Add Simplified Chinese support and a mainland-China capability

[alan8918]

The interview already ran in the user's language. Three things broke silently underneath it: the skill never activated on Chinese phrasing, the validator's English-only regexes reported zero vague criteria over a document they could not read, and a translated status value made /architect-next report a finished build on turn one. knowledge/localization.md is the binding contract for the boundary; finding 38 and Sweep 28 enforce it.

knowledge/capabilities/china-mainland.md is the 19th capability: the inside-versus-outside-the-border fork, the compliance gates as calendar time, cross-border PI thresholds, phone-first identity and the unionid trap, WeChat Pay and Alipay mechanics, the dependencies that only resolve outside the border, GCJ-02, and pre-publish moderation. Its filings are 20.1 manual gates, never 9 build steps -- every one of them completes when a regulator acts, which is exactly what finding 17 forbids as a build gate.

Co-Authored-By: Claude Opus 5 noreply@anthropic.com

Claude-Session: https://claude.ai/code/session_01SsGYobLpdbKj3hAvJBEi3V

Read this first

• This PR does NOT contain a generated blueprint or a project built from one.

This is the single most common mistake here. The Architect is the tool that writes blueprints —
it is not a place to store them. v2 writes to ./blueprints/ in your working directory, and
that output belongs in your project's repo, not this one.

• Want to show off a blueprint? → Discussions › Show and tell
• Blueprint came out wrong? → open a blueprint quality issue

PRs consisting of generated output are closed without review.

---

What this changes

Why

---

Type of change

• New shape — a file in knowledge/shapes/ (needs an accepted new shape issue first)
• Capability — new or edited file in knowledge/capabilities/
• Runtime track refresh — version pins in knowledge/runtime-tracks/
• Skills registry — an entry in knowledge/skills-registry.md
• Plugin surface — commands/, agents/, skills/, .claude-plugin/
• Interview or templates — questions/, templates/
• Docs — README.md, CLAUDE.md, .github/
• Fix — broken cross-reference, CI, typo

---

Checklist

Each item is a rule CI or review will check anyway. The reasoning behind them is in
CONTRIBUTING.md.

Always

• Version pins appear only in knowledge/runtime-tracks/. No library version anywhere in
  knowledge/shapes/ or knowledge/capabilities/ — link to the track instead. Standards
  (WCAG 2.2, Manifest V3, OAuth 2.1, HTTP status codes) are not version pins and are fine.
• Last verified: date updated on every knowledge file I touched.
• Every cross-reference I wrote points at a file that actually exists, at its exact path.
• Every build step I wrote has an observable Done when a script could check today, on this
  machine, without a human or an external approval queue.
• No removed skill names: /deep-research, /seo-audit, /pdf-design, /shadcn-ui,
  /chrome-bridge-automation, /web-reader, /humanizer. And no slash prefix on
  auto-activating skills (ui-ux-pro-max, frontend-design, playwright-cli, find-skills).
• Knowledge files are in English, opinionated, table-first, and emoji-free.
• CI is green (.github/workflows/validate.yml).

If this adds or edits a shape

• Classification reciprocity holds. Every shape my ## Is this your project? → No if
  list points at now points back at mine. I edited those files too — reciprocity is
  bidirectional and one-way edges are the reason briefs never reach a shape.
• saas-webapp.md still offers an exit to all 13 other shapes; every other shape has ≥ 3.
• Names a default runtime track and lists its core capabilities by real file path.
• Ends with ## See also linking at least two real manifest paths.

If this touches version pins (knowledge/runtime-tracks/)

• Every pin was read from the live registry or the published package, not from memory and
  not from a search result. Registry beats search; the package's own types beat both.
• The Last verified line at the top of the track states today's date.
• Anything I could not verify says so in the file rather than guessing.

If this adds a skill to knowledge/skills-registry.md

• The install command is verified — I ran it, or read it from the skill's own README today.
• The invocation form is correct: a leading / only if it is genuinely a slash command.
  No slash means it auto-activates, and writing it with a slash is a silent no-op.
• Repo, license, and star count are real and stated.
• The reference degrades gracefully — if the skill is absent, the workflow falls back to the
  knowledge base or built-in WebSearch/WebFetch and keeps going.

---

How I verified this