For an attacker to learn a cookie that's valid for a given victim "source IP", he only needs to be passive observer somewhere along the route. Even if we believe that's very hard in most cases, if it's possible at all, he has the mother-of-all anonymous reflected DoS amplifiers. http://tools.ietf.org/html/draft-ietf-tcpm-fastopen-02#secti...
So, yeah, using a key that's rotated after a short amount of time -or- number of uses (whichever comes first) seems like a good idea.
An attacker who doesn't want to do a MITM attack because that might be noticed can set up sessions to all kinds of servers outside the NAT which support TFO. Then all these TFO cookies are used in spoofed SYN packets with the source IP being set to the host behind the NAT that the attacker wants to flood. Easy enough.
Of course, some will argue that if the attacker is inside your NAT, you're already pwned.
I don't think that's a very good principle for the security design of internet protocols.