>> I can certainly wrap Gauche's low-level syscall API with guard and
>> translate <system-error> to srfi-170 error, in every srfi-170 API. The
>> overhead isn't small and I'd rather avoid it if I can.
>
> Loko Scheme's implementation of SRFI 198 uses the R6RS condition system.
> This way there is no need to use guard, because any syscall error
> anywhere in Loko already raises an R6RS condition and SRFI 198 just
> needs to recognize it. Maybe you can use a similar strategy in Gauche,
> can you teach SRFI 198 to understand <system-error>?
>
> In Loko, the foreign-error:* procedures understand R6RS conditions and
> check their argument to see which kind of condition they've been given.
> make-foreign-error translates its argument into conditions. This also
> has the benefit of letting any R6RS code handle the &who, &message and
> &irritants part of the SRFI 198 error.
>
> It works because SRFI 198 defines a constructor and accessors that hide
> the concrete type of conditions. I haven't been keeping up with the
> discussion lately, but I hope that this aspect hasn't changed.

Great! This is exactly the kind of implementation I was hoping the
abstract data type would enable. I recommend the same for Gauche as
well. I also did a Common Lisp port using the same principle.

If needed, an ADT can also have multiple implementations, e.g. one based
on a native condition type and another based on a record type. The
`foreign-status?` predicate and `foreign-status-ref` accessor will just
have to be written to recognize both. Gauche has generic functions that
can handle this if needed.