At this point the intent of 906bfcad doesn't really matter - and given the number of complaints in the month since v10 went live I'm tending to lean toward bringing the pre-10 behavior back. There is no ambiguity involved here and the breakage of existing applications seems considerably worse than the technical oddity of allowing (val) to be interpreted as a row_constructor in this situation. From a standards perspective we are strictly more permissive so no new problem there.
On a related note, the 10.0 syntax guide is wrong, it needs to break out the parenthesized single-column and multi-column variants separately: