Oops, Antoine's right, the release() semantics in the patch are broken, albeit for the precisely opposite reasons: that example will actually blow up with BufferError inside some_library_function().
I withdraw my objection - Antoine's right that release() on a memoryview object has to just mean "drop the reference to the ManagedBuffer instance". Calling release() when there are actual *exported* buffers outstanding should still trigger BufferError though (slices and copies don't count in that case, as they'll have their own independent references to the ManagedBuffer instance).