Address researcher review of the paper

[antonkayai]

Summary

• Address the annotated PDF review: clarify the extraction task, domain terms, stressors, OCR checks, and artifact availability; tighten the abstract and surrounding prose.
• Introduce metrics before construction, define micro/macro F1 and strict completeness, and document the existing claim normalization rules. Dataset, predictions, scorer behavior, and published scores are unchanged.
• Remove Author Contributions, put dataset/code links on separate first-page footnote lines, fit the appendix stressor map on one page, and exclude macOS metadata from the arXiv archive.
• Update manuscript/report consistency assertions for the revised wording.

Verification

.venv/bin/python -m pytest -q
make -C paper arxiv
git diff --check
.venv/bin/python benchmarks/check_evaluation_report.py --claims-dir data --results-dir benchmarks/results/codex_gpt56_sol_full_current_ocr_v2 --require-full-corpus
.venv/bin/python benchmarks/check_evaluation_report.py --claims-dir data --results-dir benchmarks/results/codex_full_current_ocr_v2 --require-full-corpus
.venv/bin/python benchmarks/check_evaluation_report.py --claims-dir data --results-dir benchmarks/results/claude_fable5_full_current_ocr_v2 --require-full-corpus
.venv/bin/python benchmarks/check_evaluation_report.py --claims-dir data --results-dir benchmarks/results/claude_opus48_full_current_ocr_v2 --require-full-corpus

203 tests pass; all four reports reproduce. The 29-page PDF was visually checked, and the extracted arXiv source archive compiles independently with its included bibliography.