theoretically, if the model is properly trained, it should grok that влъхвъ and magician semantically are the seam thus must land relatively close in the latent space. the same way a cat from an image encodes the same way, so that it can decode later in arbitrary language.