v3.5.1

Patch Changes

• perf: stream potential tokens in OriginalSource instead of materialising an array (by @alexander-akait in #246)

  OriginalSource.streamChunks (and therefore map() / sourceAndMap()) previously built the full splitIntoPotentialTokens array of substrings and then iterated it — even though map() and sourceAndMap() run with finalSource: true and discard every chunk substring. The scan is now streamed by offset, so chunk substrings are only allocated when actually emitted. This removes the intermediate array and, on the dominant final-source paths, all per-token slices: map() / sourceAndMap() allocate ~38–46% less memory and run ~15–40% faster.

• Reduce allocations and CPU in map() / sourceAndMap(): mappings are serialized into a reused byte buffer instead of per-mapping strings, ReplaceSource verifies original content through a line-offset index instead of splitting sources into line arrays, and ReplaceSource.streamChunks emits position-only chunks and returns the final source directly when finalSource is requested. (by @alexander-akait in #251)

• Skip sorting ReplaceSource replacements when they were added in order. (by @alexander-akait in #249)

• perf: use lookup table in splitIntoPotentialTokens for faster character classification (by @xiaoxiaojx in #240)

  Replace multi-comparison chains (4 comparisons in phase 1, 6 in phase 2) with a single Uint8Array bitmask lookup per character. This reduces per-character branching overhead, yielding ~7% improvement on typical source and ~21% on large sources.