`glob.translate` does not correctly handle `]` as first character in character class

[dkaszews]

Bug report

Bug description:

> import glob
> glob.translate('[][!]')
(?s:[][!])\Z

This specific example is explicitly mentioned in man glob:

The string enclosed by the
brackets cannot be empty; therefore ']' can be allowed between the
brackets, provided that it is the first character. (Thus, "[][!]"
matches the three characters '[', ']', and '!'.)

Therefore, the expected regex would be (?s:[\[\]!])\Z, while the actual one simplifies to just (?s:!)\Z, only matching literal '!'.

Related to #130942 , where character classes do not correctly handle path separator.

CPython versions tested on:

3.13

Operating systems tested on:

Linux

[picnixz]

Mmh I think the regex is actually correct. If you do re.match(p, s) with s being [, ] or !, the match is correct or am I missing something?

[picnixz]

Because [][!] does not simplify to !

[dkaszews]

Because [][!] does not simplify to !

Are you sure? It may be Python is using different semantics than those I am used to, but I thought regex, unlike glob, permits empty character classes and simply omits them. https://regexr.com definitely explains it that way.

[picnixz]

Yes I am (I tested it with a pure re.match(p, "[") and it matched) ECMA Script JS regex specs are different (see the differences on regex101). IIRC, in this case, PCRE and C# are also like Python but Java/Rust/JS would tell you "nope". I don't remember with Perl here. cc @serhiy-storchaka

[dkaszews]

Okay, my bad. This actually works as expected in Python:

> [ c for c in map(chr, range(32, 127)) if re.match('[][!]', c) ]
['!', '[', ']']

and also in C# (PowerShell):

32..126 | %{ [char] $_ } | ?{ $_ -match '[][!]' }
!
[
]

While it would be nice for glob.translate to generate an escaped version which works in all engines, this is totally beyond the scope.