Read a paper's own maths: stop the hang, expand its TeX shorthands

[NeoVand]

Importing https://neovand.github.io/Moire/paper/index.html hung the tab hard enough to need a force quit, and everything that survived the hang read its equations as raw source. Three defects, all reproduced against the live page.

The hang

splitLongSentence walks a long sentence in 280-character windows looking for a , to break on, backing away from any candidate that falls inside an inline equation. When the last remaining candidate sits at index 0, candidate.lastIndexOf(token, index - 1) clamps its start position to 0, finds the same comma again, and loops forever — a synchronous spin on the main thread, so no error, no timeout, nothing to interrupt it.

The window begins on an equation's first character whenever a hard cut lands inside that equation, so a caption written as …an author sees$, which is why…$ reaches the state on its first pass. Delta-debugged down to three lines; the regression test sweeps opener lengths across the window boundary and hangs on the old code.

The page's TeX shorthands

A paper published as HTML renders its own maths in the browser and hands KaTeX a macro table — \idx → \phi, \R → \mathbb{R}, \Rot{\theta} → \mathbf{R}_{\theta}. That table lives in the page's script, which extraction drops, so every custom command arrived as literal \idx: unrendered on the page, and read aloud as raw source.

tex-macros.ts reads the table off the page's inline scripts (KaTeX and MathJax forms alike — bare keys, single quotes, [body, arity] values) and expands it into the maths of the extracted markdown, before parsing, so the equations render, the deterministic verbaliser reads them, and the language model rewriting them sees ordinary TeX. A macro table is data and is never evaluated; sizes, expansion depth and total substitutions are all bounded. Expansion is confined to maths regions and skips fenced code — in prose a lone \R belongs to a Windows path.

Figures with maths in the caption

The DOM prepass folds an empty image alt down from its figcaption so the narration layer has something to describe. Defuddle reads an alt matching \command as a rendered equation image: it dropped the picture and re-emitted the caption as one enormous bogus formula — the very paragraph that hung the splitter. Left inside its figure such a caption fares no better, since the figure is rebuilt from plain text, losing the emphasis, the links, and the delimiters that made the equations equations.

Those captions now move out to a paragraph of their own, and the alt is left empty. Captions with no TeX in them are untouched, and Wikipedia's extraction is byte-for-byte identical.

Verified on the live page

Driven through the real production path in a real browser, not a harness.

	before	after
import	hangs forever	2.3 s
blocks / segments	—	507 / 1625
figures kept	15 (two destroyed)	17
unresolved custom macros	205	0
KaTeX renders / errors	—	1099 / 0

Fig. 4's caption now renders its bold, its links and its four equations, with no duplicate. Wikipedia (Quantum mechanics) extracts identically to before.

npm run lint, npm run check, 551 unit tests, 36 e2e and 12 visual snapshots across three viewports all pass; tex-macros.ts is on the coverage include-list at 97% lines.

🤖 Generated with Claude Code