Since I've spent the evening figuring this one out, I'd better log it here.
The starting point is that there are certain non-ASCII characters that
I tend to want included in the manpages I write -- in particular I tend
to say
[copyright "2011 Lars Hellstr\u00F6m"]
but I've also discovered that [vset] is useful for this sort of thing; after
[vset emdash \u2014]
I can use [vset emdash] in places where I want an emdash, and keep
my .man files readable on account of being pure ASCII. Whether
doctools manages to encode these characters correctly in the output
is however another matter.
Today, the issue is with nroff, where I see the ö being output in raw
utf-8 (which may be due to that being my [encoding system]).
Unfortunately, utf-8 does not seem to be among the encodings
officially supported by groff
(http://www.gnu.org/software/groff/manual/html_node/Input-Encodings.html#Input-Encodings)
-- while there is an option -Tutf8, this selects utf-8 as /output/ encoding,
but the /input/ encoding still defaults to latin-1. So should the nroff output
then be [fconfigure -encoding iso8859-1]d? No, a bit of digging reveals that
things are not quite that bad, but why the improvements are not
mentioned in
http://www.gnu.org/software/groff/manual/html_node/index.html
remains unclear to me.
It seems that, while groff 1.19 doesn't support utf-8 input, the
current 1.20 series does, using the preconv preprocessor:
http://lists.gnu.org/archive/html/groff/2009-01/msg00016.html
On the other hand, I get the impression that OS vendors are somewhat
slow to upgrade, so we'll probably have to live without that for some
time.
An alternative, which AFAICT works also in groff 1.19, and parallels
what doctools does for HTML output, is to use glyph name sequences
for the non-ASCII characters. The groff counterpart of Tcl's \uXXXX is
\[uXXXX], and I could confirm that replacing the ö in the generated
.n file by \[u00F6] gave me an ö in the rendered manpage (at least
with options -man -Tutf8 to groff 1.19.2).
The groff_char(7) manpage also seems relevant. My version of it does
say that "It is strongly suggested that only named glyphs are used for
all character representations outside of the printable 7-bit ASCII range."
Finally, I claim total ignorance in the matter of whether possible
inequivalences between nroff/troff/groff is anything to worry about
these days.