What change is responsible for the speedup? The cache check or removing
type(instance) from the set?  Since type(instance) is usually the same
object as instance.__class__, the set will still have only one element
(in particular for Decimal this is the case) so I don't see why that
change is necessary.

It's also important to find out where __instancecheck__ is called; I
don't see where this is happening, so I suspect it's probably somewhere
internal.