Because of exactly what Moxie said in his post. This is a relatively common occurrence in practice. Someone gets a new device. Or uninstalls/reinstalls the WhatsApp app. Or wants to read messages on their laptop, too. And so on.
Warning everyone about this all the time leads to people becoming subconsciously blind to these notifications — even to people who should care about them. The solution taken by WhatsApp is a great compromise in this situation. Not everyone will have it on, but the odds are in favor that someone they might want to intercept messages for will. And if they can't know who has the notifications enabled and who doesn't, they run the risk of tipping their hand that they're doing it at all.
So yeah, it is actually kind of like rocket science, and I guarantee you that Moxie has spent orders of magnitude more time thinking with, dealing with, and collecting data on this kind of problem than you or I combined.
Furthermore, this is an argument via authority[1]. Of course there are experts, but even an expert should explain and discuss his rationale in the interest of sharing knowledge (which moxie is doing here) - otherwise problems like this will stay "hard" for a long time.
And Moxie has explained his rationale in this thread. Argument to authority isn't always wrong — particularly in the case where the other side has no data or theory to back up their claims. For instance, I personally only know little about the actual mechanisms behind anthropogenic climate change. What I do supports the notion. But I'd be lying if I didn't acknowledge that the most compelling argument is the absolute agreement by 99.9%+ of the actual experts in the matter.
Likewise, in the absence of any obviously compelling evidence validating GP's approach, combined with Moxie's explanation above and my own experience as a security engineer, I'm going to go with the guy with literally decades of both theoretical an operational experience here.
The entities this is designed to thwart are not going to want to risk leaving behind a trail of evidence, even if the risk is small.
It also prevents fishing expeditions, since the risk would quickly add up as more targets were added.
All that said, a one-time prompt to turn on the notifications for users that care about extra-strong security seems like a good idea to me.
Warning about unusual account activity seem to be very common these days, so why not using them here.
The way the warnings are presented as part of the chat history (a very good idea) also means they could be used after-the-fact to figure out when an account was overtaken, even if the warning was initially ignore. I figure even non-technical users would like to know that, after one of their contacts tells them their account was hacked.
Additionally, why is an ignored warning worse than a warning that is suppressed to begin with? That seems to me like a landlord that decides not to install smoke alarms because "the tenants could get used to the sound" - when most of the tenants are not even aware of the concept of "fire".
Finally, I don't find the "it's important the server doesn't know" argument not convincing. If you conclude that the vast majority of people doesn't have the warnings enabled and the costs of hitting someone with warnings is low, that would make snooping still a very low-risk activity.
Summing up, I think the very least consequence Facebook should take from this is to make the warnings in-by-default instead of off-by-default.