That wouldn't be a problem as long as the vertex positions go through exactly the same transformations, as that ends with them accumulating the same rounding er

That wouldn't be a problem as long as the vertex positions go through exactly the same transformations, as that ends with them accumulating the same rounding errors. Assuming that nobody was messing with floating point rounding behavior at runtime or flat out decided to enable fast-math and live with the resulting mess.