Can you explain your broken pipe situation? Are you talking about a real-world EPIPE operating on a pseudoterminal, or just a result of using a Unix socket to emulate a PTY in the tests? Usually a broken pipe is an asynchronous condition. You cannot predict exactly when it will happen without knowing the state of the other end, OS implementation, buffering, etc. It does not seem appropriate to change the _copy() loop around unless there is an real bug.

Regarding the buildbots, if I get this patch into a state that I am comfortable committing, the buildbots will are generally set to timeout in 15 or 20 minutes. See the “make buildbottest” commands in Makefile.pre.in, test.regrtest --timeout option, etc. You can also see which platforms have buildbots, and the state of them etc: <https://www.python.org/dev/buildbot/>.

I left a bunch of comments in the code review. There is a lot of useful code in there, but also a lot of stuff that is hard to follow or that I want to clean up.