Another angle is that directives meant for external tools with truly
pathological syntax can still be encoded as a Scheme comment. If #! is
followed by datums, we can uniformly parse the relatively sane magic
comments which are essentially sequences of symbols, strings, and
numbers, maybe with lists for nesting. And we can re-use Scheme's `read`
for the parsing so it's simple to implement and tractable to understand.

If we support arbitrary strings, schemers will resort to writing ad hoc
parsers the way other language and tool communities now do. But if we
can parse the most common magic comments like ordinary Scheme datums,
some of the symbol names will look a bit funny, but otherwise those
directives mesh well with ordinary Scheme code. They are less isolated
into a separate and incompatible sub-syntax within the main syntax.