Yah, I tend to think of type witnesses as actually existing at runtime and phantom types not, but in the union trick, they don't really exist either. So thinking on it some more, seems more of a way of expressing type parameters in the first place, and well, that's what the article was about.
Now I'm wondering what phantom types would look like in C...
I've mostly seen in done in Haskell, but have used it in it Scala as well to simulate type hierarchies not present in the actual type system.
In a way, this union trick is kind of like a phantom type, since the secondary type is never actually used.