Fix Luna LM Checklist request formatting

[martinsrui-msft]

Summary

• convert assistant LM Checklist history to Responses-compatible output_text content for GPT-5 routes
• omit Luna's unsupported temperature argument without changing other models
• cover Luna, non-Luna GPT-5 Responses, and legacy Chat Completions behavior

Context

main already selects gpt-56-reasoning-nano-luna through #946. In run 37611607449, all 110 LM Checklist calls failed before scoring because autoevals==0.3.0 converted assistant history to input_text; CAPI accepts assistant history only as output_text or refusal on the Responses API.

This PR adds a narrow request-side adapter around the built-in bc-eval LM Checklist scorer. It does not change checklist assertions, grading semantics, score names, retries, or timeouts. #948 addresses the separate response-side issue where nested Responses output is otherwise converted to content=None.

The generic role-aware conversion should ultimately live in autoevals, or in bc-eval as the integration boundary if that dependency cannot be updated immediately. This local adapter keeps the currently pinned bc-eval==0.3.14 workflow operational and can be removed after the upstream integration is fixed and upgraded.

Validation

• 3 focused LM Checklist request-formatting tests passed
• Ruff check and format checks passed
• targeted Ty checks passed
• relevant pre-commit hooks passed

[github-code-quality]

Code Coverage Overview

Languages: Python

Python / code-coverage/pytest

The overall line coverage in commit 7df241b in the martinsrui-lm-checkl... branch remains at 86%, unchanged from commit d19989f in the main branch.

---

_{Updated October 07, 2026 14:03 UTC}

[Sun Haoran (haoranpb)]

How about something like #946 ?

I see something is failing in pre-commit, will make a separate fix

[Sun Haoran (haoranpb)]

You can probably close this and rebase #944 to main