fix(data): honor HumanEval+ version when loading and hashing - #316

Summary

Honor the existing version argument when loading and hashing HumanEval+.

_ready_human_eval_plus_path() resolves version, but then passes the constant HUMANEVAL_PLUS_VERSION to get_dataset_metadata(). Independently, get_human_eval_plus_hash() replaces its caller's version with "default". As a result, requesting an older release can silently read and hash the default release instead.

Forward the resolved/caller-supplied version in these two places, matching the existing MBPP+ implementation. Default-version behavior and HUMANEVAL_OVERRIDE_PATH precedence remain unchanged. No release data, scoring logic, cache format, or dependencies are changed.

Tests

Added offline tests using synthetic JSONL files in pytest's temporary directory. Only make_cache() is mocked: metadata/path construction, file loading, completeness checking, and hashing use the real implementations.

• Verify default and explicit versions for regular, Mini, and NoExtreme paths, independently for the loader and hash helper.
• Populate both requested and default caches with distinct content to detect silently selecting the wrong release.
• Verify that an override path still wins without invoking cache/download setup.
• Before the fix: 6 failed, 7 passed (wrong contents/hashes).
• After the fix: 13 passed.

PYTHONPATH=. uv run --no-project --python 3.13 \
  --with datasets --with tempdir --with wget --with appdirs --with pytest -- \
  python -m pytest tests/test_humaneval_version.py -q

Black 22.6.0, isort 5.12.0 with the Black profile, and git diff --check pass. No benchmark dataset downloads, generated-code execution, model calls, or full-suite run.

AI assistance: implemented and tested with OpenAI Codex.