Thank you for this fork of css. It's very helpful to us.

While testing, I think I have found a regression in the parser, introduced with commit 24ed6e7:

The regular expression /("|')(?:\\\1|.)*?,(?:\\\1|.)*?\1|\(.*?,.*?\)/g seems to be too greedy for cases like div[class='foo'],div[class='bar']. For this example, it captures 'foo'],div[class=', leading to an incorrect replacement of the comma that separates the two selectors and is not part of the data selector.

The original regular expression, used before this change (/"(?:\\"|[^"])*"|'(?:\\'|[^'])*'/g) handles this case correctly by matching 'foo' and 'bar'.

Ultimately, the current behavior leads to an incorrect AST, where instead of two selectors, only one (incorrect) selector is listed:

"type": "rule",
"selectors": [
    "div[data-value='foo'],div[data-value='bar']",
],

Expected:

"type": "rule",
"selectors": [
    "div[data-value='foo']",
    "div[data-value='bar']"
],

I have attached a test case demonstrating this. You can extract the archive directly into test/cases/:
case - selectors-attributes.zip