perf: split lookbehind patterns out of the partial-selector regex (WebKit 3–6× faster parse)

[tarek101]

Fixes #439.

JavaScriptCore (Safari, WKWebView) can't JIT-compile a regex that contains a lookbehind. Three PARTIAL_SELECTORS patterns have one, so the whole ~550-way PARTIAL_SELECTORS_REGEX ran in JSC's interpreter. In WebKit, removeBySelector took 150–1,900 ms per parse on long pages, against 3–5 ms in Chrome.

Change

• compileAlternation() in src/constants.ts compiles the patterns without a lookbehind into one regex and those with one into another. It returns { test }, which is true when either matches. A string matches an alternation exactly when it matches one of its parts, so results don't change.
• It applies to PARTIAL_SELECTORS_REGEX and PARTIAL_SELECTORS_ANCHORED_REGEX. Both are internal and only called with .test() (src/removals/selectors.ts).
• tests/partial-selectors-regex.test.ts checks that both testers agree with the combined regex: on edge cases for each lookbehind pattern and on every literal token (1,109 cases).

Results

WKWebView (Safari 26.2 engine), defuddle full build on a cloneNode(true) copy, median of 3 runs. Output content is byte-identical before and after on all three pages.

Page	removeBySelector before → after	parse() before → after
Quanta Magazine article	153 → 13 ms	170 → 29 ms
Wikipedia, Moons of Jupiter	448 → 48 ms	540 → 158 ms
Wikipedia, Fourier transform	1,882 → 154 ms	2,200 → 620 ms

npm test passes: all library tests, including the new file. The only failures are the website test files, which need the website/ dependencies installed. npm run lint:unused and npm run build also pass.

🤖 Generated with Claude Code