New check: re.sub[n]/re.split with regex flag as third argument w/o keyword

[jakkdl]

Context

• re.sub() doesn't substitute for all words with the replacement in the string when re.IGNORECASE is used and counter not specified python/cpython#99918 (comment)
• re.sub confusion between count and flags args python/cpython#56166

Example code

$ re.sub('a', 'b', 'aaa')
'bbb'
$ re.sub('a', 'b', 'aaa', re.IGNORECASE)
'bba'
>>> re.split(' ', 'a a a a')
['a', 'a', 'a', 'a']
>>> re.split(' ', 'a a a a', re.I)
['a', 'a', 'a a']
>>> re.split(' ', 'a a a a', flags=re.I)
['a', 'a', 'a', 'a']

The fourth parameter to re.sub[n] is in fact count, and the third to split is maxsplit, and as seen in the above linked issues this is a very common mistake.

Implementation

The check would consist of checking calls to re.split, re.sub and re.subn, with three or more parameters, where the third parameter does not have an explicit keyword.

To make it more lenient, it could require the third param to match re.[uppercase], or match any all-caps variable to cover from re import IGNORECASE, or any variable at all to cover myregexflags = re.I | re.X.
The WIP cpython PR's seem to be on track to enforcing the 3rd and 4th parameters to be kw-only though, and in case that actually becomes deprecated by 3.12 then the maximally strict variant of simply enforcing that when running <3.12 seems clearly appropriate.

[jakkdl]

Could even go so far as to require keyword on all re functions for parameters other than pattern, repl or string. So e.g. re.match('a', 'A', re.I) would raise an error. I don't think that's the direction cpython is going, and would lead to a lot of false alarms, but could possible put that version of it as an opinionated warning.

[Zac-HD]

I'm inclined to keep this simple: only apply the warning to re.sub, re.subn, and re.split, but require that count and maxsize (and therefore also flags) always be passed as keyword arguments to these functions, irrespective of value.

Requiring flags= for all other functions seems like an out-of-scope style issue, since they don't share the possibility of argument confusion. Since I dislike making users think about configuration, I'd leave this out.