A durable operation record is shared by concurrent retries. The design challenge is preventing a later timeout observation from downgrading an already committed success, while still allowing a timeout to be upgraded if the operation eventually completes. The key unknown is the exact storage and transaction mechanism, so the guidance should remain implementation-neutral and emphasize atomic monotonic settlement.