Single precision were used in meshes early on. And graphics cards use single precision natively.
But in Rhino, it has more to do with using doubles. Double has 15 decimal digits of precision. Some operations like square root cannit return completely accurate results within 15 digits, so you end up with rounding. Which makes the 15 digits less reliable, say half of them.
So we use tolerance to handle that. And this is where the trouble begins. Set tolerance down to 0.001 inches for NURBS calculations needed for small objects. That leaves only a few significant digits to the left. So if state plane coordinates take up those digits to the left, tolerance limits will fail.
The problem is not doing the civil site calculations at large numbers. Or doing high highly accurate, smaller tolerance calculations in smaller models. It is just doing both in the same model that gets to the edge of possible.