non-neutral netcharge (i.e. q=+/1 1) with kspace in LAMMPS

The first thing to check is convergence. If you greatly increase the cutoff length (say to 30 Angstrom) and decrease kspace relative error tolerance (say to 1e-15) with the Ewald sum do you get the same answer?

Remember that LAMMPS gave you a warning; using the Ewald sum (or PPPM) with non-neutral systems may be risky.