"current_curvature" is not the correct description of what the MPC is outputting in it's curvature_ego state.
The MPC is integrating it's free variable, curvature_rate, such that curvature[0] is actually the desired_curvature before any delay.
inversely: the curvature_rate_desired is the desired rate of change to the setpoint and not the actual curvature rate. 

If we were to set the initial curvature = measured curvature in the MPC initiation these names would be correct. 
This was possibly how it was initially set up but the nomenclature here is now confusing.