I would check it differently:

if isinstance(sequence, _FakeSequence):
    if sequence.size is None:
        return True
    else: 
        if sequence.size == length:
            return True
        else:
            return False 

For me a sequence fits for any size, therefore, it is always TRUE, regardless of the checked length.

This behaviour will change if you set a length. From this moment on the size is set. If you now check the wrong length it is not valid any more.

I do not like the hidden setter. I would not expect the function to set the size of the sequence.
If you really want this hidden magic, I would prefer the following:

if isinstance(sequence, _FakeSequence):
    if sequence.size is None:
        sequence.size = length
        return True
    else: 
        if sequence.size == length:
            return True
        else:
            return False 

This will fail if I ask for a different length later. In your function the size will be updated, even if the check ask for different length.

I understand what you are saying. Still, I tried to avoid this being a "hidden setter": According to the docstring, valid_sequence explicitly makes the argument a valid sequence of the given length. I did this to solve #1115: In your results, you might want to iterate over all values, real and fake sequences, in the same way.

I see three ways of doing this:

1. Proceed as suggested in this PR.
2. Rename the function valid_sequence to something like assure_valid_sequence to emphasize possible changes by the function.
3. Make this even more explicit, e.g. by
   a. using the explicit setter (_Fakesequence.size = length) when needed, or by
   b. renaming valid_sequence to make_valid_sequence. This would then raise an Error instead of returning False.
4. Completely remove the _FakeSequence and keep scalars. Then, cast on demand (to an np.array).

I'd opt to do something quick for the v0.5 release (so that the old behavior is maintained) and to remove the _FakeSequence in v0.6.