This may be true, but neither FA or MTP are lossy (aside from some possible CUDA kernel noise in FA)

MTP when done correctly is lossless. As you note: it is a draft. That draft is still checked by the original model. Drafted tokens are not accepted if the original model would not have selected them.

If you are having issues, something else is wrong. But in any case, you lack the domain knowledge to be posting authoritatively about it, and I suggest you read deeper into how these things work first.