judge_llm.py has diverged from judge.py (no Gemini support, wrong responses endpoint, 800-token cap) — collapse the duplication

src/clawbench/runner/judge_llm.py:20-22 says it is a "drop-in replacement" for judge.py with the same interface. It has diverged in ways that matter, and it is the module behind the lenient rubric that produces the published Reward-lenient column:

The default judge is deepseek-v4-pro, a reasoning model — an 800-token budget risks truncating its verdict, and truncation lands on the except path whose lenient default is match=True (see #295). On top of that, _post_json / _build_user_msg / _call_* are ~100 duplicated lines that now have to be fixed twice; the Gemini fix already only landed in one copy.

Ask: reduce judge_llm.py to its JUDGE_SYSTEM prompt and delegate to judge.py's _run_judge / _build_user_msg with the rubric as a parameter. That restores identical api_type coverage, token budget, and signature in one place.

Start with src/clawbench/runner/judge_llm.py and compare its wrapper with judge.py, especially _run_judge and _build_user_msg and the call sites in run.py:587-597. Reduce the duplicate implementation while preserving the rubric parameter; done means both judge paths share the same interface, API endpoint coverage, Gemini support, and 4096-token budget.