zsh's antisocial behaviour was fairly promptly fixed (thanks to
Stephane Chazelas and his patience).
zle-line-init and zle-line-reset seem to be user-defined widgets
and the order inverted. Put zle-line-init before reset-prompt
because some people do weird things in there.
There's a slight issue with the inverted cell representing
the cursor standing out a bit too much amongst the rest of
the characters.
The resulting experience is a lot more consistent, though.
Since I'm already dealing with the fish shell.
All of our supported shells seem to handle cursor position
in Unicode (wide character) codepoints.
It was easiest and most straight-forward to pass the data
through yet-unused program arguments.
The cursor position is marked by a Unicode glyph equivalent
to ACS_DIAMOND, although ncurses doesn't get a chance
to make any ACS translation.
That is, if it's missing from our history.
This makes for a more consistent experience when traversing
the filesystem. Arguably, it's only good for when 'h' is
bound to the 'parent' action.
Also make sure that the offset and cursor are reset when
the path is changed.
Just whether it's safe for it to keep an anchor on the current item.
I've left in the safety check for an empty entries array,
for the theoretical case of an empty or entirely filtered-out root
directory. It will likely crash in that case, anyway,
the cursor may be either -1 or 0.
Usually, when the pager or the editor fail and return a non-zero
exit status, they display a message on the standard error output.
We might also try to redirect this output to show_mesage() but
it might not fit as a whole, so, given the relative rareness
of this situation, this slight inconsistency shouldn't be much
of an issue.
It'd also be possible to redirect the output to run_pager() but
in doing so, we might make the problem recursive.
In principle correct, in practice completely broken.
I wrongly assumed basic_string::find() would behave like strcspn(),
and also used a nonsensical string index to check for the tilde.
The tilde may be escaped, and this first escape may be escaped
as well, so all paths should be expressable.
All will just work(TM) for those who need it, and those who start
filenames with backslashes are begging to have their foot shot.
It appears that even Shift JIS, Big5, GB 18030, and GBK are
compatible with Unix paths, meaning they don't redefine the slash
(0x2F), so I'm okay with my direct multibyte string handling in
unknown locales.
Legacy stuff smells.
Best complemented with a less(1) configuration that also quits on F3,
just like mcview. Or invoke sdn with PAGER=mcview, even though you
lose the ability to view the help that way (but gain filters).
