You're right! I came across this issue when I tried to use my isPrime implementation in a Kata.

The following test will fail:
Assert::That(false, Equals(PrimeNumber::isPrime(4)));

You should replace:
if (num % 2 == 0) return true;
with:
if (num == 2) return true;
if (num % 2 == 0) return false;

Actually, I forgot you were using a regex, so the max possible value of s is 99, which the if statement will handle just fine. My bad.

Good point! The thought crossed my mind while solving, but this passed the tests, so I kept it as-is.

IMO, casts are a red flag for bad design (that's also why the syntax is so ugly), so if you can get rid of one, that's a good idea. It has no runtime impact, indeed.
I am also not 100% certain about how it works, just that it makes code a little more readable as the functor is defined where it's used, instead of the reader having to look it up.
But there are really minor points.

Thanks for you comment! I used static cast to explicitly show that it will be int. As far as I know it has no impact during runtime, i.e. check performed only during compilation.
As for lambda, I use it in C# a lot, but C++ is a mishmash on its own, so this was again for clarity reasons. Besides I don't have a foggiest idea how lamda works in C++ "under the hood" :)

That static cast to int is not required due to the C++ integer promition rules: every arithmetic operation is converted to at least integer size.
Also, have you considered using a lambda instead of the acumCubes function? Worthwile.
Other than that: my compliments that you are using algorithms as much as possible! That makes this the first solution I've read that's actually easy to read and comprehend.