> But ROW values or composite types that "are null" are counted, anyway.
Well, there's nulls and nulls. The SQL "IS NULL" construct is fairly badly designed IMO, because it considers both a plain NULL and a row-of-all-NULL-fields to be "null". count(), like just about everything in Postgres other than "IS NULL", considers only a plain NULL to be null.
This is discussed somewhere in the manual, but I think it's under IS NULL, not under all the other places that'd have to be annotated if we decide to annotate as you're suggesting. (One example is that functions that are marked STRICT use the tighter interpretation.)
But count(<expression>) is among the most frequently used functions, and hardly any user reading the manual will be aware of the implications. Maybe just:
I'm with Tom on this. The behavior exhibited is the expected behavior. I haven't looked, but if anything I would make the desired point in "composite IS NULL" that this special (ROW(null) IS NULL -> true) interpretation of NULL is limited to this SQL Standard mandated operator and that when speaking generally about a composite being null throughout the documentation it is done in a scalar sense (I don't know how best to word this but select null::rel yields "" while select row(null)::rel yields "()" on printout (assuming rel has a single column)).