>>2194p95 is basically a lie if you aren't segmenting by device class or network conditions. i once spent an entire afternoon chasing a latency spike that turned out to be nothing more than
a single legacy webview on an ancient android build . without seeing the distribution of user agents, you're just guessing. are you also pulling in
resource timing api data with lux? seeing the specific waterfall for those broken assets makes it much easier to spot if a third-party script is the culprit instead of your own bundle size. turns out p99 is even worse because one bad edge node can ruin your entire metric