feat: add Rust MCP server + fix KaTeX rendering issues

[userFRM]

Summary

• Rust MCP server that uses the markxiv library directly — no dependency on markxiv.org
• KaTeX rendering fixes for common LaTeX commands that break in browser math rendering
• Figure preservation with ar5iv links — addresses Feature request: add links to figures #1

Figure Links (closes #1)

Previously, <figure> blocks from pandoc output were stripped entirely. Now:

1. extract_figure_captions() converts <figure> blocks into numbered markdown blockquotes preserving caption text:

   > **Figure 1:** Architecture of the proposed model
   

2. add_ar5iv_figure_links() enriches each figure reference with a link to the paper's ar5iv HTML page where figures are viewable:

   > **Figure 1:** Architecture of the proposed model — [view on ar5iv](https://ar5iv.labs.arxiv.org/html/2107.02789)
   

Since ar5iv renames figure files to x1.png, x2.png etc. (unpredictable from LaTeX source), we link to the full ar5iv page rather than individual images.

KaTeX Fixes

Added regex-based post-processing in sanitize_markdown():

Issue	Before	After
Missing subscript	\mathcal{X}{Y}	\mathcal{X}_{Y}
Unsupported command	\textsc{Algo}	\textbf{Algo}
Algorithm pseudo-code	\Call{Solve}{x}	\textbf{Solve}(x)
Unsupported font	\mathbbm{1}	\mathbb{1}
Angle brackets in math	$a < b$ → HTML stripped	Math blocks preserved verbatim
Display math inline	text $$x^2$$ text	$$x^2$$ on own line

Also added search() method to ArxivClient trait for arXiv keyword search.

MCP Server

A Rust-native MCP binary (markxiv-mcp) built with rmcp v0.14. Uses the markxiv library directly (no HTTP calls to markxiv.org).

Tools:

• convert_paper — full arXiv paper → markdown via local pandoc pipeline
• get_paper_metadata — title/authors/abstract lookup via arXiv API
• search_papers — keyword search via arXiv Atom API

Claude Desktop config:

{
  "mcpServers": {
    "markxiv": {
      "command": "/path/to/markxiv-mcp"
    }
  }
}

Requirements: pandoc + pdftotext installed locally. The MCP binary can be distributed pre-built so end users don't need a Rust toolchain.

How this differs from other arXiv MCP servers

Most arXiv MCP servers (~10+ exist) either fetch raw LaTeX, scrape HTML, or do basic PDF text extraction. markxiv-mcp runs pandoc locally on actual LaTeX source for much higher fidelity markdown output — the same pipeline powering markxiv.org.

Sanitize Pipeline

The sanitize_markdown() pipeline now has 4 stages:

1. extract_figure_captions — convert <figure> blocks to numbered markdown blockquotes
2. fix_katex_commands — regex fixes for unsupported LaTeX commands
3. normalize_display_math — ensure $$...$$ blocks are on their own lines
4. strip_html_tags_preserve_math — remove remaining HTML while preserving math verbatim

Test Plan

• cargo test --lib — all 50 unit tests pass
• cargo build -p markxiv-mcp — MCP binary compiles
• cargo build — main library compiles
• Test MCP tools manually with Claude Desktop or MCP inspector
• Verify KaTeX fixes against real papers with known rendering issues