Hi Romain,
I quickly look at the code (using my mobile phone… not easy 
) and I think you are right. It looks like the method returns the estimated measurements based on the state optimized at the penultimate iteration.
I see that the evaluations (i.e., the object returned by the getLastEstimation() method) is updated before the differential correction is applied to the state vector. In other words, the method gives the observed and estimated measurements used to compute the residual vector of the differential correction for the last estimation.
Taking the parallel with the Kalman Filter, they are, for the last estimation, predicated measurements, not corrected.
Best regards,
Bryan