Figure out the right set of cases in `error`

[lukewagner]

Spinning out from this comment, we should discuss what the right set of variant cases are for error. Ideally, we'd be able to align with something in the HTTP spec, but I'm not sure what that is (noting that the well-defined status codes are already in a separate status-code field in responses).

As a general principle, I think it's a good idea to have separate cases only when code might want to actually branch and handle the different cases differently; to convey extra information for debugging purposes we can use the case payloads (e.g., with a string).