Rather than pure resolution, I'd say the main issue is mesh quality and consistency of coverage.
DEM provider use algoritms to eliminate outlier measurements. A too strong averaging algorithm tends to round up terrain specific features, while a too weak process leaves erroneous spikes in the DEM.
As a result, the higher level you compile, the more unwanted spikes you may get in your mesh.
Even commercial products like Intermap's Next have such issues, like an ugly spike instead of the Matterhorn. Intermap support replied they were quite aware, but since so few people complained so far, they were not going to handle the bug (sic).
It may read a paradox, but a high resolution mesh is usually much better for rather flat landscapes than for mountainous areas.
The very high res DEM may outline very specific features in an otherwise rather flat terrain, like for instance depressed trenches of roads/railways, hills, etc. that greatly enhances the countryside of a place like Belgium or northern France, while a standard 19m mesh leaves everything flat.
High resolution DEM tends to make mountains exaggeratedly spiky, and in some cases causes erroneous artefacts that disappear when reducing the compilation level, a compromise has thus to be found.
Masks are also a touchy topic in AFS2 and I always try to avoid them when I can do without. I already happened to notice strong border effects when compiling a too small area involving masks.
If your transition high res to low res is compiled out of the very same source (like the Level 12 to Level 9 interface in my example above), then you shouldn't get a cliff.
Cheers
Antoine