I think about another one reason to use #3. I think you can have some lighting problem with obstruct objects like in case #1.
For me as a unity developer this is just too abstract to judge and this “primitive” case can’t be used to create general rule. In bigger models you can even have less vertices in case #3 than in #1 because inside part of models can be huge. And quantity of variables is quite big – amount of work, scene, number of appearances in the scene etc. Not to mention function of mesh itself: if it’s foreground or background object, if it’s animated and so on. There are some basic rules like all of rules mentioned before (less textures = better, less vertices = better etc.), but rest is project depend.
