Yesterday I had a conversation with Axel on a similar point. He noted that energy discontinuities don’t affect the results of MD simulations, only minimization and diagnostics. Force discontinuities, on the other hand, cause serious issues with the realism of simulations. For the standard lj/cut potential, this is avoided by applying the cutoff either at the minimum (r_{cut}=2^{1/6}\sigma) or at a much larger value where the force discontinuity is small, like 3\sigma. It seems that lj/smooth wants to guarantee a smooth force through both cutoffs, which can be useful in some situations.