Recent change to tuple constructor breaks subclasses of tuples

[erictraut]

A recent change to the tuple.__new__ method in builtins.pyi introduced an overload:

    @overload
    def __new__(cls) -> tuple[()]: ...

This overload is problematic for classes that derive from specialized tuples.

class MyTuple(tuple[int, str, str]):
    ...

x = MyTuple()
reveal_type(x) # tuple[()] !!!!

Because of this overload, the wrong type is evaluated.

I think this overload should return Self, not tuple[()]. Admittedly, that doesn't produce the ideal type evaluation for tuple(). It produces tuple[Any, ...] rather than tuple[()].

[AlexWaygood]

class MyTuple(tuple[int, str, str]):
    ...

x = MyTuple()
reveal_type(x) # tuple[()] !!!!

I feel like maybe a type checker should disallow this snippet. This is what happens at runtime:

>>> class MyTuple(tuple[int, str, str]): ...
...
>>> MyTuple()
()

Perhaps type checkers should disallow calls to subclasses of specialised tuples unless __new__ has been overridden?

[JelleZijlstra]

@AlexWaygood the runtime type of MyTuple() is in fact MyTuple. I don't see why type checkers should disallow that.

This change is from #7454. I already brought up this possible issue, but for whatever reason mypy doesn't infer the wrong type for subclasses in this case. Given that it causes a regression for pyright, I'm fine with reverting it.

[AlexWaygood]

@AlexWaygood the runtime type of MyTuple() is in fact MyTuple. I don't see why type checkers should disallow that.

The runtime type is MyTuple, yes, but it's not a subtype of tuple[int, str, str]. My concern is about situations like this:

class MyTuple(tuple[int, str, str]):
    ...

x = MyTuple()  # type checker is happy with this, infers type of x as MyTuple

# type checker is happy with this due to x having inferred type of MyTuple,
# and due to MyTuple being a subtype of tuple[int, str, str].
# However, this fails at runtime.
a, b, c = x

That feels pretty unsafe to me.

But I'm fine with reverting #7454 for now; I agree that it doesn't bring much benefit.