This is 100% true. A coworker of mine implemented an elevation-bitmap-to-3d-model conversion tool in 160 lines of Awk. It ran faster than our "good" Matlab tool by a factor of 10.
Awk (or Perl) doubles the usefulness of Unix. Most of the common commands in Unix are query commands. When you need to start manipulating queried data, Awk is where the rubber meets the road. Piping data through the shell stops being read-only, and becomes interactive.
Could you give a bit more details there? I don't have any experience with matlab, but I tend to think of awk as fast to write code in (and start up), though not particularly fast in execution. (Roughly on par with Python, i.e., usually good enough.)
Also: "I have since found large datasets where mawk is buggy and gives the wrong result. nawk seems safe." makes me uneasy, as does the fact that it was unmaintained for a while.
Still, the One True Awk still has my favorite opening line in its "b.c" source file:
/* lasciate ogne speranza, voi ch'intrate. */Interestingly enough, windows powershell structures its cmdlets in the same way. Makes lots of sense for stream processing as you said.
If the unit of input in this kind of stream processing system doesn't match the problem domain exactly, things get very difficult very quickly.