[highlights] fix highlights lexers

[ije]

Hangs and quadratic scans

• Elixir: a regex sigil spanning a newline no longer hangs.
• Replace repeated scans with bounded or cached lookahead in Julia, F#, Ruby, SCSS/Less/Sass, Bash, Elixir, C#, XML, and PowerShell.

Fresh local stress-test timings on roughly 256 KB inputs:

Input	Main	This branch
Julia ;f( chains	5.15 s	2.3 ms
F# interpolated string with repeated }}	546 ms	1.1 ms
Ruby string with repeated interpolations	479 ms	3.9 ms
SCSS repeated unclosed #{	422 ms	3.4 ms
Bash long string with $x expansions	374 ms	0.8 ms
Elixir heredoc with escapes	169 ms	0.7 ms
C# repeated $"…" strings	117 ms	2.0 ms
XML <!DOCTYPE [...]> subset	81 ms	0.4 ms
PowerShell dash-free input triggering Verb-Noun checks	81 ms	1.2 ms

Highlighting fixes

• JSX/TSX: comments inside open tags no longer consume the rest of the file as a string. Handle component type arguments such as <Select<Option> /> and distinguish generic arrows such as <T extends X>(…) and <const T,>(…) from JSX.
• TypeScript: fix arrow return types, uppercase names in type positions, qualified constructors, and contextual keywords such as using, accessor, and asserts. A template: property selects embedded markup only when its contents look like markup.
• Svelte/Astro/Vue: preserve attributes after {…} expressions, including across streamed lines; fix Vue interpolation boundaries. <style lang="scss|sass|less"> selects the corresponding CSS dialect.
• Markdown/MDX: fix autolinks, badge links, fence info strings (rust,ignore, js{4}), intraword underscores, emphasis containing code spans, and fences inside block quotes. MDX import/export blocks follow JavaScript lexical state.
• Fenced code: resolve all registered language names and aliases, preserve multiline state inside quoted fences, and recognize document-start syntax such as shebangs and front matter within fence bodies. Live edits restore TOML nesting state inside Markdown/MDX fences.
• YAML: preserve ${{ … }}, spaces, commas, and embedded brackets in plain scalars; recognize numeric and escaped quoted keys; allow comments after block-scalar headers such as | # comment.
• Shell and build files: fix CMake unquoted escapes, Bash arithmetic shifts and command positions, flags and paths, and Dockerfile comments or blank lines inside continued instructions.
• Ruby/Perl/PHP: fix block-comment terminators, distinguish shifts from heredocs, and recognize heredocs after Ruby member calls and Perl filehandles. PHP single-quoted strings escape only quotes and backslashes.
• SQL: recognize T-SQL temporary tables and PostgreSQL JSON operators without starting comments; distinguish table column lists from function calls.
• C family: recognize complete hex, octal, and universal escapes; preserve C++ raw-string suffixes when streamed; fix C++ namespace paths, HLSL declarators, and Zig continue expressions and slice sentinels.
• JVM/.NET/mobile: improve contextual keyword classification, trailing-lambda calls, import boundaries, and declaration handling. Comparisons and conditional operators no longer leave stale type state.
• Functional languages: fix OCaml/F#/Lisp parameters and bindings, trailing-dot floats, F# ranges and quoted identifiers, and comment-banner scanning.
• Other fixes: add Python t-string highlighting; improve CSS escaped selectors and dialect constructs, PowerShell keys, Rust paths and tuple fields, assembly registers, and grouped Go type declarations.

Performance

Combined tokenization time is approximately 4% lower across two local A/B runs of the language samples, each repeated to at least 256 KiB. These measurements use the raw Wasm record-output mode.

Language	Approximate time change
Astro	−24%
Svelte	−18%
Makefile	−16%
Vue	−14%
Lisp, Dockerfile	−11–12%
HTML, OCaml	−10%
JavaScript, TypeScript, TSX	−8–10%
Fortran	−8–10%
Swift	+5%
PHP, Perl, YAML	+3–4%

Small differences are sensitive to benchmark noise. Improvements include single-pass framework expressions, lazy JSX/signature checks, shared scanners, and cheaper byte-set checks.

Size and memory

• Wasm grows from 158,514 to 161,909 B (+2.1%); gzipped size grows from 70,919 to 75,322 B (+6.2%).
• Streaming checkpoints grow from 1,280 to 1,536 B. Byte-set storage grows by 1,024 B, and keyword-table reservations grow by 128 B.
• Total reserved space in the first memory page grows by 1,408 B. Input and the live heap still begin at byte 65,536.
• Centralize language registration, name lookup, and dispatch in languages.wat; generate Markdown fence lookup from the same registry.

Intentional behavior changes

• C: in /* c */ #define X, the # is now an operator because another non-whitespace byte precedes it on the same line.
• SQL: a spaced COUNT (*) is no longer highlighted as a function call; COUNT(*) still is.
• Rust: path calls such as Self::new() use function; dot calls such as value.method() retain function.method.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
pierre-diffshub	Ready	Preview	Sep 25, 2026 9:31am UTC
pierre-docs-diffs	Ready	Preview	Sep 25, 2026 9:31am UTC
pierre-docs-trees	Ready	Preview	Sep 25, 2026 9:31am UTC

1 Skipped Deployment

Project	Deployment	Actions	Updated
pierrejs-diff-demo	Skipped		Sep 25, 2026 9:31am UTC

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: ba3932f407

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".