LOD is certainly a very tricky problem with Marching Cubes meshes, and I don't have any great answers. Think carefully about how much you need it because it has cost me quite some time in the past. But I can give a couple of tips which I found when working with it in Cubiquity (
image here).
Firstly, I haven't done any stitching. Obviously it's the 'correct' and watertight solution but it is fairly complex. Instead I've applied a very simple hack, in that I use a different threshold value when extracting the lower LODs as compared to the upper LODs. When applied on a smooth density field this causes the lower LOD meshes to be slightly eroded and therefore slightly smaller than the higher LOD meshes. The crack still exists but it is always hidden behind the higher LOD.
This approach can of course cause visual discontinuities, but in practice I haven't seen too many problems. I haven't applied advanced lighting though and I think this might highlight the discontinuities.
Secondly, I think there is the question of how the downsampling is performed. I not sure what the TVA paper does but a common approach is just to skip ever other voxel, or perhaps average groups of eight voxels into a single one. I do not think these are mathematically sound approaches. There is a nice page
here which talks about more mathematically correct approaches, and I think applying these might reduce the need for stitching of offseting. I think it will be a few months before I really try it though.
As for your actual question, do keep in mind that geometry can disappear completely in lower LODs (e.g. a long thin structure may no longer be visible). I feel this may affect your suggestion of adjusting the meshes afterwards.