I'd like to bring up my old question that hasn't been answered.
Design of hash functions had producing unique digests (collision avoidance) as its primary goal. Any randomization properties - if they existed - were just a by-product.
Engineers noticed that hash output looked random to them, and started using hash functions as randomizers.
HMAC construct was designed to foil certain attacks against keyed hash functions. What are the reasons to believe that HMAC adds anything to the randomization property of the underlying hash functions?
(I'm not asking for a proof - just give me something that wouldn't be foolish to believe :-)