fix: harden corpus math normalization

[developer0hye]

What changed

• harden corpus text normalization so underscore-bearing identifiers stay intact during source snapshot extraction
• normalize math-heavy headings against extracted Unicode math glyphs
• make LaTeX math stripping line-aware so literal prose delimiters like $$ no longer swallow later content
• add targeted regression tests for identifier preservation, heading normalization, inline/block math stripping, and the texify delimiter edge case
• refresh the 60-entry corpus baselines after re-running validation with the new rules

Why

The validator was producing false negatives on math-heavy READMEs, especially around underscore-bearing identifiers and Markdown lines that mention literal math delimiters. That kept several corpus entries below threshold even when the rendered PDFs were fine.

Verification

• node --test validation/text_metrics.test.mjs validation/readme_corpus.test.mjs
• node validation/readme_corpus.mjs bless --tier all --force
• Full corpus result: 60 total, 0 hard failures, 5 advisory-only remote asset warnings

Related: #13