Intermittent ECPG test failures on Windows buildfarm machines

Observe the following buildfarm failures:

https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=whelk&dt=2018-03-29%2013%3A41%3A13
https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=woodlouse&dt=2018-04-18%2016%3A42%3A03
https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=dory&dt=2018-04-27%2016%3A15%3A25
https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=dory&dt=2018-04-30%2000%3A45%3A25
https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=dory&dt=2018-04-30%2002%3A00%3A25

The common feature is that a single ECPG test case emitted an empty stdout
file. There is no other indication of a problem: the corresponding stderr
output files are correct (and no, the "correct" contents of those aren't
empty), the test process exited with status zero, and there's no sign of
an issue in the postmaster log. And it's a different test case each time.

I trawled the buildfarm logs back to the beginning of 2017 and can't find
any similar failures before these. So it seems like we broke something
fairly recently, probably during March; but what? The woodlouse case is
on 9.6 not HEAD, suggesting that whatever we did wrong got back-patched,
and for sure I see no recent commits in relevant-seeming code in 9.6.

Baffled ... any ideas?

regards, tom lane