Integrating the sieve into the client would be possible IMHO - it takes some time to accomplish it, of course.
But you're right, factoring directly helps the people searching for a prime, as a factor found prior to the PRP run makes the latter obsolete. Of course, sieving basically has the same properties as mentioned above, but it works on all k/n pairs, not on the current one specifically.

Hm, sounds a bit like capitalism vs. communism to be...

Sieving has the advantage that the computational effort can be set with a very fine granularity (down to just a bit more than 1p). One can choose whether it takes 5 minutes or 5 days on a certain PC.

Integrating sieving into the main client makes it possible to use the optimal tool for the respective PC architecture - assuming the user sets a preference like "help the project most".
Maybe slow computers and non-SSE2 CPUs are used mainly for sieving, whereas SSE2-equipped ones get to factor and PRP. Thinking of a resource allocation model, I guess a lot of the (faster) non-SSE2 CPUs will factor/PRP as well...