Fix Luna LM Checklist response parsing

[martinsrui-msft]

Summary

• fix Luna LM Checklist scoring by extracting nested Responses API assistant output_text before autoevals converts it to content=None
• scope the compatibility client to the exact Luna path while delegating non-Luna requests unchanged
• preserve supported Responses parameters and raise an explicit error when no output text is returned

Failure reproduced

In run 37596204440, job 112734686674, all 110 LM Checklist calls failed when bc-eval 0.3.14/autoevals 0.3.0 converted the nested CAPI response into choices[0].message.content=None. The built-in scorer then called json.loads(None). Kusto's missing test_passed error was downstream because no row received any scorer output.

This PR is the new top layer above #944.

Validation

• 6 focused pytest cases passed
• Ruff check and format passed
• full Ty check passed
• relevant pre-commit hooks passed
• local integration with cached bc-eval 0.3.14/autoevals 0.3.0 reproduced the original content=None behavior and then produced test_passed=1.0 through the real scorer
• real BCal test run 37613146969 passed on all 4 entries, logged zero LM Checklist exceptions, produced the test_passed score, and completed Braintrust/Kusto storage

[github-code-quality]

Code Coverage Overview

Languages: Python

Python / code-coverage/pytest

The overall line coverage in commit 55a8f41 in the martinsrui-fix-luna-... branch remains at 86%, unchanged from commit d19989f in the martinsrui-bcal-resi... branch.

Show a line coverage summary of the most impacted files.

File	martinsrui-bcal-resi... d19989f	martinsrui-fix-luna-... 55a8f41	+/-
src/bcbench/commands/result.py	93%	89%	-4%
src/bcbench/results/display.py	93%	92%	-1%
src/bcbench/results/summary.py	96%	96%	0%
src/bcbench/res...lts/__init__.py	100%	100%	0%
src/bcbench/res...ceval_export.py	100%	100%	0%
src/bcbench/com...ands/dataset.py	25%	26%	+1%
src/bcbench/res...completeness.py	0%	96%	+96%