I'm new on LAMMPS and don't understand the physical and computational difference of handling long-range electrostatic interactions through lj/cut/coul/cut or lj/cut/coul/long.
You need to recognize the definition of short-range and long-range interactions. Consider an interaction given by the potential function u(r). If the integral:
U = int (rho * u(r) * 4 pi r^2 dr) between 0 and inf
converges, that interaction is considered short-range. Otherwise, it is a long-range interaction. If you consider u(r) ~ 1/r^n, then for n > 3 we have short-range interactions. So, the Coulomb interaction with n = 1 is obviously long-range, and if you calculate it in a simulation the same as a short-range interaction like Lennard-Jones, you get wrong results. There are special techniques for numerically integrating Coulomb interactions over the infinite periodic images of the simulation box, among which Ewald summation and Particle-Particle Particle-Mesh methods worth mentioning. LAMMPS can enable the use of such methods, if you use the keyword long, and specify the parameters of the long-range handling method. Or, you can accept the error and tell it to cut Coulomb interactions as it would any other interactions with the keyword cut.
I catch underlying reason of coul/long and coul/cut, however, both keywords need a cut-off value to be specified in LAMMPS. http://lammps.sandia.gov/doc/pair_lj.html
Why coul/long needs such parameter? As you noticed me, this poses an approximation which is incorporated by coul/cut keyword.
Usually, the methods for long-range integration of electrostatic interactions are composed of two parts, which are summation of forces in real space, and summation of the effects of periodic boxes in Fourier space (the k-space in Ewald summation notation). So, you can specify the cut-off for real space interactions in this LAMMPS command. Then you need to specify the parameters of k-space summation separately. Quoting from LAMMPS documentation:
"The Coulombic cutoff specified for this style means that pairwise interactions within this distance are computed directly; interactions outside that distance are computed in reciprocal space."