What was wrong?

In the following integration tests that test custom registrations, our method of testing is a bit stateful since it registers then unregisters coders for the custom data type. If one test fails, that could lead to a confusing failure in another test when the registry complains that a registration already exists for the custom type:

https://github.com/ethereum/eth-abi/blob/master/tests/test_integration/test_custom_registrations.py#L69-L99

How can it be fixed?

We could modify those tests to use a fixture registry instance that is created per test run. If we want to continue to do true integration testing via encode_single and decode_single, we will probably need to modify the API to support use of a custom registry instance.