yes you have to split based on normals in the modeler if it supports smoothing groups , but I am refering to the question , if its a concern that more vertices are then needed .

and I believe its no concern as long as the vertices stays below 16 bit range , the engine would only need to have all of the mesh data in a single mesh buffer for a single draw call ..

if in fact the model goes over the limit , the mesh can be seperated in to multiple nodes , if needed by the modeler I think .

and I believe the engine splits triangles such that each has its own vertices , it might be the case for blocks , I am not sure but I remember having difficulty with vertice counting sometime and figured out sometimes theres duplicate vertices ,because some triangles have seperate vertices ,which led me to believe the engine splits them .

anyway , I could be very wrong .