Your first two points both tie into improvements in (d)encryption speed, and that is a fair point. But even with the current gen 32bit cores VPN transmission and disk encryption on mobile devices is already done in real time with no noticeable impact. And some mobile SOCs may have a separate ASIC that does the (d)encryption anyways.
I'm very skeptical as to your third point. Please elaborate on how 64bit general compute improves on recognition accuracy? It doesn't. At best it may speed things up a bit, but again not appreciably
As for your last point, there will only be a boost to FP calculation speed if you use 64 bit variables. But as I said in my original post you wouldn't want to unless it was a STEM based task and then you wouldn't be doing it on your mobile device anyways. And unless you really needed the extra precision of 64bit, which you usually don't, using 64 bit variables is actually a detriment because best case your double precision FLOPS are 1/2 of what your single precision FLOPS would be which means you are wasting compute.
I'm not saying that going 64 bit in a mobile chip is bad or worthless, in fact its not, what I'm saying is that its not as big a deal as Apple made it sound. For example, if the A7 was exactly the same except the cores were 32 bit instead of 64 bit, all the numbers the quoted for improvement over the last one would still be the same. The performance improvements for the A7 over the A6 had very little to do with the 64 bit general compute processors.