Using the command line provided in elastic#81983 I could not reproduce the issue.
Anyway, changing the timeout from 5 seconds to just 1 second I could
reproduce the issue. At the end, keeping a timeout of 1 second I could fix
the issue moving the refresh operation just before running the match_all
query. My understanding is that the timeout is used to give the rollup job
enough time to complete the rollup operation. Anyway, other than that, after
the rollup job has written data into the rollup index we need to refresh it
in order for the subsequent search operation to hit the expected rolledup
document.

Instead of refreshing from the yaml test we refresh once the
indexing operation is complete in the callback.

This reverts commit 7968c70.

The client is used to call the refresh action too. As a result, using
'any()' as the action would match both the search and refresh actions
while we just need to handle the search action.

We also need to mock the refresh response for the client
to actually propagate the result of the refresh call.