Unfortunately this approach doesn't feel that great down here in Australia, definitely a function of latency.
I think you could get a lot closer by framing this as an optimization problem, where you use the full alphabet dictionary, but add a residual prediction which aims to cover as much of the remaining domain name tree as possible weighted by popularity. This tree could then be pre-baked and stored with the same system. This would probably get you p99 0ms even in Australia.
The perceived latency starts from keydown, not keyup. Redefining latency to start at keyup reduces measured latency, not perceived latency, and delaying the visual display to keyup makes perceived latency strictly worse, not better. Even sticking with the keyup definition, just displaying the result as soon as it is available gives the possibility of negative (defined) latency.
I think you could get a lot closer by framing this as an optimization problem, where you use the full alphabet dictionary, but add a residual prediction which aims to cover as much of the remaining domain name tree as possible weighted by popularity. This tree could then be pre-baked and stored with the same system. This would probably get you p99 0ms even in Australia.