refactor(tasks): migrate 10 benchmarks from llm_judge to the verifiers pipeline

[Luodian]

Summary

Migrates 10 benchmarks from the legacy lmms_eval.llm_judge server API to the new lmms_eval.verifiers pipeline introduced by the stacked feat/verifiers-module PR.

Stacked on #1379. Review against feat/verifiers-module; merge after it lands (this branch then gets rebased/retargeted so the duplicated base commits drop out).

Each task's LLM-judge call is replaced by a lazily-constructed VerificationPipeline (threaded double-checked singleton) built from the package's OpenAIVerifier provider, with per-task extractors/prompts/response-parsers preserved. Scoring behavior is preserved with three deliberate exceptions, listed under Behavior notes below.

Tasks migrated

Task	Verifier	Notes
mmmu	OpenAIVerifier(binary)	reasoning process_results
mmvu	StripReasoning + OpenAIVerifier(binary)	open-ended GPT rescue now actually grades (see Behavior notes)
mathvision	StripReasoning + OpenAIVerifier(binary)	gpt_eval process_results
videoevalpro	StripReasoning + OpenAIVerifier(custom prompt, A/B/C parser)	full judge rewrite; drops raw openai client + GPT4oJudge
llava-in-the-wild	OpenAIVerifier(comparative parser)	replaces evaluate_comparative
llava-bench-coco	OpenAIVerifier(comparative parser)	preserves system-instruction prefix
browsecomp	StripReasoning + OpenAIVerifier(binary)	gated by metadata use_lmms_judge; exact-match fallback on judge failure
corecognition	StripReasoning + OpenAIVerifier(binary)	gated by metadata use_lmms_judge; direct-comparison fallback on judge failure
voicebench	OpenAIVerifier(1-5 score parser)	partial: only the open-ended judge migrated
ami	OpenAIVerifier(1-5 score parser)	edinburghcstr AMI meeting corpus

All migrations use only the openai provider (present in the verifiers package).

Behavior notes (deliberate changes)

1. mmvu (open-ended): the legacy path passed the judge a rubric-only custom_prompt with no question/answer/prediction substituted — build_binary_prompt returns such prompts unchanged and the provider ignores the request fields, so the GPT rescue never actually graded anything (a latent no-op). The migrated prompt substitutes the real fields, so open-ended mmvu scores will move. This is a bug fix, not drift.
2. browsecomp / corecognition now run StripReasoningExtractor before judging — reasoning blocks no longer reach the judge.
3. ami / voicebench: the legacy judge system messages are preserved as prompt prefixes (same approach as llava-bench-coco's migration).

Validation

• py_compile passes on all 10 files.
• All 139 yaml !function utils.* references resolve.
• ruff: 6 findings, all pre-existing (identical to base, line-shifted only); migration introduces zero new lint.
• Verifiers imports resolve; 5 representative modules load against the real verifiers package.
• Hermetic mocked-judge tests: corecognition/browsecomp fall back correctly on judge failure; ami/voicebench prompts include the original system text.

Notes

• videoevalpro carries two small aggregate fixes also present in fix/task-utils-improvements; a trivial re-merge may be needed depending on merge order.
• voicebench intentionally keeps its llm_judge import: a second, un-migrated judge path still uses it.

Review pass (2026-07-06)

• Merged the verifiers-module review fixes (judge-failure judge_failed metadata, composite confident contract) into this branch.
• Judge infra failures are no longer scored as wrong answers: corecognition falls back to its direct string comparison and browsecomp to its documented exact-match path when the verifier reports judge_failed (the pipeline swallows judge exceptions, so the old except-based fallbacks had become unreachable).
• Restored the ami/voicebench judge system messages (Behavior note 3) and removed an orphaned _create_judge_prompt helper.