That's harder than you might think. How would you test for all possible uses? Even if you looked at something fundamental like an N-port, you'd still have to define metrics for usefulness and many of the topologies in the paper would evade most obvious circuit metrics.
Well, you are right to some extent, but in practice you don't need "all possible uses" and instead you are looking for a specific behavior, which you can easily encode in a test that uses a simulator.
The article mentions that circuit 23 was found using an "exhaustive search", so presumably using an automated means.
But I don't think that works in the general case to find all "useful" circuits. For example figure 27 is a pA source and requires 2 precise control voltages and a clock source. 39 is a peak detector. Does not seem that easily encoded.