Make parseGraderSelection permissive; keep validateEvalYAML strict for unknown-field detection

[Copilot]

parseGraderSelection used KnownFields(true) when decoding the pass-1 LLM response, causing it to silently drop valid grader selections whenever the LLM included extra keys (e.g. reason:, explanation:) alongside graders:. Unknown-field validation belongs in validateEvalYAML, not here.

Changes

• parseGraderSelection — removed KnownFields(true) from both structured and bare-list decoders. Extra keys in LLM output no longer prevent grader-type extraction.
• validateEvalYAML — unchanged; retains KnownFields(true) as the authoritative place to catch unknown/invalid fields in the final generated eval YAML.
• New test — TestParseGraderSelectionExtraFields covers the previously-broken case:

# LLM response with extra field — was silently dropping "code" and "text"
graders:
  - code
  - text
reason: these graders cover the task

---

🔒 GitHub Advanced Security automatically protects Copilot coding agent pull requests. You can protect all pull requests by enabling Advanced Security for your repositories. Learn more about Advanced Security.