Yeah, it's not a very good choice that it returns null for a zero-D array. But it's been like that for 20-some years, so the question is whether we are prepared to take the compatibility hit from changing it.
If we were willing to break things around zero-D arrays, I don't think that's the only thing to change. It's equally silly that array_dims() returns NULL for such arrays, for instance; their dimensions are certainly not unknown. Perhaps an empty string is the right result, though I've not thought about it hard.
I'd also argue that an out-of-range AARR_NDIM result is grounds for raising an error; returning NULL is a poor substitute for reporting data corruption.
In short, if we're to touch this, I'd want somebody to go through all the array functions/operators and see if anything else is weird with zero-D arrays.
Although I see a cost of compatibility break, I agree so NULL in this case is confusing.
The empty array can be taken as possible unlimited dimensional with zero sized dimensions.