This test has been observed to fail on gerrit-verify jobs as well as in remote cluster tests.
From Alexander Behm:
The problem is that the remote fragment may start executing immediately, so the time difference between "Rows available" and "All fragment instances started" could be less than the time it takes for rows to be produced by the remote fragment.
The "All fragment instances started" timeline event was added relatively recently. So the existing test did make sense at some point in the past - but it is definitely broken now.
(See the full comment below.)