Good idea
"I did a quick survey of the make of machines in use in the cafe. Two out of the three times Apple MacBooks came out at over 50% share, the other time at 40% share. Total sample size is now probably about 50 machines, so statistically significant.
What does this say?"
Absolutely nothing. If I go to a redneck beer-drinking-and-fishing hangout, I can "prove" that 90% of the population drives Ford pickup trucks.
-------------------
That said, moving to an ARM would be a smart move. They have superior power use (something like 1 watt at 1.2ghz), the modern designs do have a pretty strong FPU, they have "Neon" instruction set and a DSP which do virtually all video decoding work (it can decode videos using about 30mhz of processing power, and that is without using any video decoding support of the video chip.) A lot of OSX is heavily multithreaded, it'd be real easy to stick 8 or more dual-core ARM CPUs into a notebook, and still use less than the power budget of the existing Intel processor.
"If you recompile OSX for ARM64 and you keep the APIs identical, why would you need to emulate anything for additional software?"
The CPU must be emulated, a program is not just a string of API calls. CPU emulation technology is good but this can still be a significant slowdown. Of course, you are right though, none of the other hardware has to be emulated a bit. (Although I haven't tried it...) qemu for Linux can do exactly this -- for instance, put some Linux for Intel libraries and binaries on an ARM, and run them on the ARM... most system calls are the same, and qemu can "convert" a few platform-specific ones (usually different for historical reasons).
OSX supports "FAT" binaries (inherited from NextStep) -- these are binaries that contain, potentially, SPARC, Motorola 68K, PA-RISC, Motorola 68K, PowerPC, Intel, and (probably already) ARM code in a single binary (probably the Mach-O file format already has ARM support, since iOS is a mutant OSX).
So, if Apple flubs it you'll end up with these Macs that "can" support native code, but actually are constantly running everything under emulation, not getting the performance it should. If they do moderately well, by the time the ARM machiunes ship they can at least make sure video playback support and Apple's own apps use NEON and any available DSP (ARM video acceleration for instance, very effective)) *and* supports whatever video acceleration the video chip supports. If they get it right, they will release ASAP XCode, gcc, etc. that support Intel + ARM "FAT" binaries, as app makers rebuild the apps they naturally gain ARM support*, so by the time the ARM machines ship a fair amount of software can be ARM native.
*I've run Linux distros on PA-RISC, PowerPC (Mac) *and* an IBM POWER system, DEC Alpha, as well as Intel, I have a command-line only Debian install on my (ARM) phone too. None of these feel like a stripped down port, portable code is portable. I think most apps will be similar going from Intel OSX to ARM OSX, it'll just be a matter of hitting "build project" or whatever, not some troublesome porting process.