fix(judge): collapse judge_llm.py onto judge.py's transport

[vaibhavdabas16]

What does this PR do?

judge_llm.py declares itself a "drop-in replacement" for judge.py (:20-22) and then re-implements _post_json, _build_user_msg, and the api_type dispatch. That copy has drifted, and it is the module behind the published Reward-lenient column:

	judge.py	judge_llm.py (before)
judge_context kwarg (passed at run.py:592)	yes	absent from the signature — a TypeError on this path
openai-responses	/responses (:240)	folded in with openai-completions → /chat/completions (:188)
google-generative-ai	supported via /v1beta/openai (:235-239)	NotImplementedError (:195)
max_tokens	4096	800

The judge_context gap is the one with the quietest failure mode: imported rubric-based tasks were judged on the lenient path with the rubric withheld from the judge. The 800-token cap matters because the default judge is deepseek-v4-pro, a reasoning model that spends hidden tokens before emitting its JSON line — truncation lands on the unparseable path.

This reduces the module to its JUDGE_SYSTEM prompt, which is the one thing that legitimately differs between the rubrics, and delegates to judge._build_user_msg / judge._run_judge. 176 lines of duplicated transport become 28.

judge_request keeps its positional signature, so rescore.py's judge-function dict (rescore.py:261-269, which calls both judges positionally) is unchanged. It gains judge_context and retries as keyword-only arguments matching the strict judge, and the returned dict keeps its "rubric": "lenient" key.

Two behaviour changes fall out of the collapse and are intended — flagging them explicitly rather than burying them:

1. Retries now cover any transport exception, not only HTTP 429/500/502/503.
2. A failed call reports judge_call_failed: ... as its reason instead of an empty string.

Both are judge.py's existing behaviour; this PR adopts it rather than preserving the divergent copy. Happy to keep the old retry policy instead if you would rather that stayed put.

Corpus

• v2
• v1
• both
• not applicable

The lenient rubric is the V2 reward signal (judge_llm.py:15-18), though the module is reachable from clawbench-rescore for any corpus.

Test plan

• New tests/test_lenient_judge_parity.py (10 tests) pins each of the four divergences, asserts the two rubrics differ only in the system prompt on the wire (same URL, headers, user message, and token budget), and guards against the transport being copied back in.
• The new tests were run against the previous judge_llm.py: 9 of the 10 fail, each with the error the issue describes — TypeError: unexpected keyword argument 'judge_context', the openai-responses call never reaching /responses, assert None is True for Gemini, and max_tokens 800. A parity test that only passes after the change proves nothing, so this control mattered.
• Verdict parsing was checked for equivalence before collapsing it: both _parse_verdict implementations were run over the union of their test corpora plus markdown-fence, prefix-prose, and stringly-typed edge cases (15 cases) and agreed on every one. So _parse_verdict is re-exported from judge.py, no published number moves, and the existing test_lenient_judge_parse_verdict_is_tri_state keeps guarding this path unchanged.
• Full suite: 209 passed, 3 skipped. The single failure, test_host_tasks.py::test_checked_task_json_files_parse_and_validate[v1-lite], is a local Windows artifact (those task files are git symlinks, mode 120000, that Windows checks out as text) and is green on CI.
• ruff check and ruff format --check clean on both files.
• Verified conflict-free against main and against all four of my other open PRs (fix(batch): one truncated run-meta.json no longer discards the batch summary #312, fix(rescore): resolve paths from the workspace instead of a maintainer's home #313, fix(judge): a bad --judge model no longer discards a completed run #314, fix(runner): bound container wait so one wedged run can't stall a batch #316) with git merge-tree.

No network is touched by the tests: they monkeypatch judge._post_json and assert on the captured payloads.

Related issues

Fixes #300.