feat(scorers): support multiple python_scorer_* entries in run YAML

[saurabh-net]

Motivation

In evaluation workflows, users often need to evaluate model outputs against multiple distinct Python evaluators (e.g. execution accuracy, code style, compliance, or custom domain judges) within a single evaluation run configuration YAML.

Previously:

1. score.py only checked for exact key if key == "python_scorer":.
2. Standard YAML parsers overwrite duplicate mapping keys, so repeating python_scorer: in YAML only preserved the last entry.
3. Custom keys like python_scorer_accuracy or python_scorer_style were skipped by score.py.

Implementation Summary

• Key Prefix Matching: Updated evalbench/scorers/score.py to match any dictionary key under scorers: starting with python_scorer (if key.startswith("python_scorer"):).
• Full custom_name Fallback Resolution: Retained full custom_name resolution logic in score.py (scorer_name -> script_path basename -> key fallback).
• Analyzer Resolution Alignment: Updated evalbench/reporting/analyzer.py (analyze_one_metric) to resolve comparator_name using the same fallback rules as score.py. This guarantees that df["comparator"] matches comparator_name cleanly regardless of whether scorer_name, script basename, or key is used for comparator naming.
• Config Protection: Shallow-copied scorer_config before passing to PythonScorer to prevent in-place mutation of input dictionaries.
• Documentation: Updated docs/configs/run-config.md and docs/gemini_cli_agent_testing.md with YAML usage examples for single and multiple Python scorers.

Testing & Verification Context

1. Automated Unit Tests

• Ran pythonscorer_test.py: ./.venv/bin/python -m unittest evalbench/test/pythonscorer_test.py
  • Outcome: Passed 6/6 tests. Added coverage for multi-scorer comparator instantiation, custom_name resolution, and analyzer.py summary metric aggregation.
• Verified hybrid_xa_judge_test.py: ./.venv/bin/python -m unittest evalbench/test/hybrid_xa_judge_test.py
  • Outcome: Passed 5/5 tests. Confirmed evalbench/scorers/judges/hybrid_xa_judge.py and datasets/bird/example_hybrid_run_config.yaml remain 100% backward-compatible.

2. Manual Integration Testing

• Setup: Created two custom Python evaluators (scorer_accuracy.py returning 100.0% and scorer_style.py returning 85.0%) and configured a test YAML containing both under scorers: (python_scorer_accuracy and python_scorer_style).
• CLI Execution: Ran direct end-to-end evaluation using evalbench/evalbench.py CLI:
  python evalbench/evalbench.py --experiment_config=/tmp/evalbench_noop_test_run.yaml
• Result: evalbench.py executed all pipeline stages (DB Setup -> Prompts -> SQLGen -> SQLExec -> Scoring -> Analyzer -> CSV Exports). Both Python scorers ran natively, logged individual scores, and were recorded in scores.csv and summary.csv.