vmem_xalloc() is a public interface which allows callers to request particular alignment. Can't it be used instead?

At least, vmem_alloc_aligned() can simply use vmem_alloc() if the requested alignment is satisfied by all members of the arena, and vmem_xalloc() otherwise. Then you'll hit the quantum cache when possible. But presumably the caller would know when that's the case and can decide itself whether to use vmem_alloc() or vmem_xalloc().