It seems the buildfarm has been failing not only on Win32, but also on other machines! The fact is that on Win32 the failure is only much more probable. So probable, in fact, it seems to happen everytime.
The failing test is the stats test, which tries to measure whether the stat system is actually counting operations on the tables. Initially it was thought that autovacuum was the cause, but some investigation suggests that not to be the case.
Magnus has been helping me pinpoint the problem. The first thing we tried was to have autovacuum use a "reasonable" setting for vacuum_cost_delay. "Give it 10 milliseconds," I told him, confident that such a low setting was enough to cause the scheduler to let the stats system to run and thus increment the counters. The theory was that autovacuum being enabled caused stats not to have time to run in the 2 seconds that the test sleeps.
It didn't work though, so he raised it to 100ms and then 1000ms, to no avail. The test still failed.
Next, he raised the 2 seconds sleep to 10 seconds. It didn't work either. So he turned autovacuum off, and reran the test. Guess what? The test still failed!!
Then he checked the test manually, and it turned out that the pgstat views show the table to always have counters on 0!
We haven't been able to pinpoint the exact cause, but now it's looking like the autovacuum change wasn't the culprit; maybe it was the autovacuum change plus something else. We're not sure.
Still investigating ...