fix: normalize list score dicts in remote evals

[Everett Smith (evrtt)]

Context

Valid multi-score remote scorer results from remote eval scorers are rejected. Remote eval scorers can return list[Score], but that result crosses an HTTP/JSON boundary and the caller receives list[dict] rather than real Score objects. await_or_run_scorer normalizes a single score dict with Score.from_dict(), but is missing normalization for list[dict] responses.

Changes

• Normalize dict entries inside scorer result lists with Score.from_dict().
• Keep existing validation for malformed score entries.
• Add regression coverage for a scorer returning a list of score dicts.
• Verify both per-result scores and summary scores are populated correctly.

Testing

• Added test_run_evaluator_normalizes_list_of_dict_scores

  • creates an evaluator with two data rows
  • uses a scorer that returns a list of two score dicts:
    • {"name": "per_result", "score": 1.0}
    • "name": "summary_only", "score": 1.0}
  • runs the evaluator locally with run_evaluator(...)
  • asserts each eval result contains both scores
  • asserts the final summary contains both score names with score 1.0

• Ran the focused core regression test

• Ran a local smoke test where a scorer returns multiple score dicts, confirming the eval completes and emits both expected scores.

[Abhijeet Prasad (AbhiPrasad)]

does this issue effect the other sdks?

[Everett Smith (evrtt)]

Abhijeet Prasad (@AbhiPrasad) yeah Go and Ruby both similarly fail with a list of scores currently