This question is very old, but I have just spent some time on the very same problem so I thought I might post my work-around. I agree the sign of the argument of $D_{-1/2}(x)$ is not mentioned in https://dlmf.nist.gov/12. Then the defining differential equation is second-order and even in $z$. So the vector space of solutions has to be of dimension two, and to contain one purely odd and one purely even solution. Then I'm afraid what mix of these is provided by one particular numerical implementation ends up being a matter of convention.
What I've figured out regarding the scipy.special.pbdv implementation (which turned to be just what I needed) goes as follows (some equalities I use are not in https://dlmf.nist.gov/12, so I use the numbering from Abramovitz & Stegun):
- As you said, for real positive argument, (A & S, 19.3.7), $D_{-1/2}(x)$ returns $U(0,x)$, which turns out (A & S, 19.27.5), to be equal to $\sqrt{\frac{x}{2\pi}}\mathrm{K}_{1/4}\left(\frac{1}{4}x^2\right)$.
- For real negative argument, use is made instead of (A & S, 19.3.8), which states that $U(0,-x) = \sqrt{\pi} V(0,x)$. So for $x\leq0$, $D_{-1/2}(x)$ returns $\sqrt{\pi} V(0,-x)$. Which turns out (A & S, 19.27.4), to be equal to $\frac{\sqrt{\pi |x|}}{2} \left[\mathrm{I}_{1/4}\left(\frac{1}{4}x^2\right)+\mathrm{I}_{-1/4}\left(\frac{1}{4}x^2\right)\right]$.
Then, using (A & S, 19.27.5), and noting $\mathrm{H}(x)$ the Heaviside step function, one can summarize all this as:
$$
D_{-1/2}(x) = \sqrt{\frac{|x|}{2 \pi}} \mathrm{K}_{1/4}\left(\frac{x^2}{4}\right) + \mathrm{H}(-x) \sqrt{\pi |x|} \mathrm{I}_{1/4} \left(\frac{x^2}{4}\right).
$$
I have no idea what happens out of the real line.