fix: stop rejecting skills whose SKILL.md uses thematic breaks - #3297

What Problem This Solves

Fixes an issue where publishing a skill fails with

Skill content is too thin or templated. Add meaningful, specific documentation.

even though the SKILL.md is fully documented. The trigger is a SKILL.md that has
no YAML frontmatter and uses --- as an ordinary Markdown thematic break. The
publish is rejected outright, so the skill never reaches the catalog.

Newly created accounts are hit hardest: the reject floors are highest for the
low trust tier, which covers every account under 30 days old or with fewer
than 10 published skills.

Why This Change Was Made

The quality gate removes frontmatter before measuring the body:

raw.replace(/^---\s*\n[\s\S]*?\n---\s*\n?/m, "")

The m flag makes ^ match at every line start rather than only at the start of
the document. Frontmatter is optional on publish — publishSkillVersion takes the
display name from its own arguments, and parseFrontmatter returns an empty
record for a document that has none — so a SKILL.md may legitimately open with a
heading and use --- as a horizontal rule. When it does, the pattern latches onto
the first rule and deletes everything up to the second one.

Dropping the flag anchors the pattern to the start of the document, which is what
the rest of the repository already does. The canonical parser
parseFrontmatter in convex/lib/skills/index.ts guards with
normalized.startsWith("---"), and the three other frontmatter patterns
(convex/skillsShMirror.ts, packages/schema/src/clawPackage.ts,
packages/clawhub/src/schema/clawPackage.ts) carry no m flag. skillQuality
was the only one reading the same document differently from the publish path that
calls it.

The truncated body also fed toStructuralFingerprint, so template-spam
similarity was being compared over a fragment rather than the real document. That
is corrected by the same change.

Non-goals, kept out to hold the change to one concern:

• Tuning the reject or quarantine thresholds. They are unchanged.
• Sharing a single frontmatter helper across the four call sites.

User Impact

A skill whose SKILL.md uses --- as a thematic break now publishes normally
instead of being rejected as thin content. Spam detection also compares whole
documents rather than truncated fragments, so the fingerprint is no longer
skewed toward whatever a skill happens to place above its first horizontal rule.

Evidence

Base commit: 79ef4af1. The fix is not on main — stripFrontmatter there still
carries the m flag.

Signals and the gate decision read from computeQualitySignals and
evaluateQuality, for a 109-word SKILL.md with two headings, four bullets and two
--- thematic breaks, evaluated at the low trust tier with no similar recent
publishes:

94% of the body was being discarded before measurement.

This module had no unit test. convex/lib/skillQuality.test.ts adds three cases;
two of them fail on the parent commit and pass with the fix:

$ VITE_CONVEX_URL=https://example.invalid bunx vitest run \
    convex/lib/skillQuality.test.ts        # parent commit 79ef4af1

 × computeQualitySignals > keeps the body of a frontmatter-less SKILL.md that uses thematic breaks
AssertionError: expected 1 to be 3 // Object.is equality

 × evaluateQuality > does not reject a documented frontmatter-less skill from a new account
AssertionError: expected 'reject' to be 'pass' // Object.is equality

 Test Files  1 failed (1)
      Tests  2 failed | 1 passed (3)

$ VITE_CONVEX_URL=https://example.invalid bunx vitest run \
    convex/lib/skillQuality.test.ts        # with the fix

 Test Files  1 passed (1)
      Tests  3 passed (3)

The third case (still strips real frontmatter) passes on both sides: it compares
a document carrying frontmatter against the same body without it and asserts the
measured bodyWords and bodyChars match, so the intended stripping is still
happening.

Consumers of the module, run together:

$ bunx vitest run convex/lib/skillPublish.test.ts convex/lib/skillQuality.test.ts \
    convex/lib/skills.test.ts
 Test Files  3 passed (3)
      Tests  61 passed (61)

Gates run locally on Windows:

• bunx tsc --noEmit — clean
• bun run lint — clean
• bun run deadcode:ci — clean
• bun run format:check — the only two files it reports are CLAUDE.md and
  .agents/skills/autoreview/CLAUDE.md; both report identically on an unmodified
  checkout of 79ef4af1, so they are pre-existing and untouched here.
• bun run ci:unit — 5542 passed. The 24 failures are pre-existing on this
  platform: an unmodified checkout of 79ef4af1 fails the same 14 files with the
  same 24 tests (5539 passed there; the difference is the three tests added by
  this PR). They are the scripts/ worker, CLI and security-dataset suites that
  shell out to bun, plus convex/lib/githubAccount.test.ts and
  src/routes/-management.test.tsx. Linux CI is the authoritative signal.

The Vercel preview check will need OpenClaw Foundation team authorization, as with
other fork pull requests.