Benefits of SDFs over the standard Boundary Representation (used in Freecad and similar) are that you can do "pattern" operations with domain repetition, which means making N copies of a feature is O(1), vs O(N^2), you can deform objects with domain deformation, which means if you have a closed-form representation of how you want to deform space you can basically directly apply that to your object, procedural surface texturing is easy, CSG operations are easy.
The big drawback is that it is hard to provide any workflow based around "selecting" faces, edges, or vertices, because you don't naturally have any representation for these things, they are emergent from the model's SDF.
I have some blog posts on my progress: https://incoherency.co.uk/blog/stories/sdf-thoughts.html and https://incoherency.co.uk/blog/stories/frep-cad-building-blo...
I am solving the "selecting faces" problem by having the SDF propagate surface ids as well as distances. So the result of the evaluation is not just the distance to the nearest point on the surface, but the id of the specific surface that is nearest.
My next big frontier is reliably providing fillets and chamfers between arbitrary surfaces. I have a handful of partial solutions but nothing complete yet.
The most promising idea is one that o3 came up with called "masked clones", the idea is roughly to make a clone of the 2 surfaces you want to blend, mask them by intersecting with an object that is like a "pipe" along the intersection of the 2 surfaces, apply the blend within the pipe, and then add this "blend pipe" as another child of the lowest common ancestor of the 2 blended surfaces.
And after that I need to work on more standard CAD stuff like constraint solving in the 2d sketch editor.