fix(ui): render dollar math next to punctuation without matching prices

[opencode-agent]

Problem

marked-katex-extension only opens $...$ at the start of the text or after a space, and only closes it before whitespace or ? ! . , :. Agents routinely put math against punctuation, so text such as ack ($440\text{ ms}$ IPC p50, …) renders as raw LaTeX. The same rule also rejects the $x$-axis, is $x$; y, and 因此$x^2$是正数.

Its inline start also retries the regex from each $ on every text token, and that regex can scan to the end of the line. Paragraphs with many $ therefore take seconds or minutes to lex. This happens in both the markdown worker and the live streaming lexer.

Change

Replace the extension with three small marked extensions in marked-parser.tsx, and drop the marked-katex-extension dependency.

• $...$ follows Pandoc's tex_math_dollars boundaries:
  • The opening $ must be followed by a non-space.
  • The closing $ must follow a non-space, and must not be followed by a letter, digit, or underscore.
  • The character before the opening $ does not matter.
  • Pandoc only forbids a following digit. The extra letter/underscore rule keeps $HOME/$USER and $PATH:$HOME as text. The cost is that $n$th does not render, which is already true today.
• $$...$$ renders as display math anywhere in a line.
• \(...\) keeps its existing tokenizer.
• $$ / $ blocks keep the extension's block regex unchanged.

Now renders:

• ($440\text{ ms}$), (i.e. $x$), the $x$-axis, is $x$; y, $x$'s, 因此$x^2$是正数, ($$x$$)

Stays text:

• prices: $5 and $10., $5-$10, $5, originally $10, $5 (or $6), $5/$6., between $5 and $6 dollars
• shell variables: $HOME and $PATH, $HOME/$USER, $PATH:$HOME
• code spans: `$x$`
• escaped dollars: \$x$
• $ x $ (whitespace inside the delimiters is now rejected, as in Pandoc)

Regex performance

/\$\$(?!\$)((?:\\.|[^\\\n])*?(?:\\.|[^\\\n$]))\$\$|\$(?=[^\s$])((?:\\.|[^\\\n$])*?(?:\\\S|[^\s\\$]))\$(?!\w)/y

Single match attempt

• The repeated alternatives cannot overlap: \\. starts with a backslash and the other class excludes it. There are no nested quantifiers, and the trailing unit is tried once per position, so backtracking is linear.
• $...$ content cannot contain an unescaped $. A failed attempt stops at the next $.
• A failed $$...$$ attempt stops at the end of the line. Any later reachable $$ on that line would close it successfully.
• recheck reports the new regex as safe.

Across a paragraph

• marked calls start(src) again for every text token, passing the rest of the paragraph. The extension's start retried the regex from every $ in that remainder, which is O(tokens × $ count × line length).
• start now returns the next $, like the existing \( extension. When the tokenizer fails there, marked merges the text tokens back together.
• Reference check: the lexer output from this version is identical to one using a regex-validated start, across 300k fuzzed strings and every .md file in this repo and anomalyco/anomaly.

Benchmark

Median lexer time with Bun 1.4.2 and the repo's marked:

• none: marked with only the \( extension
• old: marked-katex-extension 5.1.13 plus \(
• new: this PR

input	size	none	old	new
realistic agent prose	100 KB	9.8 ms	14 ms	15 ms
`a` $5 repeated	8.7 KB	1.9 ms	15 s	4.1 ms
It costs $5, repeated	108 KB	108 ms	6.8 s	136 ms
$a repeated	100 KB	95 ms	11 s	187 ms
$$a\$ repeated	100 KB	159 ms	669 s	132 ms
$a\\\ repeated	8.6 KB	1.8 ms	13.7 s	4.4 ms
\$$a repeated	100 KB	173 ms	n/a	180 ms

Across every shape tested, new stays within about 2× of none. The remaining superlinear growth on single-line 100 KB paragraphs comes from marked itself: none grows from 0.45 ms at 10 KB to about 97 ms at 100 KB.

A guard test lexes six of these shapes at 10 KB within a 500 ms budget. It takes about 36 ms with this change and about 50 s if start retries the regex from every $.

Tests

• bun test src in packages/ui and packages/session-ui
• tsgo typecheck for both packages
• I could not run the Playwright component tests locally (the shared host ran out of threads). The math story only uses \( and $$ blocks.

Requested by: @Brendonovich (Brendan via Slack)