Eleven tests changed something that fires an unstructured task, slept a fixed interval, then asserted on whichever request happened to be last. Two problems at once: a probabilistic wait, and asserting on "the last one" rather than on the request the test is actually about.
Three of them were passing vacuously. Two assert that the reload after a
filter change carries no --cursor — but with the reload not yet landed
they were asserting against the initial request, which has no cursor
either, so they passed whatever the code did. The third asserted the
screen was not in a failed state, which is also true of a reload that
never started. Each was confirmed failing against injected faults before
being trusted.
They now poll for the request they mean, identified by its argv, and
assert on that one. A shared until helper replaces the sleeps: faster
in the common case, and deterministic rather than merely likely.
Separately, fifteen sites indexed stub.seen directly, so a regression
that produced fewer requests crashed the test process instead of failing.
They now go through #require, which reports and lets the rest of the
suite finish.
No assertion was weakened and no application code changed.