Summary

• strip browser annotation artifacts before extractor runs while preserving wrapped text
• align Trafilatura defaults with upstream and make the txt variant use trafilatura.extract()
• pass metric_config into metrics so LLM-enhanced formula splitting is actually used
• add a 545-sample leaderboard reproduction script and refresh README results

Validation

• python -m py_compile examples/run_545_leaderboard.py webmainbench/evaluator/evaluator.py webmainbench/extractors/base.py webmainbench/extractors/trafilatura_extractor.py webmainbench/extractors/trafilatura_txt_extractor.py webmainbench/metrics/base.py webmainbench/metrics/base_content_splitter.py webmainbench/metrics/calculator.py webmainbench/utils/html_cleaner.py tests/test_html_cleaner.py tests/test_metric_config.py tests/test_trafilatura_config.py
• reran 545-sample baseline rows with LLM metric splitting enabled