The premise (title) is "if you want less traffic build fewer roads"
I would suggest that this is only true when one adds an additional transaction route (with inherent/intrinsic response time and throughput characteristics) to a model that starts out with an exclusive OR choice of one of two routes, each with only one bottleneck, such that the participants now have a "choice" to take a route that now includes the heretofore impossible case of two bottlenecks.
Queuing systems behave non-linearly after all.
Now, if the "bypass" in the example model had been one that bypassed BOTH existing bottlenecks, it is easily see that even if the new route is moderately slower in it's throughput/response time than the existing two choices; the sum of the system in net, improves. And, of course, if the bypass is intrinsically faster (as used in the example), the system as a whole improves even further.
So the Premise was not only not proven, it was a bit disingenuously stated. In fact: if you want less traffic (i.e. congestion, etc.) you should build more, BETTER (intelligently routed) roads...
Enjoyed the read though