The rationale is that you can use them along with those other versions of UUID that contain nonrandom data.

Assume Bob uses one such not-entirely-random UUID variety and actually makes use of the embedded information. Charlie uses random UUID. Bob and Charlie merge their data.

If Charlie had filled the version bits with random garbage, the merge would poison Bob's use of the embedded information.

Basically, UUIDs are a discriminated union type.

By violating the spec you're assuming you'll never want to use a tool that strictly adheres to the spec. Postel's Law works best because it accommodates some people not following it.