The most common error when dealing with Kalman filters (and especially EKF) is to think that the convergence of the P matrix is equivalent to real convergence of the estimation.
You need to look at the normalized innovations.
The innovation is the difference between the expected (predicted measurement) to the actual one:
Innov = y - h(x_predicted)
The normalized innovation is the Mahalanobis distance of the innovation, which correlates with the convergence once the P matrix is small:
d^2 = Innov.transpose * Cov(Innov).inverse * Innov
Where Cov(Innov) = Cov(y - h(x_predicted)) = R + H * P_predicted * H.Transpose