[Martin v. Loewis[ > ... > The problem appears to be in the tp_clear. The task of tp_clear is to > clear all references that may participate in cycles (*not* to clear > all references per se). That's the key insight, and one we all missed. Thanks for sharing your brain, Martin!