Okay. I seem to have localized the problem. There's also a cure at hand.

What has happened is the following: Since I usually use 32^3 grid points per refinement level in octant mode, information from all coarse grid points is needed for prolongation into the refined region (when using full Berger-Oliger; which I am). So in the run in question, the simulation screwed up on the coarsest grid and the faulty data got prolongated onto finer grids and spread there. BANG.

The only way to cure this is to use bigger grids. 40 zones in one direction in octant, 80 in bitant mode.