fix(halueval): read Yes/No verdicts as whole words, not substrings

[Zhuoxi2000]

What

HaluEvalAdapter.match_score credited a sample when reference in filtered_prediction.strip().upper(). For a NO label, any word that contains "no" (not, now, note, know, knowledge, ...) turned a Yes reply into a correct answer. A reply naming both verdicts (Yes and No: ...) was credited for either label, and a reply with no verdict (I do not know.) was credited for No. aggregate_scores derives the predicted class from acc, so precision, recall, f1_score and yes_ratio were affected too.

This PR applies the rule from the official evaluator (RUCAIBox/HaluEval evaluation/evaluate.py): a reply is correct only when it contains exactly one verdict and that verdict equals the label. Replies containing both verdicts or neither are incorrect.

How

• New extract_verdict() returns 'YES' / 'NO' when the reply contains exactly one verdict and None otherwise. match_score credits the sample only when that verdict equals the reference.
• Verdicts are matched as whole words, either as the prompt writes them (Yes / No) or in capitals (YES / NO). The official evaluator uses a case-sensitive substring check ("No" in ans), so it treats Note ... and Nothing ... as a "No". Matching whole words avoids that. Lowercase prose such as "there is no mention of ..." after a Yes is still read as a single Yes verdict, as in the official script. This is stricter than the reference about word boundaries and otherwise follows its exactly-one-verdict rule. It does not reproduce the reference byte for byte. One visible change: a bare lowercase yes / no was credited before and is not now, which matches the official script and the prompt (The answer you give MUST be "Yes" or "No"). If you prefer case-insensitive matching, I can switch, at the cost of reading lowercase prose such as "no mention" as a second verdict.
• aggregate_scores is unchanged. It derives the prediction from acc and is correct once acc is.
• HaluEval evaluation_version is bumped from v1.0 (default) to v1.1 because the scoring semantics change, following the Evaluation versioning rules in AGENTS.md. BenchmarkMeta.description is unchanged, so no _meta / docs regeneration is needed.

Tests

New tests/benchmark/test_halu_eval_scoring.py (14 cases): Yes replies containing not / now / Note with a No label, a both-verdict reply under each label, a reply with no verdict, plain verdicts, and single verdicts followed by an explanation (including YES, NO. and lowercase no in prose).

# before (main @ cfb9af3, test file only)
python -m pytest tests/benchmark/test_halu_eval_scoring.py -q
6 failed, 8 passed

# after
python -m pytest tests/benchmark/test_halu_eval_scoring.py -q
14 passed

python -m pytest tests/api/test_evaluation_versioning.py tests/api/test_benchmark_registry_contract.py tests/api/judge/test_gates.py -q
48 passed

pre-commit run --files evalscope/benchmarks/halu_eval/halu_eval_adapter.py tests/benchmark/test_halu_eval_scoring.py   # all passed
lint-imports   # 5 kept, 0 broken

The same reference in filtered_prediction.strip().upper() pattern exists in pope, hallusion_bench and drivelology_binary. Their prompts ask for a bare YES/NO, so they are less exposed. This PR is intentionally scoped to HaluEval, and I can follow up on those if wanted.

AI assistance: this change was drafted with an AI coding assistant (Claude) and verified locally with the tests above.