Normalize duplicated arXiv LaTeX metadata

[hqhq1025]

Summary

Fixes arXiv metadata formatting issues caused by mixed text/MathJax extraction and raw TeX display.

This is broader than a single paper title: it normalizes duplicated LaTeX/math fragments at seed time and startup, and routes all user-facing arXiv title/abstract/comment/journal-ref displays through a shared formatter.

Examples handled:

• \gtrsim 100\times\gtrsim 100\times -> ≥ 100× for display, and a single cleaned TeX fragment in stored metadata.
• z=1.37z=1.37 -> z=1.37 for display, and $z=1.37$ in stored metadata.
• \mathbb{P}^1\mathbb{P}^1 -> P^1 for display.
• \mathcal{N}=1\mathcal{N}=1 -> N=1 for display.
• GL(d_1)\times GL(d_2)GL(d_1)\times GL(d_2) -> one GL(d_1)× GL(d_2) display fragment.
• Raw URL TeX commands such as \url{...} and \href{...}{...} no longer appear in displayed metadata.

Affected display surfaces now use display fields:

• abstract page
• search results
• category pages
• listing pages
• author pages
• account recent library
• library and starred pages
• export detail
• simulated PDF preview

BibTeX/export download paths keep the cleaned raw metadata instead of the display-formatted text.

Adds:

• sites/arxiv/metadata_cleaning.py
• scripts/check_arxiv_metadata_latex.py

Verification

• python3 scripts/check_arxiv_metadata_latex.py
• python3 -m py_compile sites/arxiv/app.py sites/arxiv/metadata_cleaning.py scripts/check_arxiv_metadata_latex.py
• Flask test-client render checks across:
  • /abs/2604.07983
  • /search?query=2025mkn&searchtype=all
  • /list/astro-ph.CO/recent
  • /category/astro-ph
  • /author/Lemon,%20Cameron
• Verified rendered pages do not contain duplicated \gtrsim 100\times\gtrsim 100\times, adjacent duplicated math blocks, z=1.37z=1.37, \url{...}, or \href{...}{...}.

Fixes #16.