feat(eval): add lightweight skill contract checks

[Scared-Heart]

Summary

This PR adds a first lightweight eval layer for the Skill repository. The goal is to protect important Skill instruction contracts from accidental Markdown regressions, not to introduce full model-based evals yet.

Changes included:

• add a zero-dependency Node.js runner for local contract checks
• add initial cases for teambition and teambition-statistics
• check key instruction contracts such as required env vars, pagination guidance, project inference, priority lookup, statistics tool order, and filterTql rules
• document how to run and extend the checks
• add a GitHub Actions workflow to run the checks on PRs and pushes to main

Scope

This is intentionally a small v0. It only reads local Markdown and JSON files. It does not:

• call Teambition APIs
• require real tokens or production data
• run model-based evals
• verify actual agent tool-call behavior

I think this still provides value as a baseline regression guard, because the current Skills are mostly instruction-driven and important behavior can regress through seemingly small Markdown edits.

Future direction

If this direction looks useful, the next layer could add mocked tool-call evals or optional integration-style evals against a test workspace. Those would be better suited for validating actual agent behavior, while this PR focuses on keeping the existing instruction contracts stable.

Verification

npm test

Result: 2/2 eval cases passed.