feat(skills): add HomeMux design system skill

[kunchenguid]

Vendors the HomeMux design system as a Claude-visible skill under .agents/skills/homemux-design (the .claude/skills symlink already exposes it). It drives the holistic HomeMux design pass. This is vendored design reference content (tokens, components, guidelines, the iOS UI-kit mock, brand assets), not firstmate-authored code, so it did not go through the no-mistakes code pipeline.

[juniorlovestmh]

Evidence disclosure: The accepted intent is live proof of the observed error class plus dedupe, and live pipeline 41 first-run emission plus second-run silence satisfies it; pipeline 40 scrolled out of the bounded last-10 window because new pipelines were created since - environmental drift, not a code gap - and the mocked tests already prove distinct multi-pipeline wakes.

[x45dev]

Correcting my own comment above: the first reason I gave does not hold.

I read the policy as forbidding an autonomous merge on a change to the constitution's methodology. It does not. AGENTS.md:50 on origin/main reads "knowledge/constitution.md's invariants or to this file's methodology or merge policy" - the methodology and merge-policy clause attaches to AGENTS.md itself, not to the constitution. Ticking Definition of Done boxes is not an invariant change, and the invariants are confirmed untouched, so that trigger never fired.

The hold still stands, on the two grounds that do hold. The substantive one is unchanged and is the real reason: the deferred naming set explicitly names an invariant inside knowledge/constitution.md as carrying a named institution, so that file genuinely sits inside the question already in front of the author. This branch touches a different section of it, which is why the change itself reads as safe - but it is his file to rule on this week. Under the policy, what formally earns the hold is the last high condition, a named reason for wanting a second opinion, and that is what I am recording it as.

Nothing about the verification above changes, and my recommendation is unchanged: merge.