I am almost as far as you can get from an expert (and I am sure there is one here who can explain it better and hopefully correct me) but:

For example the TNP issue derives from OCCT (or something in the stack close to it, I am not exactly sure) not really handling face naming at all.

So if you want to avoid topological naming issues (which is a hard problem in CAD), you apparently have to do some work to track before and after and reconstruct your face naming from either side of the OCCT black box.

https://wiki.freecad.org/Topological_naming_problem

https://forum.freecad.org/viewtopic.php?t=27278

Then there are various fairly entrenched issues to do with filleting and chamfering. Basically, both these operations will fail if a chamfer or fillet would completely consume an existing edge. It also sometimes creates impossible objects when filleting, or used to.

Booleans can be slow.

And more generally, it seems if you track the FreeCAD project that OCCT can be inscrutable when things fail; error messages aren't the greatest etc.

The flip side of OpenCascade is that it seems to be highly portable and has for example been compiled to JS with Emscripten for this astonishing thing:

https://zalo.github.io/CascadeStudio/

It's a monumental open source project, for sure, and it's definitely not nothing that we have an open source CAD kernel; these are projects that perhaps have to extend beyond the working life of an individual developer if they are to be stable. And there are loads of projects built around it. So it's absolutely consequential and we're lucky to have it.

Truck and Fornjot are incomplete and not quite ready for prime-time. libfive is FREP, not BREP.

OCCT is the useable of all. But it also has old architecture, all sorts of imperfections (some of them described by the other guy here), and the code quality isn't great, I'm told (by much more experienced people).