> have problems that do not occur with comment parsing (e.g. when using
> non-ASCII-superset multi-byte encodings such as UTF-16 or Shift-JIS)?

I had that wrong -- from skimming Wikipedia, Shift JIS, EUC-* and Big5
are almost strictly ASCII-compatible in the 0..127 range (trail bytes
that can overlap ASCII characters are almost the only exception).

Would it be problematic to read the first S-expression from a Shift JIS
/ EUC-* / Big5 coded source file assuming that it's (extended) ASCII?
Then if that form is something like:

     (declare-file
       (coding "Shift_JIS")
       ...possibly other declarations here...)

The file would be read again as Shift_JIS?