turns out there was a problem with signed values after all
failing testcases with k=5 produce the indices of positive values correctly, but negative ones don't seem to be compared correctly (always producing index 0?)