I've already made custom LOD and physics models as I do that for all of my work - I rather enjoy it as it's my way of trying to beat/outdo the initial Li estimate.
The problem is that I don't think the LODs know what to do with the linked objects as, more often than not, it'll just use the parent where-as the model itself can differentiate between the two.
I've even tried it with a simple cube. I made one cube and made it the child to another. Then I made the LOD for that cube. The problem is that the LOD, in the SL engine, only recognises the parent and ignores the child. It seems incapable of accepting that there are supposed to be two objects. I think it's because you're unable to actually change the way the LOD looks in the engine itself so it HAS to upload a singular model.
The problem there is that the main model needs to be a parent and child - a link-set but LODs don't seem to work that way. As I said I work in a modular way for the things that I create so I suppose I'll just have to manually position things without the pinpoint accuracy... I like to work with precision when making things like this. Don't even get me started on how this would work for animated parts.
It works if I use the "generate" method but I don't like to use that. When I use my custom LOD it just chooses the parent by default. I've done research on it and tried different things but nothing works. Naming things _LOD2, _LOD1, _LOD0, etc. Nothing seems to work because then, all of a sudden, the LOD system gets confused between the second object and the actual LOD itself.
I guess I'll just have to do it the old fashioned way.