This is the real algorithm. The types are generalised here.

We publish simple types to the user which assume they're working with a 2D grid.

But in reality, we aren't tied to (x,y) positions and we don't care what the terrain model is. Positions can be any type you like (as long as it's comparable so that we can use it as a dictionary key).

But...that makes the type signatures pretty unintuitive to newer users, so we hide them away here.