My answers below, having in mind that peak detection is error prone. For ypool it's easy to measure quantity of miners or shares per second. But it's very hard for the network to measure the block rate with fast response time. The measurements done right now take the last 12hs of blocks, meaning it will get the average of the high/low peaks. If you adjust faster, the adjustments get more erratic because variance plays a higher role in the measurements.
what good is getting new miners when the superblock can't even make the full use hashpower at its peak (nevermind the records breaking)?
Getting new miners is always good because it improves the security of the network. The superblock may not always come at the hashrate's peak but it will, eventually. Since it cannot be predicted in time, statistically it should drift until it hits the peaks again.
Would you agree improving the odd of finding record breakers by having hashpower used at least by 95% is better than 35% all the time?
Yes, but if it implies the risk of overestimating hashpower and leaving the network without blocks for a long time, then no. So, no.
Or do you prefer the current 35% level to climb up at steady rate to reach the current 95% level sometime later and then ask yourself why is it records breaking is so technically tough?
Yes, I'd like that because it would mean we have more miners, and when the time of the superblocks drifts to the peaks we would have much larger records.