I attempted to do this in a manner that preserved the line numbers to reduce the
review effort on the resulting diff, but we still have to deal with the
ramifications of how a future-incompat lint behaves compared to a hard-error (in
terms of its impact on the diagnostic output).
