Fix patching of io.TextWrapper in tests

[akaihola]

Description

patch("io.TextIOWrapper") wouldn't actually cause io.TextIOWrapper() to return the StringIO object created by the test. This didn't cause any problems though until I added a test in #2484 in which content= is non-empty:

>               f.write(dst)
E               ValueError: I/O operation on closed file

src/black/__init__.py:849: ValueError

In the debugger we can see that io.TextIOWrapper used on src/black/__init__.py:842–844:

            f = io.TextIOWrapper(
                sys.stdout.buffer, encoding=encoding, newline=newline, write_through=True
            )

is not the patched lambda:

(Pdb) io.TextIOWrapper
<class '_io.TextIOWrapper'>

With the changes in the PR, it is properly patched:

> src/black/__init__.py(850)format_stdin_to_stdout()
-> f.write(dst)
(Pdb) io.TextIOWrapper
<Mock name='mock.TextIOWrapper' id='139749471475408'>

Checklist - did you ...

• Add a CHANGELOG entry if necessary?
• Add / update tests if necessary?

[JelleZijlstra]

Can you clarify why the previous way didn't work? It looks to me like it should.

[akaihola]

Can you clarify why the previous way didn't work? It looks to me like it should.

@JelleZijlstra, a very good question, which was a tough one to sort out!

Indeed, in the main branch, it works just fine. But in #2484 I had to introduce the extra call to decodebytes() in format_str() (see diff). And since decodebytes() instantiates an io.TextIOWrapper() and closes it, the io.TextIOWrapper() invocation we intended to patch in format_stdin_to_stdout() wouldn't work anymore.

In other words, this is a case of a common problem when patching and asserting library classes/functions: You can't be sure that the patched item won't in the future get extra invocations in addition to the one you're originally testing for.

For some reason, patching only the io module imported in black/__init__.py fixes this issue. I can't really figure out now why that is the case, since decodebytes() is in that same module.

As a conclusion from looking into this deeper, the modification to patching in the test function really belongs in #2484 since it's only the changes there which break the patching. So closing this PR.

[JelleZijlstra]

Thanks for the thorough explanation!