feat: honest image backends, a visual QA loop, and Iris as persona #17

[d6veteran]

Implements proposals 1, 2, and 5 from docs/proposals/image-generation.md, then formally closes out 3 and 4 based on what the implementation revealed. The proposal doc is included as the first commit and its status header records every decision.

Why

profiles/kai.md claimed image tooling the product never installs: Hugging Face MCP dynamic_space with FLUX.1-Krea-dev, Qwen-Image, and FLUX.1-Kontext-Dev, plus four Figma MCP tools. install.sh copies profiles, agents, and commands and never installs or checks an MCP server. No Claude model generates raster images either, so Kai's real ceiling was markup it writes itself. The persona described tools that were not there and then improvised.

What shipped

Proposal 1, honest capability boundary (persona behavior only; the claude-team doctor command was declined). Domain Expertise is backend agnostic, and a new "Declare the Backend" behavior requires naming the backend before promising an image, or saying none is connected and offering what is actually deliverable.

Proposal 2, visual QA loop. Open the artifact and inspect it before showing it: score against named criteria, name the single worst failure, revise one thing, up to three passes. Claude is multimodal on input, so this needs no vendor.

Proposal 5, split the lane. Iris — Brand & Illustration (#17, claude-opus-4-8) owns logos, wordmarks, icon sets, illustration, marketing graphics, and asset licensing. Kai keeps screens, layout, and design systems, and specifies the assets a screen needs rather than generating them. The QA loop split along its seam rather than being deleted: Kai kept the render-and-screenshot half, Iris took the generated-asset half. Both personas ended at four behaviors.

What was dropped, and why

Proposal 3 (a checked-in BRAND.md + assets/MANIFEST.md + claude-team brand init) is dropped. Two of its justifications did not survive checking:

• It claims brand init would mirror branch start and session start. Those write exclusively to ~/.claude/. This CLI has never written a committable file into a user's working tree; its only project writes are .git/hooks/pre-commit and .git/info/exclude, both untracked by design.
• It claims to unblock the repo's "missing" publish/style-guide.md. That file was deliberately deleted in 04ace9e ("consolidated to code-katz/style-guide.md"), so re-adding a per-repo brand file re-introduces what was removed on purpose.

Half of it had also already shipped inside Iris's Brand Brief and Asset Provenance Record, and "persona names an artifact the repo does not define" is the established convention here (Kai and Sasha have pointed at DesignSystem.swift with no schema since they were written). Accepted tradeoff, stated in the doc: Iris keeps improvising a manifest shape per project.

Proposal 4 (wire a backend) is folded into Proposal 6 (companion skill) rather than shipped here. They were decomposed as two items and should not have been. A vendor MCP matrix in this repo would be documentation the project does not control, drifting on someone else's release schedule, and it breaks the convention that every companion capability (devlog, roadmap, plans, todo, publish, conductor) lives in its own repo while "Works Well With" stays one line and a link. Backend mechanics also serve Reiner, Ernie, Toni, and Kai, not only Iris.

Prior art for that skill is recorded in the doc. The load-bearing finding: neonwatty/logo-designer-skill generates SVG concepts with no image backend at all, which means the vector-mark problem needs no vendor. Iris now routes by asset class accordingly, treating hand-authored SVG as the primary path for logos and icons rather than a consolation prize, with a connected backend needed only for raster illustration.

Test hardening

CI ran shellcheck and the suite on two platforms, but never ran generate-agents.sh and never compared agents/ to profiles/ — the count check only counted files. A profile edited without regenerating passed green while the subagent served stale text. The same hole existed for hand-maintained commands/.

Both are now checked per persona. Verified they fail on a deliberately unregenerated profile (stale: robin(1), drift: robin(1)) and pass once restored. Also adds the missing claude-team list assertion for Sage.

Verification

• bash tests/run.sh — 128/128, up from 123
• shellcheck on all five CI-linted files — clean
• grep -rn "Recraft\|Gemini\|Hugging Face\|Canva\|FLUX" profiles/ commands/ agents/ — no vendor names in persona content
• 17/17/17 across profiles/, commands/, and agents/

Notes for review

• No "Works Well With" row was added for the image skill: that repo does not exist yet, and linking a nonexistent tool is the same class of bug this PR set out to fix.
• Prior-art licenses could not be verified in-session (api.github.com 403s through the proxy). The doc flags them as look-don't-copy until checked.
• Left as history: README.md:853 (v0.5 release notes) and the DEVLOG.
• Already stale before this work and untouched: gtm.md:416, commands/silicon-valley.md.