Reactors can keep their own state including their current position in the event stream. When a replay is initiated it ignores events older than it's current "head."
What happens when the write of the current position fails?
It's the same problem as presuming that ACK messages in message brokers/queues are guaranteed to not fail.
Since the message transport and other durable resources are rarely able to be enlisted in the same atomic transaction, and since distributed transactions would largely be an antipattern, it would seem that a reactor that records its current position can't be presumed to be an infallible way of ensuring that messages aren't processed more than once.
In the end, it always comes back to ensuring that handlers are idempotent, having something that can be used as a stable and consistent idempotence key, and accepting that the idempotence logic is the responsibility of the handler coders rather than something we can count on generalized infrastructure for.
