rejects >2048 chars with HTTP 400).
• PWA icons sourced from brand/icon@2x.png → public/icons/icon-{192,512}.png (opaque, maskable safe-zone). Fixes the manifest.json /icons/* 404. Logo/app-icons are no longer FLUX jobs.
• Manifest 79 → 18 scene jobs: illustrations made light-only (24 duplicate *-dark jobs dropped); nav-icons (illegible at 24px → keep MUI icons) and tank fill-levels (numeric → FLUX renders literal "40%/60%" text → better as a parametric SVG) dropped.
• 15 reviewed, background-removed transparent-PNG illustrations committed (states, onboarding, dashboard-empty, phases, 9 features).

Pending — regeneration blocked on daily Cloudflare FLUX quota (HTTP 429)

Three renders need a re-roll when the free-tier window resets (tomorrow):

• dashboard-welcome — a second face appeared on the plant (double face)
• phase-curing — disjointed face (eyes near the rim)
• feature-glossar — face-on-plant regression

Follow-up (separate PR)

Barrel integration (illustrations/index.ts named exports) + consumer wiring + optional vtracer-SVG conversion + retiring the old face-on-plant SVGs. This tracking issue stays open until that lands.

Generalist-handled: the work is visual-iteration-bound (prompt tuning + per-image conformance review require vision), which the text-only design specialists can't perform.