Regardless of what seems to happen in practice, none of the BigUint operations (including comparison) are guaranteed to work in constant time. Since the implementation uses a Vec whose size depends on the size of the integer, this could easily have significant timing differences in other situations.
I've tested a scalar of about 50% 1s against a scalar with almost no 1s, measuring at the nanosecond resolution, with 100 samples each.
The p-value of the 2-sample T-test was greater than 1%--there's no evidence that one takes less time than the other. That can be replicated by playing with the benchmarks yourself. If you have any evidence to the contrary, I'd be happy to look at it.
