Narrowed field types are rejected

[nickdrozd]

A.x has a union type. B and C inherit from A, and B.x and C.x each have non-union types belong to the union.

class A:
    x: int | str

class B(A):
    x: int

class C(A):
    x: str

mypy --strict accepts this, but mypyc raises errors:

asdf.py:5: error: Type of "x" is incompatible with definition in class "A"
asdf.py:8: error: Type of "x" is incompatible with definition in class "A"

It would be cool if Mypyc would accept this. Barring that, the error message could be better. Something to the effect of "the types are compatible, but Mypyc requires that they be identical".

[JelleZijlstra]

This is correct as this code is technically unsafe (if you have an object of type A, you can do either b.x = 1 or b.x = "1", but if it's actually an instance of B or C, then one of these calls will fail). Mypy generally allows this unsoundness for practical reasons, but mypyc has to be stricter.

However, the error can likely be improved.

[nickdrozd]

Is there a better way to express this logic? In my circumstance, every instance of A will be an instance of B or C; IOW, there are no unspecialized instances of A. Methods defined on the subclasses will usually be defined on A as well, and called as a backup if certain conditions are not met. There is no mutation anywhere.

[JelleZijlstra]

Probably mark it as read-only on the base class by using @property. Read-only attributes are covariant, read-write attributes are invariant.

[nickdrozd]

That suggestion works. Feel free to close the issue, or keep it open for the error message improvement.