Rizwan Saleem

The source-manipulation layer inside Vite and Rollup. Fixed an infinite loop where two overlapping move() calls spliced the chunk list into a cycle and hung the build (1.2.1); a hasChanged() regression reporting change on byte-identical strings, caught differentially over 20,880 generated operation sequences before it reached a release (1.2.2); and a zero-length-match loop that exhausted memory on any global regexp matching the empty string (1.2.3). A fourth, correcting seven divergences from String.prototype.replace (the worst inserting the literal text undefined into output for a capture group that did not participate), is open for review, with 32,902 of 136,000 differential comparisons disagreeing before the fix and zero after.

Restored recast’s documented print(parse(source)) === source identity, broken when a comment hid an enclosing parenthesis from the printer; verified across 5,076 generated inputs, 2,540 identity violations fixed with zero regressions (0.23.20). Two further fixes are merged but not yet in a published release: ?? combined with || or && printing outright SyntaxErrors, and a hardcoded ecmaVersion silently overriding the caller’s parser options.

Unclosed [, ( and a trailing | threw a raw TypeError instead of the parser’s own structured error, losing the source position that makes a CSS parse failure actionable (7.1.5).

no-cycle aborted an entire lint run with a TypeError whenever the linted file’s own path did not resolve: an unsaved editor buffer, or a resolver that cannot handle absolute paths. Rebuilt the graph rooted at the imported module and cached it, so the common path skips the work entirely.

Exposed defineConfig on a dedicated ./config entrypoint, so a typed configuration no longer has to import the whole tool (knip@6.33.0). The separate-entrypoint design was a commenter’s proposal rather than mine; the implementation and the exports-map work were the contribution.

The fix and the deterministic regression test for a P1 scroll-position bug. The diagnosis here was the reporter’s; the test, which pins a measurement race so it fails reliably rather than occasionally, was the harder half.