I think a common-sense approach is to cite the original paper (whatever the language) in order to give credit and attribution but only rely on arguments from papers you can understand in your proofs (so you don't violate the golden rule). 

Regarding reviewers, the worst that can happen (I think) is that you use a crucial argument from a paper you can understand but that the reviewer cannot understand. In that case, I think the problem is the same whether the reviewer cannot understand it because he is unfamiliar with the math or because he is unfamiliar with the natural language. In both cases, you, as  the author, should try to present relatively clear references, and that includes translations when appropriate I guess, but ultimately this is a failure of the reviewer. If I were reviewing a paper and found myself in this situation, I would politely ask the author if there is a translation available. If not, I would tell the editors I am not competent, but wouldn't blame the author.

It is a bad idea to upset reviewers, but banning reference in languages other than English (or any other language) even for attribution purpose is an outrageous suggestion that should not be complied with.