Turn TextIOWrapper(buffer) into a protocol

[srittau]

Closes: #11418

[srittau]

This could also use some regression tests, but I'm currently a bit short on time.

[AlexWaygood]

Why do we need all those empty __init__.py files in the test_cases directory? They appear to be causing some confusion for our regr_test.py script

[srittau]

Why do we need all those empty __init__.py files in the test_cases directory? They appear to be causing some confusion for our regr_test.py script

Because otherwise mypy gets confused about check_io.pyi and typing/check_io.pyi. But renaming the latter should help as well.

[AlexWaygood]

Why do we need all those empty __init__.py files in the test_cases directory? They appear to be causing some confusion for our regr_test.py script

Because otherwise mypy gets confused about check_io.pyi and typing/check_io.pyi. But renaming the latter should help as well.

Ugh. Yeah, renaming seems the easiest option there for now... thanks!

[AlexWaygood]

Hmm, looks like we've been here before:

• Use a Protocol as the argument for io.TextIOWrapper #10229

The differences that PR has to this PR are:

• The read() method in the new protocol returns ReadableBuffer rather than bytes in that PR.
• That PR overrides TextIOWrapper.detach() and TextIOWrapper.buffer() to return instances of the new protocol

[srittau]

Hrm, cc @JelleZijlstra. Since this PR seem to fix the issues without adding to the primer output. I suggest to merge this first, and then have another look at Jelle's old PR.

[AlexWaygood]

Hrm, cc @JelleZijlstra. Since this PR seem to fix the issues without adding to the primer output. I suggest to merge this first, and then have another look at Jelle's old PR.

That sounds reasonable to me, though if it would be more accurate to have the read() method return ReadableBuffer, I think it would be safe to make that change. That change should reduce the number of false-positives in user code, if anything.

[JelleZijlstra]

Thanks for picking this up. The major remaining difference from my PR seems to be that you're not overriding buffer and detach. This helps to get rid of some of the primer hits.

[JelleZijlstra]

Suggested change

	def read(self, size: int = ..., /) -> ReadableBuffer: ...
	def read(self, size: int, /) -> ReadableBuffer: ...

It always calls read with an argument. Note this change makes the protocol more permissive.

[srittau]

Unfortunately that's not true: https://github.com/python/cpython/blob/81e140d10b77f0a41a5581412e3f3471cc77981f/Modules/_io/textio.c#L1977.

[AlexWaygood]

Maybe add a comment linking to that line in the source code?

[github-actions]

Diff from mypy_primer, showing the effect of this PR on open source code:

streamlit (https://github.com/streamlit/streamlit)
- lib/tests/streamlit/web/server/component_request_handler_test.py:176:38: error: Argument 1 to "TextIOWrapper" has incompatible type "str"; expected "IO[bytes]"  [arg-type]
+ lib/tests/streamlit/web/server/component_request_handler_test.py:176:38: error: Argument 1 to "TextIOWrapper" has incompatible type "str"; expected "_WrappedBuffer"  [arg-type]

[AlexWaygood]

LGTM. #11420 (comment) is optional.

[AlexWaygood]

I think this also fixes #6061?