Agreed. If (approximate) 1st-order information can be obtained in any way (even by *carefully* deployed finite difference), then gradient-based methods should be used. It is wrong to use DFO in such cases.
> DFOs tend to have issues past a few dozens variables.
It seems that the PRIMA git repo https://github.com/libprima/prima shows results on problems of hundreds of variables. Not sure how to interpret the results.