> I’m just not sure I see a way to fix this without tracking them all.

IIRC we do skip the GC flags for user-created types only when the subtype is not adding new variables *and* the base class is not a GC class by itself. This includes the case with __slots__ but is not limited to. If I recall correctly, there is a bunch of metaclasses that fall into this category and some other minor things so maybe is not that bad to unconditionally make all user objects tracked.

In any case I think is prudent to run the performance test suite with PGO/LTO + CPU isolation to get an idea. Unfortinately we already have some unwanted 3.9 performance regressions of unknown origin and I would like to not add to it if we can.