fix(registry): extract skill content from ZIP archive

[mgoldsborough]

Summary

• Skill .skill files are ZIP archives containing SKILL.md, but the announce endpoint was reading the ZIP binary as raw UTF-8 text and regex-matching for YAML frontmatter — which always failed silently, storing content as null
• Skill detail pages (e.g., /skills/@nimblebraininc/upjack-app-builder) only showed tags/metadata with no skill body content
• Now uses AdmZip (already a dependency) to extract SKILL.md from the archive before parsing frontmatter

Test plan

• Verify pnpm typecheck and pnpm test pass (65/65 tests passing locally)
• Re-announce an existing skill (e.g., upjack-app-builder) and verify the API returns non-null content
• Confirm the skill detail page renders the markdown body on the Overview tab

[JoeCardoso13]

Clean fix — correct root cause, minimal diff, good test coverage.

Minor nitpicks (non-blocking):

1. Observability — When extractSkillContent() returns undefined, the route handler proceeds silently. A log.warn would help debug content extraction failures in production.
2. Robustness — The frontmatter regex ([\s\S]*? non-greedy) could mismatch if YAML content contains a literal \n---\n line. Unlikely in practice; same regex as before, so not a regression.
3. Test coverage — No test for deeply nested paths (e.g., a/b/c/SKILL.md). The endsWith logic handles it, but it's undocumented by tests.
4. Error handling — The try/catch in extractSkillContent only wraps the AdmZip constructor, but getData() can also throw. Consider widening it to cover the full function body so the string | undefined return contract holds. The route handler's outer try/catch prevents any real damage.