humanspeak / svelte-markdown / 25363806545

93%

main: 81%

push

github

fix(extensions): broaden KaTeX delimiter and boundary handling (#282)

* fix(extensions): broaden KaTeX delimiter and boundary handling

Two regex tweaks in `src/lib/extensions/katex/markedKatex.ts` that
fix real-world LLM output we were silently dropping:

1. Block rules accept single-line delimiters in addition to own-line
   - `^\\\[\s*([\s\S]+?)\s*\\\](?:\n|$)` — was `[ \t]*\n…`
   - `^\$\$\s*([\s\S]+?)\s*\$\$(?:\n|$)` — was `[ \t]*\n…`
   LLMs (Claude, ChatGPT, etc.) overwhelmingly emit `$$x = \frac{...}$$`
   on a single line. The own-line-only rule meant those formulas
   survived as paragraph text because the inline tokenizer also
   rejects `$$` openers via `(?!\$)`. The relaxed `\s*` keeps the
   own-line form working too.

2. Inline `$...$` boundary lookahead now accepts `)`, `]`, `}`
   - `(?=[\s?!.,:)\]}？！。，：]|$)` — was `(?=[\s?!.,:？！。，：]|$)`
   The original boundary class (ported from upstream
   marked-katex-extension) didn't include closing brackets, so
   `($e$, $i$, $\pi$, $1$, $0$)` lost the trailing `$0$` because
   it was followed by `)`. Currency strings like `$5,000 across $42`
   still don't match because digits after the closing `$` aren't
   in any boundary class.

Adds five new unit tests covering single-line `$$x$$`, single-line
`\[x\]`, and the `)` / `]` / `}` boundary cases. Existing currency
and AMS tests still pass.

Adds `/test/katex-issue` route as a regression playground — the
exact dataverse transcript that exposed both bugs is rendered top
to bottom, plus 10 isolated fragments with `data-testid` hooks for
direct Playwright assertions.

Verified: 778/778 unit tests, 95/95 KaTeX e2e (5 browsers), live
Playwright walk-through of `/test/katex-issue` confirms all 10
fragments render correctly while currency strings stay text.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* docs(extensions): align KaTeX delimiter table with new behavior

CodeRabbit pass on PR #282 caught two real follow-ups from ... (continued)

476 of 545 branches covered (87.34%)

Branch coverage included in aggregate %.

3 of 3 new or added lines in 1 file covered. (100.0%)

973 of 1006 relevant lines covered (96.72%)

5507.15 hits per line