Although technically "less pure", I think it would be helpful for these tests to show the expected regexp in addition to having pass/fail test patterns. That would, e.g., either have caught the ** issue or made it easy for me to convince myself I was mistaken.

Their implementation seems very odd, I don't understand it either.

(This approach arguably gets \**/log.txt wrong, I guess? Probably not worth fixing.)

Maybe it would be cleaner to structure this like:

if ($escaping) {
  $escape = $long_list; // Full list with everything that gets escaped.
} else {
  $escape = $short_list; // Just the stuff that's always escaped.
}

if (in_array($char, $escape)) { ... }

Then all of the subclauses wouldn't need additional conditions.

This would apply unnecessary escaping to commas, but that seems fine.

I would expect [ and ] to also get escaped.

The delimiter, # must be escaped too (or you can use ( and ) as delimiters without additional escaping).

Won't this convert ** to .*[^/]*? (That is, process both stars?) While I don't think it produces the wrong behavior, I'd expect an $i++ after the .* case.

There's no check for $in_curlies being 0 here, so I'd expect this to produce an invalid regular expression for an input like {. It would be cleaner to throw an error here.

There's no check for $escaping here, so I'd expect this to produce the wrong result for an input with a terminal backslash, like asdf\.