At first this appeared to be a race condition in sytest, but I don't think it is.

There are a few integration tests that sometimes fail; e.g.

Running tests/52user-directory/01public.pl...
...
  Testing if: Users appear/disappear from directory when history_visibility are changed... 
   | User initially not in directory... OK
   | User appears in directory after history_visibility set to public... OK
   +--- FAIL:
 | Timed out waiting for test at ./run-tests.pl line 542.
 +----------------------
...

In current sytest/develop this appears to be a race when running it, because it's flaky and variously comes and goes, but adding a try_repeat_until_success loop to it just means it either succeeds quickly or times out and fails every attempt. So it feels more like it sometimes succeeds or fails within the synapse itself.