superclass information can be lost in import cycles

[rwbarton]

The following test case fails:

[case testSuperclassInImportCycle]
import a
import d
a.A().f(d.D())
[file a.py]
if 0:
    import d
class B: pass
class C(B): pass
class A:
    def f(self, x: B) -> None: pass
[file d.py]
import a
class D(a.C): pass

with the error Argument 1 to "f" of "A" has incompatible type "D"; expected "B".

This test case is a reduction from mypy itself (a = mypy.types, A = Type, B = TypeVisitor, C = TypeTranslator, D is any type transformation defined in another module d). The test does not fail if the mypy-import of module d in module a is removed. Nor does it fail if the order of the imports in the main module is reversed. I found this counterintuitive at first, but upon reflection I guess it's because the order of traversing imports changes from main(a(d())) to main(d(a())), so whichever module is imported first is actually processed second.

I got as far as determining that something causes the MRO of D to be computed before the base classes of C are known. So the MRO of D ends up being computed as [D, C, object], which does not contain B. Seems that we could know when processing D that it has a base class in a module a that we are still in the middle of processing, and update the MRO of D when we finish processing a.

(#481 is the general ticket for issues involving import cycles.)