It is a little hard to tell from just those traces.
But I think it looks like you need a lot more P, It looks like you are running mostly open loop now, only relying on the feed forwards to move the machine. See how the error grows at a constant rate until the end of the move, then it slowly shrinks as the very weekly tuned PID loop slowly reels the error back in.
I like to also look at the velocity feed back, beside the velocity command trace. Place them right over top of each other so that you can easily see how closely the the velocity feedback follows the velocity command. If you are using a Mesa card the hal pin should be something like hm2_5i25.0.encoder.nn.velocity. If your encoder reader does not have a velocity feedback, (pretty sure most all do) you can get a mathematically calculated one from the PID on the pin pid.N.feedback-deriv.
