Thanks for that patch: it's a good move forward for i18n. Some suggestions:
* Today I fixed the bug with "%%" and the 'error' function, so there's no need
for a FIXME or a workaround any more.
* In strings.texi, reorder the format spec description so that it matches the
textual order of a format spec. This should lessen confusion.
* Allow field numbers in a %% spec. All other components of a format spec are
allowed in %%, so odd to report an error for just field numbers.
The reason I banned that initially is that the behavior for the case "%1$% %d" is confusing: will the %d take argument 1 or 2? (We should ban such mixing instead, see below.)
* There is no need for a special diagnostic for field numbers greater than
PTRDIFF_MAX. Just use the same diagnostic other too-large field numbers use.
This avoids a need for an alloca.
* Reword "Invalid field number `0'" to "Invalid format field number 0" to make
it more obvious that it's a format and there's no need to quote the 0.
Proposed further patch attached (it addresses the above points), along with a
copy of your patch rebased to current master for convenience.
Thanks, feel free to push.
Two further things:
- Probably there's a bug lurking because the info[n] ought to be indexed by specification index, not argument index. Something like (format "%1$c %1$d" ?a) will probably do the wrong thing (untested).
- We should ban mixing explicit and implicit field numbers, like POSIX printf(3) does. The gain from allowing to mix is negligible, and it makes the implementation and the documentation needlessly complex.