As suggested by Thomas Andrews in the comments, the resolution of the apparent discrepancy is that there are different measures.
Length, area, and volume are different measures. If you take a disk of positive radius in the $xy$-plane in $\mathbb{R}^3$ and try to compute its volume (3-dimensional Lebesgue measure), you get zero. But it has a nonzero area, and if you use the 2-dimensional Lebesgue measure on $\mathbb{R}^2$, you will get the nonzero area.
If you compute the "area" of a circle, you get zero, but if you compute the length, you get $2\pi r$.
So what measure are we using when we compute the arc length of a curve? Is it Lebesgue measure on $\mathbb{R}^1$?
No, it the measure defined by the arc length formula $$s = \int \lvert r'(t)\rvert dt$$
I go into more detail about this measure in my answer in Surface measure=Lebesgue measure on $\mathbb{R}^{N-1}$?
So for your question, the measure we use on 0-dimensional sets is (signed) counting measure. This is also the measure induced on boundaries by arc-length or Lebesgue measure of intervals on $\mathbb{R}^1.$
In order to make Stokes' theorem and the fundamental theorem of calculus, it is a signed measure, which takes into account orientation. With that in mind, we have that $\{a\}^{-}\cup\{b\}^{+}$ has nonzero measure, and $$\int_{\{a\}^{-}\cup\{b\}^{+}}F(x) = F(b) - F(a),$$
as required.