<antocuni> arigato: yes, that's my theory as well. It looks like the default tp_dealloc for heap types is wrong
<antocuni> so it should be correct to always specify our own
<antocuni> my concrete use case is that I want to automatically DECREF the HPyFields
<antocuni> uh, I think it's more complicate than that
<antocuni> e.g., subtype_dealloc has special logic to call tp_finalize
<antocuni> and if you use a custom dealloc, tp_finalize seems to be NOT called 😱