Why does setting a descriptor on a class overwrite the descriptor?

Barring any overrides, B.v is equivalent to type.__getattribute__(B, "v"), while b = B(); b.v is equivalent to object.__getattribute__(b, "v"). Both definitions invoke the __get__ method of the result if defined.

Note, thought, that the call to __get__ differs in each case. B.v passes None as the first argument, while B().v passes the instance itself. In both cases B is passed as the second argument.

B.v = 3, on the other hand, is equivalent to type.__setattr__(B, "v", 3), which does not invoke __set__.

You are correct that B.v = 3 simply overwrites the descriptor with an integer (as it should). In the descriptor protocol, __get__ is designed to be called as instance attribute or class attribute, but __set__ is designed to be called only as instance attribute.

For B.v = 3 to invoke a descriptor, the descriptor should have been defined on the metaclass, i.e. on type(B).

>>> class BMeta(type): 
...     v = VocalDescriptor() 
... 
>>> class B(metaclass=BMeta): 
...     pass 
... 
>>> B.v = 3 
__set__

To invoke the descriptor on B, you would use an instance: B().v = 3 will do it.

The reason for B.v also invoking the getter is to allow user's customization of what B.v does, independently of whatever B().v does. A common pattern is to allow direct access on the descriptor instance, by returning the descriptor itself when a class attribute access was used:

class VocalDescriptor(object):
    def __get__(self, obj, objtype):
        if obj is None:
            return self
        print('__get__, obj={}, objtype={}'.format(obj, objtype))
    def __set__(self, obj, val):
        print('__set__')

Now B.v would return some instance like <mymodule.VocalDescriptor object at 0xdeadbeef> which you can interact with. It is literally the descriptor object, defined as a class attribute, and its state B.v.__dict__ is shared between all instances of B.

Of course it is up to user's code to define exactly what they want B.v to do, returning self is just the common pattern. A classmethod is an example of a descriptor which does something different here, see the Descriptor HowTo Guide for a pure-python implementation of classmethod.

Unlike __get__, which can be used to customize B().v and B.v independently, __set__ is not invoked unless the attribute access is on an instance. I would suppose that the goal of customizing B().v = other and B.v = other using the same descriptor v is not common or useful enough to complicate the descriptor protocol further, especially since the latter is still possible with a metaclass descriptor anyway, as shown in BMeta.v above.