Add lookup-schema eval test cases

[GiorgioUghini]

Summary

Adds the first eval test suite for the lookup-schema skill — a reference/query skill that looks up Copilot Studio YAML schema definitions using three commands: lookup, search, and resolve.

Since this is a stdout-based skill (no YAML files created), all checks use stdout_contains, stdout_not_contains, and exit_code.

Test Cases

#	Name	Command Tested	Description
1	Lookup known definition — SendActivity	lookup	Looks up a common, well-known kind. Verifies the response includes SendActivity, text, and kind.
2	Lookup trigger kind — OnRecognizedIntent	lookup	Looks up a trigger definition. Verifies the response explains intent and trigger concepts.
3	Lookup top-level kind — AdaptiveDialog	lookup	Looks up the root topic kind. Verifies beginDialog and trigger are described.
4	Search for model-related definitions	search	Searches for model-related schema elements. Checks for specific terms like modelDescription and properties (not just the generic word "model").
5	Resolve $ref references in a definition	resolve	Resolves QuestionWithOptions with its $ref sub-definitions. Verifies fully resolved output includes properties.
6	Lookup non-existent definition — graceful fallback	lookup (negative)	Looks up FakeNonExistentKind. Verifies the skill gracefully reports "not found" without crashes (ENOENT, stack trace, FATAL).

Coverage

• ✅ All 3 skill commands covered (lookup, search, resolve)
• ✅ Positive and negative test cases
• ✅ Uses basic-agent fixture (consistent with other eval suites)

[adilei]

Hey @GiorgioUghini — the evals refactor in #130 changed the eval structure, so this will need a small update before merging:

1. Move evals/skills/lookup-schema.json → evals/scenarios/lookup-schema.json
2. Rename top-level key from skill/evals to scenario_name/evals (see evals/scenarios/agent-settings.json for reference)
3. Rewrite prompts as natural language — e.g. instead of "Use the lookup-schema skill to look up SendActivity", use something like "What properties does the SendActivity kind have in the Copilot Studio YAML schema?"
4. Add routing checks — skill_invoked: "copilot-studio:lookup-schema" and optionally agent_invoked if it should route through a sub-agent
5. Drop any changes to evaluate.py — the harness was rewritten in Refactor evals: scenario-based testing instead of skill isolation #130

The actual check types you're using (stdout_contains, stdout_not_contains, exit_code) are still supported, so the test cases themselves are fine — it's mostly a structural move.