But it isn't just the buffer size of the pipe (which defines how much can be sent atomically); I'm assuming that the amount of data is so much that the kernel exhausted some resource trying to buffer it all.

You're right, I merged the concepts, even if some lurking inside the 2.4.x Linux kernel convinced me that Linux considers them to be the same, as far as I can remember.

Anyway, the fractioned approach you suggest should be used more generally - testing if write/syswrite actually wrote all requested characters should become an habit much like testing if open was successful.