Out of curiosity, could we enable javassist as the bytecode provider & rerun the test to see whether we get the same StackOverflow? I'm curious if we don't, maybe that can help us knowing the two providers are doing something different here.