In practice, I have spent an entire week during a universitary course with various problems from such problem site and eventually learned that the problem descriptios were different in a subtle way or certain edge cases were not considered or the test suites were flat out wrong. (This was particularly amusing, because they had added wrong test cases after certain solutions were accepted beforehand. Thus, It was clearly visible that the problem was solveable, because there were programs accepted).
Also I don't entirely agree with the argument that my program has to work in an unknown environment with unknown controls. In such a situation, I would at least get a logfile to see what was going wrong. Without that, debugging anything non-trivial is just not worth it without getting paid for it.
Anyway, we're open to suggestions on how to improve the process. :)
For one, even if this is the actual test input you use, I will not know it.
For two, it might not contain the trigger for my errors, but if I can get enough such samples, I should eventually get a sample which contains the problematic input.
For three, from what I see, the problems should be possible to generate randomly, especially if the certified solver rejects malformed input. E.g. for the missle problem, throw together N numbers and look what happens until the certified solver outputs a proper solution.
Given this, this system could be extended using a rating-system by users ("Click here if this input helped you solve the issue"). This would overall result in a bunch of inputs being generated (basically whenever someone needs a new input), and some edge case inputs will overall bubble to the top, because they are rated helpful more often than others.
Anyway, I'm happy to help. Good luck :)