feat(scorers): add NamedScorer wrapper to support custom scorer names

[nmath-goog]

Introduces the NamedScorer wrapper class to support custom top-level YAML keys in run configurations (e.g. rubric_pass_fail: {type: python_scorer}).

Summary of Changes

1. NamedScorer Wrapper (evalbench/scorers/namedscorer.py):

• A lightweight Comparator wrapper that stores the custom metric name (name) and delegates compare(...) to its base_scorer.

2. Resolution Logic in score.py (evalbench/scorers/score.py):

• get_scorer_instance checks if a YAML key exists in DEFAULT_SCORERS. If not, it inspects the type: attribute (e.g., rubric_pass_fail: {type: python_scorer, script_path: ...}) or nested dictionary (rubric_validator: {python_scorer: ...}) and wraps the underlying comparator in NamedScorer.

3. Summary Reporting in analyzer.py (evalbench/reporting/analyzer.py):

• Dynamically includes custom named comparators from df["comparator"] in summary reports.

Test Plan & Verification

1. NamedScorer Unit Tests (evalbench/test/named_python_scorer_test.py):

• test_type_python_scorer_and_nested_dict: Tests configuration resolution for custom top-level keys using both type: python_scorer attribute syntax and nested dictionary syntax.
• test_named_scorer_wrapper: Verifies delegation of compare() calls and metric name property forwarding.

2. Resolution & Execution Tests in score_test.py (evalbench/test/score_test.py):

• test_get_scorer_instance_named_scorer_type_attr: Asserts get_scorer_instance returns a NamedScorer wrapping PythonScorer for top-level keys specifying type: python_scorer.
• test_get_scorer_instance_named_scorer_nested_dict: Asserts get_scorer_instance returns a NamedScorer for nested dictionary declarations.
• test_score_compare_execution: Asserts score.compare(...) executes named comparators and correctly sets comparator = custom_name in output results.

3. Full Suite Execution:

• All test cases passed.

[prernakakkar-google]

/gcbrun

[nmath-goog]

/gcbrun

[graceqi-g]

/gcbrun