Ninja now output the failed output nodes right after the failed edge.

Recollecting those to classify causes of compile failures will help make decision.
Original measurement was using regex and was over a year ago.
Thus we'd like to know up-to-date and reliable info on this.