It's the same essential problem as with e.g. Wine - if you're trying to reimplement someone else's constantly evolving API with a closed-source implementation, it takes a lot of effort just to barely keep up.

As far as portability, people who care about that already have the option of using higher-level APIs that have CUDA backend among several others. The main reason why you'd want to do CUDA directly is to squeeze that last bit of performance out of the hardware, but that is also precisely the area where deviation in small details starts to matter a lot.