Look at the articles marked "-R": I was reading them just some
minutes ago, but selecting them didn't change %O from - to +.

Look at the last two articles: when I entered the group they were
marked as not downloaded, but reading them downloaded them and
changed the "- " to "+R".

I'd almost say that "." (new/unseen) marked articles get frobbed
correctly.  It's even true.  But also there is the second line in the
summary window: Ami's article got changed from "- " to "+R".

It is all very confusing.

Maybe it is useful to know that the internal data structures in Gnus
are not corrupted: after reading all new downloaded articles in a group
(changing "+ " to "+R" in the process), I can exit the group and then
re-enter it and then the old downloaded articles change their marks
from "- " to "+ "!  So the old articles are shown okay when there are
no new ones, but the presence of new ones screws things up for the
old ones.

In the previous paragraph, both new and old articles are unread and
downloaded.  New articles were downloaded in the same session (or
perhaps by the last `g J s' invocation), whereas old articles were
downloaded in a previous session.  (I'm not sure if exit Emacs/start
Emacs is what distinguishes old from new articles, or just another `g
J s'.)
-- 
A turnip curses Elvis