[christso]

Summary

Add bun run eval:humanizer to compare the actual Humanizer skill with write-naturally on the same public cases, writer settings and graders.

The extension fetches a pinned skill revision and its MIT license, verifies both hashes, and prepares a separate temporary workspace. The runner changes the skill-loading instruction and checks while retaining the common rewrite instructions and grading criteria. It also supports caller-supplied suites without storing their fixtures in this repository. Documentation covers focused runs, provenance, host downloads and cleanup limits.

Verification

• Catalogue and local links validated; all 21 existing unit tests passed.
• Generated public and companion private configurations validated without model calls.
• Temporary CLI fixtures preserved all cases and default/case-specific graders, checked pinned downloads and skill-read assertions, and verified cleanup and exit status on success and failure.
• HTTP, transport and checksum failure fixtures stopped setup and cleaned the workspace.
• Independent review found no remaining blockers.

No comparative model evals were run. These changes add the comparison setup; they do not establish which skill produces better writing.