Fix random test failure introduced by #4385

[randombit]

Fixes #4415

[coveralls]

coverage: 91.078% (+0.001%) from 91.077%
when pulling 914e969 on jack/fix-4415
into dc326ed on master.

[reneme]

I'm a bit uneasy about the recursion, frankly. Alternative would be wrapping this into for(size_t attempts = 10; attempts > 0; --attempts) {} with a continue instead of the recursive invocation and a "graceful" abort in case something goes south. But I'm guessing this is unlikely enough so that we don't end in long chain of recursive calls.

[randombit]

Another simple response would be if we find ourselves with x == 0 return a completely random scalar instead.