Consider adding claude-code integration tests

[ViktorHofer]

To find issues like #416 immediately.

[ViktorHofer]

Skill-validator's eval command relies heavily on CopilotClient (github.copilot.sdk) but we don't necessarily need to test that part with claude code. I would probably just add basic E2E tests that make sure that the marketplace+plugins are installable on claude code and have maybe a santiy check that an agent or skill can be triggered (either from context or directly).