Second change: Use a string search instead of RegExp to match ansi codes

Result:

  original:
    23 692 ops/s, ±0.73%   | slowest, 17.61% slower

  for loop:
    27 710 ops/s, ±0.94%   | 3.64% slower

  for-loop + no regex:
    28 756 ops/s, ±0.58%   | fastest

Third change: Don't spread the string into an array, loop through codepoints directly

Result:

  original:
    24 256 ops/s, ±2.36%   | slowest, 41.81% slower

  for loop:
    28 719 ops/s, ±2.07%   | 31.1% slower

  for-loop + no regex:
    30 076 ops/s, ±1.17%   | 27.85% slower

  no regex + no spread:
    41 685 ops/s, ±0.89%   | fastest

4th change: Avoid converting ansi codes to number for end code lookup

Result:

  original:
    27 402 ops/s, ±0.76%   | slowest, 41.57% slower
    
  (snip)

  no regex + no spread:
    43 911 ops/s, ±0.81%   | 6.36% slower

  no conversion between number and string:
    46 895 ops/s, ±0.98%   | fastest

5th change: Don't use string.split to access first char before first ";", use indexOf and slice

Result:

  original:
    22 828 ops/s, ±1.84%   | slowest, 49.89% slower

  (snip)

  no conversion between number and string:
    42 327 ops/s, ±0.90%   | 7.08% slower

  indexOf instead of splitting:
    45 553 ops/s, ±0.90%   | fastest

I think it'd be a lot easier for maintainers to review this PR and much more likely to merge your changes, if it was split up into much smaller PRs, focused on making each individual performance optimization. Right now it's hard to understand, because the whole library basically get rewritten, which would be a tough sell.

Guess I can rip out the last two commits and put them up as an alternative. The others consecutively build on each other and sometimes undo what a previous one did, so splitting them into individual PRs doesn't make much sense to me.

Edit: done