Dear Axel,
please note, that a regular 3d-ewald is assuming a roughly
spherical shape (=> near cubic for parallelepipeds) of the
charge distribution.
It's interesting. I've never heard about it. I thought that a sufficient number of k-space vectors along each axis provides
a needed precision anyway.
i would expect that you can achieve the same effect without
changing the LAMMPS source code and without much additional
computational cost, by simply using the replicate command
to use real copies instead of periodic images.
Maybe you're right. I used to avoid such replication to minimize a
computational cost. But maybe in LAMMPS an overhead will be indeed paltry.
I'll make a comparison as soon as Dr. Plimpton will make a fix he promised 
Best wishes,
Vadim