Obviously when rendering, the render layers are taken into account, but they are not used when viewing the animation in the viewport. This means not only the objects that can be seen are different, but also there can be collisions between objects that are meant to transition in the animation (e.g. a solid object which turns into a fractured one (which are actually different objects)).
I’ve worked around this a little bit with collision groups, but this seems to only partially address the issue. Any better workflows?