It's not because they wanted to, I guarantee it. It's because it's impossible to not do it. Critically important details, down to the engine's frame timing, how replication is deeply integrated into the engine, and how the default Actor movement code works means you have to not gut just a single implementation detail, but you have to cut whole swaths of code out of use.

I'm dying to be proven wrong, but the intersection of people who know multiplayer networking well, Unreal Engine 4 or 5, and have successfully implemented client-side prediction with reconciliation and rollback in said software and out of it as well, are vanishingly small.

I failed to understand how one could use the existing engine's loop which is variable framerate, the existing replication system, which favors an implementation opaque RPC system over explicit payload structures, and rewrite Actor movement and engine timing subsystems without having a parallel network loop that uses none of the above.

My guess is that it is possible to use Unreal's variable framerate and replication system, but only rewrite Actor movement implementation details by creating a class that inherits from it and overrides the behavior entirely, but you're still stuck with the fact that because the Unreal Engine uses a variable framerate, you have an inconsistent rollback timeline across every client.

You want consistent frame times, in my experience, to have rollback. Does it mean that it's impossible otherwise? No, but you can get caught rolling back in-between game states that don't actually exist on the granularity of the timestep you're running at.