I don't think pre-including headers is a good practice.
It can be solved with #define now. A possible way to prevent it is making the reference solution a local lambda.
User's solution can change the vector before it's passed to the reference solution (with an altered function signature). The reference solution should be called first.
One random test is too few.
Although it most likely will work anyway in a case like this, random_device is supposed to be used with distributions.

Thanks for report, I have just corrected it, but no sure why it doesn't show up.

Hi, I've made an C++ translation, please review and approve it.