As suggested, I'm filing this in a new bug report for wider review.

The updated patch attached.

The previous discussion was in bug#64735, and the benchmarks (which more 
or less hold for the new patch as well) can be viewed here (last table): 
https://debbugs.gnu.org/cgi/bugreport.cgi?bug=64735#506. Except the new 
version somehow performs a little better at read-process-output-max=4096 
as well, despite seemingly doing more.

Let me know if I DRYed this too much, or if there are better names for 
the extracted common routines, or etc. Or for the new variable 
(read-process-output-fast).