I looked into CSG a couple of years ago as I wanted to write my own implementation. In retrospect, in terms of time, I would've been far better off using a library. However, after doing a fair amount of research, the simplest approach turned out to be the most stable in terms of cost and flexibility; most importantly, it was the most suitable for me. The paper on the Laidlaw approach can be found here.
Mind you that debugging CSG code takes a lot of time and testing - you can't just assume that when you try one boundary case (eg vertex-edge) it'll work if you swap the meshes. Also, due to raycasting, Laidlaw's approach limits you to closed meshes (which is what you generally want, but not always). As far as I remember the paper has a small mistake in it, but I can't recall what it was. In any case, be ready to have to write plenty of additional code from a custom memory manager to your own triangulator (in the very least you'll need to perform convex primitive decomposition) and merger (to re-create mesh faces after CSG), which in and of themselves will take lot of time and testing. I found writing the merger probably the hardest to do so it wouldn't be dead slow, seeing as you're going to want to minimized overlapping edges and you'll be working with annoyances like T-junctions. The triangulator can be really tricky as well (I ended up implementing the slowest, but simplest approach - ear clipping) as you'll need to manage things like vertex ordering, repeated/collinear vertices, etc.
Overall, if you really want to implement your own CSG routine, be really methodical about it from get-go: index your vertices, pre-decompose everything into triangles for speed and thoroughly test everything at every step. The sad thing about the CSG procedure itself is that you can't really test it until you've written the bulk of it, which in Laidlaw's case is a lot of if statements, which need to be painstakingly correct.
If you're not absolutely bent on writing your own code, I heartily suggest you make use of a library. Ironically, since I was stubborn to write my own, I never actually compared all the different libraries out there...
PS - I never put enough effort into getting this working properly, but once you've done your initial decomposition, Laidlaw's approach is separable and can be extended to two threads (one per mesh), which I find to be of great value compared to other methods. Do not go with the BSP approach - it'll be dead slow as soon as the meshes have any decent number of polygons.
I ended up getting my code working at real-time framerates for box-cylinder (20 sides) intersection. I'm sure it can be further optimized to get real-time performance out of 40 + 40 tri meshes.
I never did this, but overall I would consider three ways:
1) create a number of outlines of various shapes in any 3D modeling software that form a single whole. When fragmenting, simply run CSG on the object. I'm not sure how to avoid things like a chair's legs ending up in a single fragment, though (you'd probably need to scale the fragment meshes to best fit the size of the fragmented object). You could pre-rotate the fracture meshes to create fake variations of it.
2) pre-fracture the mesh manually (in most cases the fracturing is an offline process anyway)
3) consider a voxelization approach: voxelize the object when loading it and split off arbitrary "chunks" as needed; I'm not sure how to perform hole capping right off the bat, though
4) I wanted to find a link I saw a couple of years back regarding a pretty cool destructible material system, but couldn't find it. It was closed and still in beta AFAIR, but it handled stuff like wood chipping, etc. I did come across this, however (haven't read it yet, but hopefully it helps)