when I modified it to read results back in a
"read until end of file" loop.

The bug was that instead of trying to read 1 KiB chunks of data
from the cta front-end, the chunk size was only 8 bytes (the size
of 64-bit pointer).  The calculation of the chunk size was
incorrectly:

   sizeof(buf - 1)

instead of:

   sizeof(buf) - 1