"engines": {
    "node": ">= 14.0.0"
  }

I guess this version of rollup requires you to have 14 or greater, as I switch, this bit me.

👋 nice job with the tokenizer

Description

From specs: https://drafts.csswg.org/css-syntax/#consume-number

If the next 2 or 3 input code points are U+0045 LATIN CAPITAL LETTER E (E) or U+0065 LATIN SMALL LETTER E (e), optionally followed by U+002D HYPHEN-MINUS (-) or U+002B PLUS SIGN (+), followed by a digit, then:

1. Consume them.
2. Append them to repr.
3. Set type to "number".
4. While the next input code point is a digit, consume it and append it to repr.

The logic looks okay.. but I think you forgot to consume the sign:

&& isInteger(cssText.charCodeAt(shut + 2))

It's + 2 because you didn't consume the sign (didn't do ++shut, so shut still points to e|E). Otherwise the check is always false.

Steps to reproduce

1. Open https://csstools.github.io/parser/
2. Type

#c {
  border: 1e+1px
}

3. see AST displays

Additional

From specs, consume number (called by consume numeric token) should return 6. convert repl to number which is described here: https://drafts.csswg.org/css-syntax-3/#convert-string-to-number
Since the tokenizer is written in JS, you can pretty much ignore it all and just do parseFloat(this.value). Could CSSNumber return, for example:

CSSNumber {
   value: "12e+1",
   number: 120, // Result of parseFloat(value)
   unit: "px",
}

Anyways, this last thing is just a suggestion and I'm not sure how parseFloat would affect performance 😅

Is there a reason this isn't on npm?