As excited as I am about this jump from the fuzzy NeRFs/gaussian splatting to real meshes, I'm not holding my breath for BREP generation. Mesh to BREP has always been fraught because for anything beyond "find the cylinders", it becomes really subjective what a good representation is, and your average mesh likely doesn't have any simple representation that captures the full organic shape with analytic definitions.
With mesh faces now supported in BREP, I'm more optimistic about a mixed modeling approach, where you can do the braindead find-the-cylinders conversion but keep the rest mesh, not needing to force it into some eldritch contortion of BSurfs.
I'm sympathetic to using mesh approaches, hence the last part of my comment focusing more on the physics, etc. vs. requiring BREP.
The main advantages of BREP are:
(1) You capture design intent because of the explicit dimensioning and constraint model, which of course is still not used enough and 2D drawings are useful.
(1a) This intent is often needed (even if just implicitly) during the manufacturing process as machinists, toolmakers, etc. + their software (CAM, G-Code, etc.) convert them into physical parts.
(2) They are human understandable and editable.
(3) The legacy ecosystem uses old CAD, and it's very hard to switch - ie Boeing isn't putting 777 in a new CAD unless it has some massive advantage.
So having BREP, or perhaps a mixed approach like you suggest with the feel of BREP (feature tree, direct modeling, etc.) approach would ease the transition.
With mesh faces now supported in BREP, I'm more optimistic about a mixed modeling approach, where you can do the braindead find-the-cylinders conversion but keep the rest mesh, not needing to force it into some eldritch contortion of BSurfs.