• [Fix] Stop #flush from running middleware twice on messages re-buffered after a failed flush (regression of #474 reintroduced by the failure/requeue path), which double-encrypted/compressed/serialized payloads and duplicated headers on retry.
• [Fix] Re-check the producer liveness under @buffer_mutex in #buffer/#buffer_many, so a producer closed between the liveness check and the append raises ProducerClosedError instead of accepting the message into a closed producer's buffer and losing it silently.
• [Fix] Restore both buffers when #flush fails before anything is dispatched (a middleware step raising, for example) instead of discarding the whole batch. Only ProduceManyError and MessageInvalidError were covered before.