Kspace ewald (coul/long potential) generates strange forces

You are only turning off the Lennard-Jones interactions. You still have Coulomb interactions and atoms with charges and thus there should be forces due to the charges from the Kspace style. That one always computes interactions between all atoms in the system. It is not a pair-wise computation, so how you assign pair styles in the hybrid pair style is irrelevant.

This can be easily confirmed by changing your thermo_style command to:

thermo_style    custom step etotal epair evdwl ecoul elong

Which produces the following output:

Per MPI rank memory allocation (min/avg/max) = 9.339 | 9.339 | 9.339 Mbytes
   Step         TotEng         E_pair         E_vdwl         E_coul         E_long    
         0  -1.5575404     -1.5575404      0              0             -1.5575404    
Loop time of 2.21e-06 on 1 procs for 0 steps with 2 atoms

As you can see, the contributions are from the long-range solver.