p100 would include those where the uplink is so slow that the page takes minutes to render completely. Those users are useful to see, but the browser is much less of a bottleneck for them, and the data should be broken out into a separate bin.

It would also include users who happened to close their laptop lids and suspend for an hour, right during pageload.

That's if it's measured by wallclock time, but they should be using a monotonic timer that ignores sleep and clock changes.