Thanks Masayuki for the updated patch, and especially for the additional test cases.

Looking at the patch, it occurs to me that this solution seems to be the minimal needed to fix the issue that we were originally trying to fix, without adding too much additional complexity or new semantics to how TLS keys are used in the interpreter.  In other words, this preserves the existing usage with minimal changes except to better support opaque key types.  So I think it's a point in favor of this change that's managed to remain focused.

I'll work on updating the PEP draft to reflect the additions.