Yeah, some failures are not detected as failures. When failures happen instantly, you get impossibly low timings.

Please feel free to play around with the source code and see if you can improve it. I'd be glad to merge your diff.