Global real-space refinement simply isn't very useful at high resolution, and you shouldn't need a large radius of convergence when the R-free is already 13%. I'm a big fan of automating everything possible, but this is one case where fixing the sidechains manually in Coot (which should do an excellent job) is by far the easiest option. Once you fix the sidechains the C-beta outliers will probably disappear.
PS. A common reason for C-beta deviations at high resolution is residues with alternate conformations being split at C-beta rather than splitting the entire residue, which allows the backbone to move to accommodate the optimal placement of the sidechain (rather than straining the geometry to get the sidechain atoms to fit the density). Always split the entire residue - the backbone is almost never rigid in disordered regions.