very broad theorems like CAP (and e.g. The halting problem) are still very useful for proving things by reduction. Instead of talking about nodes failing, you could talk about a network partition isolating just that one node (sure, the node might process some things on its own, without input, or not --because it crashed-- but for CAP that is totally out of scope, and part of the underlying algorithm). Instead of talking about multi object transactions, it talks about anything that implements a register - which includes multi object transaction systems.

CAP therefore shows what real-life trade offs are feasible in the first place ( of course, there are still systems allowed by CAP that might be impossible for other reasons).