The problem with using a lower LOD is that you get invisible walls/flythrough surfaces with many of the LODs currently in existence (I know Wanderer tested that out earlier this week too).
And this is basically what I said to Wanderer in his PM. Because of the geometry differences you couldn't use anything but LOD1 and even then only if it were a very well designed model and LOD setup. It's not just the invisible surfaces problem though, it's things like decals, particle effects, laser hits, etc. Collision detection not only tells you whether something hit but where it hit as well and those things are more interested in the "where". But really, if all of the other things in the list were taken care of, the only models that would really need a separate collision model would be ones in the 40,000+ range. Things like really large and detailed ships, planets, or terrain in other words.
If you take #2 to be converting the BSP data into an efficient octree for collision checking then you can end up cutting down on the number of verts to check to almost nothing in comparison. Right now the collision code checks the bounding box and if that scores a hit then it will check each vert until it figures out the closest possible hit. With an octree you can cut that down to just looking at a small group of verts that you know for sure is going to have a hit on it. So this means that a 10,000 vert model may be able to pass through the collision detection code only having had to check about 100 verts. That equals out to a massive time savings. Combining that with scene management and using index buffers you end up with a haul-ass collision detection check.
And that is also where this gets on topic. That octree setup will need to be done when a model is loaded, further increasing mission load times. But most people seem to overlook model loading as a part of the code which could offer some of the most benefit from being multi-threaded. It is largely contained so it should be one of the easiest parts of the code to convert (not that it would actually be "easy" though), bitmap loading is postponed until later on so OpenGL context issues should never be a problem, and it can take up over 90% of the load time if IBX files need to be generated (about 40% otherwise). That way you get to load a model in the background and still have a responsive UI in the foreground. It offers a speed advantage for multi-core users and better user experience for single core users. In other words: win-win.
In retrospect, it might have been a good idea to suggest people used detail boxes more heavily and do away with the standard LOD system. I know that's impossible, it's just conjecture, but that way you could have divided out just the 'hull' of the ship for ship collision detection, and ignored the details.
I'm not sure that the collision detection code even knows anything about detail boxes actually. Even if it did it doesn't change the fact that it gets walked during collision setup anyway. It would offer some benefit, but in general not really enough to make any difference.