The difference is indeed subtle. I understood it like this:
Suppose you had a hash and a BadHash. You pick x' and a random nonce r', compute y' = BadHash(r'|x'), and give me y'. It turns out that if x' is UTF8-encoded English text, I am able to tell you what x' was, and (though not strictly necessary), I can even tell you r'. If x' happened to be just a random bit-string, I'd be out of luck. But no matter, this is clearly not a hiding hash.
Now the puzzle: I give you a value Y', and a randomly chosen value R' (say "11110011...100"), and ask you to find an x such that BadHash(R'|x) = Y'. Good news: it turns out that Y'= BadHash(00101...0001 | UTF8("Bitcoin is deflationary")). So because BadHash is non-hiding (plus), you can determine an R (namely 00101...0001), and and x (namely UTF8("Bitcoin is deflationary")), such that BadHash(R|x) = Y' But this doesn't help you, because the puzzle specified that you need an x that works with a different R, namely "11110011...100". So you haven't solved the puzzle.
You see, then, that the two properties aren't equivalent.