There is nothing special about your UserDict here, the same happens for just NumPy arrays:

a = np.array(3, dtype=object)

np.asarray(a)    # just returns a
np.asarray([a])  # returns a list containing a

Unless you want to distinguish np.asarray(a, dtype=object) here, I am not sure there is much to be done about it. The trust is that NumPy has to effectively guess if you want to put in the array as an element, or assign the contents of that same array.

In the first case:

• np.asarray(array) needs to return input arrays unchanged, we can't avoid that really.
• np.asarray([a], dtype=object) often can't avoid packing the whole array, so unpacking it is also awkward... Thus the behavior.

I suppose you could invent a rule that may be slightly more spot-on (with some extra complexity in the code), such as "if all entries are arrays and 0-D, unpack them".
In particular np.asarray([None, a]) is an example where I think it would be arguably wrong to default to unpacking a, since you don't unpack the None (as you can't).

There may be some old discussions about this. I was once convinced that this is as good as it gets, but probably with the constraint "I don't want to change more than necessary".