I think the emphasis on the documentation is a bit of a red herring. Sure, documentation for most things could definitely be better. But when you are building a system that is a composition of subcomponents, the bulk of the complexity is almost always going to come from the interactions. Those aren't things that can be readily captured in the documentation of any single component.

This conclusion is actually pretty obvious when you consider it from a testing/dimensionality perspective. If you have 2 components with N inputs, an interesting application using both may have close to 2N inputs. Sounds pretty tame, but the size of the space grows _exponentially_ with dimension. And for a bounded space, almost all of the volume will be near the surface, which would mostly correspond to non-trivial combinations of the inputs. I think we tend to lump most of the code needed for this complexity into "business logic".