GH-89812: Add pathlib._PurePathExt

[barneygale]

Move __fspath__(), __bytes__() and as_uri() methods from PurePath to a new _PurePathExt subclass. This new subclass is inherited by PurePosixPath, PureWindowsPath and Path.

Because PurePath isn't directly instantiatable (you get a PurePosixPath or PureWindowsPath instance back), this shouldn't change user-facing behaviour.

The methods must not be inherited future tarfile.TarPath and pathlib.AbstractPath classes, which will subclass PurePath.

• Discussion: https://discuss.python.org/t/make-pathlib-extensible/3428/94

• Issue: Add tarfile.TarPath #89812

[barneygale]

The test naming is pretty janky. #104829 will help.

[barneygale]

Would anyone be willing to review #104829? It blocks this PR. (please and thank you!)

[barneygale]

#104829 has landed, so test naming should be much improved! Think this is ready for review again.

[barneygale]

I've revised the implementation such that the methods are actually removed from PurePath. This will allow our future AbstractPath class to inherit PurePath, as it should conceptually:

(Contrast this with the earlier proposed hierarchy where PurePath and AbstractPath are siblings.

The methods are moved to a new _PurePathExt class, which is inherited by PurePosixPath, PureWindowsPath and Path.

This change will not be visible to users who try to instantiate PurePath, because they'll always get a PurePosixPath or PureWindowsPath instance back.

It will be visible to users who subclass PurePath, but that has only just become possible in 3.12. As such I think it's worth backporting this into 3.12 for consistency.

[barneygale]

Here's a version of the diagram that includes the private _PurePathExt class. The diagram in the post above excluded it in order to show the public view of the class hierarchy.

[merwok]

Just a note that I’m meditating on https://hynek.me/articles/python-subclassing-redux/ and linked articles at the moment and wondering if controlling subclass explosion in pathlib should be a guideline 🙂

[barneygale]

Oh I totally agree. I'd prefer to add onlypathlib.AbstractPath, but I'm prevented by the fact that PurePath has an __fspath__() method :)

And if I were to do pathlib over again, I'd scrap the *WindowsPath and *PosixPath classes. That could be controlled by an argument to the PurePath initialiser instead. Oh well...

[barneygale]

I've converted this back to a draft as I'm still in two minds about how to implement this. It's possible I'll add _AbstractPath into this PR, or close this PR and log another, or something else. Sorry for the noise

[barneygale]

Friendship ended with #104810, now #106043 is my best friend.

[CAM-Gerlach]

@barneygale