feat: add microfrontends skill

[tim123abc]

Adds the microfrontends skill for building, configuring, and deploying microfrontends on Vercel.

Contents

• skills/microfrontends/SKILL.md + 6 reference docs (configuration, path-routing, local-development, managing, security, troubleshooting)
• Regenerated generated/skill-manifest.json and generated/skill-catalog.md (27 skills)
• Updated vercel.md ecosystem graph

Triggering

• pathPatterns: microfrontends.json, apps/*/microfrontends.json
• bashPatterns: vercel microfrontends / vercel mf, and npm/pnpm/bun/yarn install of @vercel/microfrontends
• importPatterns: @vercel/microfrontends

Note

This rebases the work from #35 onto current main as a single signed commit, to satisfy the repository's "commits must have verified signatures" rule (the original branch carried main's pre-rule unsigned history, which the rule re-flagged). Supersedes #35.

🤖 Generated with Claude Code

---

Design rationale (re: triggering & references)

How does the skill get triggered? (and is it over-eager?)

Deliberately conservative on the automatic (hook) side:

• No promptSignals → never injected by exact prompt phrase-matching.
• Hook injection fires only on concrete signals you'd hit when you mean it:
  • editing microfrontends.json / apps/*/microfrontends.json (pathPatterns)
  • running vercel microfrontends / vercel mf (bashPatterns)
  • installing or importing @vercel/microfrontends (bashPatterns + importPatterns) — added in this PR
• The only fuzzy prompt path is the lexical fallback scored against the retrieval block (aliases/intents/entities), gated by a min-score threshold, a hard floor, dominant-topic suppression, and a 2-skills / 8KB cap.

So it won't trigger constantly. Note this PR intentionally broadens triggering vs. the original by adding import/install matching (the most common real entry points, matching every other package-centric skill, e.g. vercel-sandbox, vercel-storage, ai-sdk) — still precise signals, not loose ones. The softest surface is the long description text (what Claude reads to self-invoke); that's the lever to trim if it ever feels over-eager.

Why a references/ directory instead of one flat file?

• SKILL.md is a 32-line router: core concepts + CLI surface + a topic→file table with an explicit "do not read all references upfront — grep the relevant one."
• The 6 reference docs total ~1,058 lines (trimmed from ~1,122 — see the boilerplate-trim commit).
• A flat ~1,150-line skill would blow the 8KB UserPromptSubmit budget and crowd the 18KB PreToolUse budget, getting truncated to its summary. Progressive disclosure keeps the always-injected payload tiny and pulls depth on demand.
• A flat file is only preferable when a skill fits budget (~150–200 lines) with no natural topic split. This skill is well past that, so the split matches the nextjs / vercel-cli pattern.

[Melkeydev]

approved - just bump the versions for all the providers please