Mex does not accept -g?

Hello, I have suddenly collided with a rather frightening problem. I have a code that is about an even split between a compiled/gpu-accelerated core and matlab high level stuff. The flow to creating the mex files goes as (ignoring most of the compiler options ofc):

nvcc -g -cuda foo1.cu % creates cuda fatbins in .cpp files
g++ -fPIC -g -c foo1.cpp % compile cpp to object code
(...)
mex -g -c interface.cpp % compile interface
mex -cxx foo1.o foo2.o fooN.o interface.o -output bar % link

And this rewards me with:

'/home/erik-k/code/bar.mexa64' is not a MEX file. For more information, see <a href="matlab: helpview([docroot '/matlab/helptargets.map'],'ENOTMEX')">File is not a MEX file</a>.

make: *** [Makefile:77: bar] Error 255

After some frustration and bisection, I eventually found that removing -g from the interface.cpp compile "fixes" the problem. The biggest thing I cannot understand is, the output from nm shows that the mexFunction symbol IS PRESENT in the "bad" mex file. Does -g break the little hidden "version info" object file that mex makes/links with the input code?

I noticed this began after I upgraded from matlab 2018b to 2019b.