Filter out non-ASCII characters in automatically processed headers.
Do this in a way that minimizes the code change: keep manipulating
strings, but strip off non-ASCII characters when reading lines, which
should only remove characters in comments that we don't parse anyway.