cursor.validateManifest was a no-op; the real structural check only existed
in the dev-only ajv schema test, so a user running `pluginpack validate
--target cursor` got materially weaker protection than every other target's
shipped validator. Adds a name-pattern check and unknown-top-level-key
rejection, mirroring the vendored schema's required/additionalProperties
rules, plus a regression test.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

- Installs the real claude CLI in CI so `claude plugin validate --strict`
  (the strongest documented conformance oracle for the claude target) runs
  on every PR instead of silently skipping.
- Adds Node 22.12.0 to the CI matrix alongside 24, so package.json's stated
  `engines` floor is actually exercised, not just declared. Fixes CLAUDE.md's
  contradictory "Node >= 24" line to match the real floor.
- Adds scripts/verify-pack-contents.mjs and verify-packed-install.mjs: the
  former asserts the npm tarball's file list (not just that `npm pack
  --dry-run` exits zero); the latter installs the real tarball into a
  scratch project and runs the installed binary, catching a broken `bin`
  shebang or `files` misconfig no existing test could see.
- Adds a `labels` CI job requiring one of RELEASE.md's changelog labels on
  every PR.
- Removes the now-redundant `undici` override (release-it now declares it
  directly); keeps the `esbuild` override, which is still load-bearing
  (GHSA-g7r4-m6w7-qqqr requires >=0.28.1, outside tsup's own declared range).
- Adds @vitest/coverage-v8 and a `test:coverage` script (no threshold gate).

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

…I caveats

- README badge: Experimental -> Prerelease. Shipping a 1.0.0 tag under an
  "Experimental" badge (Glean's own stability doc: "not recommended for
  production use") directly contradicted the tag; Prerelease ("feature-
  complete, being validated before GA") matches where this actually stands.
- skills/add-pluginpack-target/SKILL.md described the pre-registry
  architecture (src/targets.ts, src/validate.ts, ~5 touch points) that was
  deleted in the target-registry migration; rewritten to match CLAUDE.md's
  current architecture.
- Added OpenAI Codex CLI to the three skill docs and the README Quick Start
  example that omitted it despite it being a fully supported 5th target.
- Documented in README's Programmatic API section: the target set is closed
  (no third-party target-registration API), the package is deliberately
  ESM-only, and Artifact/ResolvedProject's Map fields don't survive
  JSON.stringify. Added the missing getInstallSnippetCitation/Citation
  entries to the same table.
- Unexported SourceProvider from src/index.ts: nothing in build()/loadConfig()
  accepts a source override yet, so the exported type committed to public
  surface no one could actually use.
- Documented the deliberate choice not to use zod's .strict() on config
  schemas, and what counts as a "breaking change" for a tool whose output is
  files written into a consumer's repo (RELEASE.md).
- Re-verified the vendored Cursor schemas are still byte-identical to
  upstream and recorded that check in SOURCE.md.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

…e, clean, docs

Only `build` and `install-info` were ever exercised as the real built binary
via bintastic; `validate`/`diff`/`prune`/`clean`/`init`/`docs` bypassed
cli.ts's commander wiring entirely, so a broken option parse or wrong exit
code in any of them had no test that could catch it. `init` in particular
had zero coverage at any level, since it isn't a library function.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>