Huh, it's interesting that the final test set gives results which significantly differ from results of the sample tests attached to the task (it looks like this is the case for some other submissions as well, not only mine).

Which is even more interesting is the fact that real time improvement is significantly lower than improvement for cpu time - on sample tests, I had these two values almost identical locally. And, moreover, locally they became closer to each other compared to baseline, because there was decrease in sys time. Maybe final tests have some specific which makes newly re-written arena allocator ineffective (which might increase sys time, and, therefore, overall real time).