Add MedFact-Bench zero-shot evaluation environment

[KouateMuhamed]

Motivation

MedFact-Bench (Jin et al., Med-V1) is a 14,274-example benchmark for medical claim verification and evidence attribution across five component datasets: SciFact, HealthVer, MedAESQA, PubMedQA-Fact, and BioASQ-Fact. It fills a gap in MedMarks, which currently has no claim-versus-evidence verification environment. It is also directly useful for downstream MedARC fact-verification work, including AMFV, where verdict accuracy serves as a training and evaluation signal.

Summary

• Adds the medfact_bench evaluation environment with pinned Hugging Face and local Parquet loading.
• Implements the Med-V1 five-point score parser, three-way verdict mapping, and invalid-prediction handling.
• Adds benchmark-specific raw-results reporting, including per-component accuracy, parsing diagnostics, coverage, class metrics, and paper-comparable macro-accuracy.
• Adds MedFact-Bench to the smoke and verified benchmark configurations, with protocol and reference-result documentation.
• Completes local end-to-end smoke validation with Llama-3.2-3B-Instruct and Qwen2.5-3B-Instruct.

Validation

uv run pytest tests/test_environments/test_medfact_bench.py
uv run pytest tests/test_environments/test_medfact_bench_reporting.py
uv run ruff check environments/medfact_bench tests/test_environments
uv run ruff format --check environments/medfact_bench tests/test_environments
uv run medarc-eval bench --config configs/medmarks-smoke.toml --dry-run
uv build --project environments/medfact_bench

Also verified:

• Ruff check and format check pass.
• medarc-eval bench --config configs/medmarks-smoke.toml --dry-run includes medfact_bench.