I believe that the choice you should make depends on how good you are at tennis (at base level). If your only chance of getting even one point against Federer is if he makes a double fault, then you should choose 6-0, 6-0, 6-6 (6 - 0). If you are good enough to perhaps win a game against Federer with non-vanishing odds, then your optimal strategy may be different. The reason for this is Bayesian updating.
- Supposing you choose 6-0, 6-0, 5- 0 (40 - love). If you are really bad, then Federer will adjust his game once he realizes how bad you are. He may choose to serve differently with an intention to avoid double faults as much as possible at the risk of making a bad serve. He has essentially adjusted his per-point win probability. If you, dear reader, are really bad, then one must adjust one's model to account for this Bayesian updating. Thus, 6-0, 6-0, 6-6 (6-0) might be the more realistic solution. Once FedEx realizes you are terrible, you will be toast.
Another thing to consider - this one very tricky judgement bias and psychological effect, namely, the "Overconfidence effect". Consider the following two examples :
(1) A survey of 1 million high school students showed that 70 percent think they are above-average leaders (only 2 percent rated themselves below average). 
(2) In another study 94 percent of college professors claimed that their research was above average. (See Seed magazine reference below : )
People may think that they may actually have a chance of taking even one point off Federer whereas in reality, that probability p may be much, much closer to zero. The "6-0, 6-0, 5- 0 (40 - love)" solution may suffer from this bias for people who are somewhere between the good to great spectrum.
Another point, this one related to the framing of problems.
A puzzle is not the same as a mathematical problem. Whereas a mathematical problem is in the abstract, a puzzle may add additional constraints because it is set in the real world. In my solution, the chances of Federer actually dying may seem like a humorous suggestion or an aside but such a probability must, in fact, be taken into account in the real world. Another thing to consider is that Federer may be tired (this point, which is Amit Chakrabarti's, is different from the situation of Federer actually dying).
Again, a puzzle is not the same as a mathematical problem. If constrained optimization is what you want, then the puzzle setter would want to state the problem more carefully. Settings in outer space are an oft-used techhnique for framing such problems so as to ensure strict mathematical correspondence. Use them. Robots are another useful device. Use them too.
Update (July 8): fixed typos. updated the post a bit.