@cgohlke Does this still fail on rc4?

It still fails. Only #2086 was fixed.
The test_contour failure is harmless. Not sure it's worth fixing the test or marking it knownfail on win-amd64-py2.x.

I don't understand that error message. I had a cursory look at the test and this is the function call:

segs = path.iter_segments(transforms.IdentityTransform, clip=(0.0, 0.0, 100.0, 100.0))

The first argument is indeed of the right type, so this is rather confusing. As for it failing on the while loop, I have no idea what's going on there. Both i and len_vertices are just integers.

Could it be an incorrectly handled exception in the C++ extension? If I add try: raise except: pass statements after the call to _path.cleanup_path, before the line while i < len_vertices, the test passes.

Maybe I am missing something, but according to the docstring the transform parameter to path.iter_segments should be a instance of matplotlib.transforms.Transform while the test passes the class. In fact, using segs = path.iter_segments(transforms.IdentityTransform(), clip=(0.0, 0.0, 100.0, 100.0)) fixes the test failure. But how/why did passing the class work on platforms other than win-amd-py2.x?

Oh that's a good catch!

I have no idea why that doesn't work on windows. I feel like if that failed on windows it should fail everywhere.

@cgohlke: That fix certainly seems reasonable, and logically it definitely should be an instance as you proposed. When the transform is ultimately passed to C++, it tries to convert it to a transform, and failing that falls back to the identity transform... so one would expect to see the same behavior even though the input is nonsense. I haven't looked, but I suspect you're right that it's in the handling of the exception (or not) that the difference between Windows and other platforms is revealed. I've marked this as blocker and will endeavor to fix the underlying problem as well as the obvious one (fixing the test) before the next release candidate.

I have what I hope are fixes for all of the above. I'm not clear on why it wasn't failing on Linux/OS-X, but there was definitely an uncleared Python exception sticking around. @cgohlke: Can you confirm this fixes these tests?

I can confirm that the change to agg_py_transforms.cpp fixes the TypeError.

The changes to test_contour.py also seem to work although I get some seemingly unrelated new ImageComparisonFailure. I'll postpone full testing to the new release candidate...