Yes, the 95th percentile means that only 5% exceed the specified size, but day by day this figure fluctuates. Regarding fee optimization, I tried manually setting fees on all channels to approximately 800 ppm, and then progressively, over many days, lowering them and manually recording the number of successful payments, the total payment amounts, and so on in a spreadsheet. When you look at the results over a short period, it seems like you've found some optimal fee, but as more time passes, you realize that all these fluctuations have some other internal nature.

I tried to think about how optimization could be done, but it seems to me that this task is impossible because full optimization would require knowing the complete network graph with all node balances, which is fundamentally impossible. Moreover, this situation is constantly changing. Without knowing the balances of all nodes at a given moment, it's impossible to accurately predict which node will be optimal at any given time, which fee will be optimal, and so on.

Furthermore, payment paths and fees are set by the sender. The sender draws their information from the network graph, which always differs for each observer due to the slow propagation of gossip information.