There is nothing like "the golden nounce" which could accidentally get skipped.
Theres a bunch of nounces who satisfy given network diff for the actual block with given transactions to be included in the block.
Which nounce solves the block changes all the time, as new transactions get included in the data we are hashing.

As for slush, you can no longer set the diff manually, the server sets your workers diff based on the workers speed.
Workers are not working with/on/for a special diff value, because the diff is per design something which can only be known _after_ the hash is calculated - thats the trick, after all, you cant decide which diff your result should fulfill, you have to try until you find one fitting your needs.

Your worker-specific diff value just does two things:
Tell your mining software, which shares to send to the server (all with diff equal or higher to your userdiff) and which to drop (all with lower diff).
Tell the server, how high your score for each share should be, the higher your diff, the higher the score.

That way, with higher diff you send less shares to the server but get higher score for each, and with lower diff you send more shares but get lower score. Over some time, this equals out. Higher diff is a little more variance, but a little less traffic to the server, so what the pool does by setting your diff is forcing you to accept a marginal higher variance in your earnings, which will still equal out quite fast, but reducing the server load at the same time, cause the pool will no longer get flooded with low diff shares by miner which hash too fast.

So i know that the salsa core used by the scrypt algorithm uses 128.5 kB or memory (would be smart to use 256 kB per thread) and from what i've gathered, a single thread (running on a spartan 6) would use 576 slices (roughly) or about 3680 blocks of logic, and produce a hash rate of around 12.5 kH/s using DDR2 memory speeds.