The ghost and buffer zones are also used for time interpolation, but since you suggest to only evaluate the constraints when the coarse level is active, you are circumventing this.
The largest issue I see is that e.g. a regular L2 norm over the simulation domain is not very useful. It emphasizes very much the coarse grid, which is very large, and where the boundary conditions might do strange things to the constraints. Similarly, you will need to be careful to exclude the interior of the horizons (see CarpetMask) to avoid looking near the punctures. CarpetMask is not smooth in time, so looking at the time evolution of the constraint norm is then also problematic.
You might be more happy if you look e.g. at the "isum" or "inorm2" output. This output is not weighted by coordinate volume, but by grid cells, so that the coarse grid is deemphasized.
-erik