Hi Dongzhe,
This is partly intended behavior since in our implementation, the calculation of the J_ij matrix elements require that all the atoms are located within a single unit cell. Since your configuration have some negative fractional coordinates, it will internally wrap the configuration and trigger a restart. Note, however, that the second update uses the already updated configuration and only takes 5 steps to converge (vs. 28 for the first update).
Also, I just found out now, that the check for doing a restart is too tight meaning that the program will often perform a restart even if it isn't really necessary. But again, this is just a few additional SCF steps.
Regards,
Troels