I saw failures for this test case when introducing new logic to generate collated indexes in PBTs. This was a false failure; it didn’t change query behavior, but it changed the queries generated because it changed how the random bits from the fixed seed were being consumed. The new queries matched fewer documents on average, causing the failure.

We should make this test less fragile. One way to do that is to split the assertions being made in the test into new test files, and increase the number of repetitions in each case. This would hopefully allow for a more stable average, decreasing the chance that an “unrelated” change in the PBT model will cause failures in the test