Your wolf analogy is the direct opposite of natural selection. There's nothing random about it. But let's get to the core issue.
Your point about the stages of color change is well taken. And some traits do confer a higher survivability rate, but since the rate of change starts small and changes slowly (as you say), so does the survivability advantage of the change. So let's say it's a hundred stage process from say a green (normal) salamander to a bright yellow one. Each iteration improves it's chances of survivability, as you say. Lets say, with birth and death rates, that there are ten million of these salamanders around at any one time. We'll use our previous numbers, and head to a new generation. 95% of the new ones aren't going to mutate at all, so we're down to 50,000. 95% of mutations are meaningless, so only 250 (1 in 4000) get a mutation that does anything at all. But did it occur on your gene? The are 10,000 of them, so your group of 250 has a one in 10,000 chance of getting the "one hundredth closer to bright yellow" gene. Meaning each one has only a 1 in 40 chance of getting to step one. Dice have no memory, so the next generation doesn't help you. In the next generation, one salamander out of ten million still only has a 1 in 40 chance of of getting one one-hundredth of the way to bright yellow. Go a hundred thousand generations, and the next one still rolls the same odds.
But let's say you get your guy. His dad gets naturally selected, he gets the trait, and he's produced. He's 1/100th less green than his buddies, he has a 1/100th better chance for survival than his compatriots. His generation, however, faces the same long odds at the genetic level. There's only a 1 in 4000 chance that he'll get a mutation at all, etc etc. It's only when we talk about this guy surviving to give offspring that the improvement kicks in and natural selection gets to play a part again. It's exponentially, astronomically unlikely that we get to step one hundred. Genetic drift will help a little, but not at first. Genetic drift is somewhat logarithmic, and the impact of it on early genetic mutation approaches zero.
And I've drastically oversimplified the problem. You've got millions of salamanders over millions of generations. Fine. It's a huge number. I'm trying to show you that the problem is exponentially greater.