I'm in favor of a minimalist approach as far as featuritis: it's much
easier to add features than to remove them later, and I'm especially
wary of features that have seen no convincing demonstration of
utility.
(I'll contrast my minimalist approach to features to my more liberal
opinion about the importance of getting structural changes right
earlier than they might be seen in 'common practice'.)
I find Bert Bos' de(con)struction of proposed link relations
appealing. I'm even a little wary of 'TOP' though, without a more
complete description; if you presume PARENT, CHILD, 'TOP' is where you
get when PARENT runs out.