If I've understood https://bugs.python.org/issue35290 and its resolution in https://github.com/python/cpython/commit/f6e323ce322cf54b1a9e9252b13f93ebc28b5c24 correctly, that's a concrete example of why I consider the early coercion approach that avoids any kind of redecoding dance significantly more robust against issues arising from future changes in the CPython startup sequence.

I'm away for AWS Re:Invent until Dec 2, but after I get back, I'll finally sit down and take one last pass over this to see if I can come up with a version that Victor is actually willing to review and approve.