We assume an idealized behavior of a support in order to simplify the analysis. Generally, one would assume a x-y analysis separate from an x-z.
I think your instinct is correct in that a roller support can realize a bending load/moment perpendicular to the translation direction allowed by the support though this can be minimized by design.
Assumptions in analysis:
Roller supports are free to rotate and translate along the surface upon which the roller rests. The surface can be horizontal, vertical, or sloped at any angle. The resulting reaction force is always a single force that is perpendicular to, and away from, the surface. Roller supports are commonly located at one end of long bridges. This allows the bridge structure to expand and contract with temperature changes. The expansion forces could fracture the supports at the banks if the bridge structure was "locked" in place. Roller supports can also take the form of rubber bearings, rockers, or a set of gears which are designed to allow a limited amount of lateral movement.
A roller support cannot provide resistance to a lateral force. Imagine a structure (perhaps a person) on roller skates. It would remain in place as long as the structure must only support itself and perhaps a perfectly vertical load. As soon as a lateral load of any kind pushes on the structure it will roll away in response to the force. The lateral load could be a shove, a gust of wind or an earthquake. Since most structures are subjected to lateral loads it follows that a building must have other types of support in addition to roller supports.