Compare the retained result's attempt identity with durable processed history, including what survives the monitor and producer restarting. A high level alone cannot tell you it is a new completion. Keep unassignable results for review rather than clearing them or sending the work again.
That is one concrete persistence gap. Check the reference's lifetime and reuse rules with controls before using it as a permanent key, then test repeated delivery and reopening against the stored history and displayed count. Moving the set to disk alone does not prove every result identity is unique.