Lotta reasons to not worry about this one. SUSA is a great pollster, but the crosstabs are too far off, especially on early voters. Throw it in, mix 'em up. NC will be close, that's a reality. Know that this poll exists and keep watching the numbers.
IDK, this is definitely is an outlier, but one should generally not look too much into crosstabs.
Why? Because it is a lot of them: youngs, older, liberals, conservatives, black, white, blue, red, male, female, rural, suburban, urban, educated, non-educated, low-income, middle, early voting etc, etc. Each group is about 100-400 in an average poll, which gives 5-10 MOE, and ~10-20 MOE of difference. But that sort of cancels out each other (for example, Trump might do to good among EV, but too bad among non-EV). At least in theory. Undersampled groups might though cause bigger errors (but EV in this particular poll wasn't undersampled, I guess).
That basically means, that you always find a odd-looking subsample. By chance. Always. I never look at subsamples < 400. That's why I don't like Monmouth (they totally has just ~400 LV); they are good on average, but pretty often give weird swings.
But OK, you don't like early voters, right?
Totally EV = 659*40% = 264
Then Hillary's and Trump's MOE would be ~ 6.04
The MOE of difference would be ~11.95
So does this result of EV look so strange, given those MOE?
Both Nates wrote about "how to read polls". Atlas should read it