This is how the mathematics of complex signals work.
The proof begins with Euler's formula:
$$ e^{i\varphi} = \cos \varphi + i \sin \varphi \tag 1 $$
For signal processing, instead of $\varphi$, we are usually thinking about some sinusoidal oscillation at angular frequency $\omega$ that varies with time $t$, which we can write as:
$$ e^{i\omega t} \tag 2 $$
This is what the signal generator block outputs, when in sine mode and with a complex output. By (1) above you can see both the real and imaginary parts are sinusoids at angular frequency $\omega$, just offset 90 degrees in phase.
When you multiply two of these complex sinusoids together, at frequencies $\omega_1$ and $\omega_2$, you get:
$$ e^{i\omega_1 t} e^{i\omega_2 t} \tag 3 $$
which simplifies to
$$ e^{i (\omega_1 + \omega_2) t} \tag 4 $$
which, again by (1), is a single complex sinusoid at frequency $\omega_1 + \omega_2$. There is no difference term.
A consequence of this math is that $\omega$ can be negative. Which is why in GNU Radio if you have a complex stream at a sample rate of say 48 kHz, that can represent 96 kHz of bandwidth: from -48 kHz to 48 kHz.
The sum and difference terms when heterodyning real-valued functions comes about because a real function can not unambiguously represent positive and negative frequencies, but mathematically, they are still there.
How? Consider two complex sinusoids, at frequencies $\omega$ and $-\omega$, summed together:
$$ e^{i\omega t} + e^{-i\omega t}
= \cos \omega t + i \sin \omega t
+ \cos -\omega t + i \sin -\omega t
\tag 5 $$
Considering the trigonometric identities:
$$ \cos x = \cos −x \\
\sin x + \sin -x = 0 \tag 6
$$
Now (5) simplifies to:
$$ e^{i\omega t} + e^{-i\omega t} = 2\cos(\omega t) \tag 7 $$
Which means when you multiply two real sinusoids to heterodyne a signal:
$$ \cos \omega_1 t \times \cos \omega_2 t \tag 8 $$
Then by (7) and neglecting the factor of 2 (since it only changes the amplitude of the result, and that's not important), equivalently you're doing:
$$ (e^{i\omega_1 t} + e^{-i\omega_1 t})
(e^{i\omega_2 t} + e^{-i\omega_2 t}) \\
= (e^{-i(\omega_1-\omega_2)} +
e^{i(\omega_1-\omega_2)}) +
(e^{-i(\omega_1+\omega_2)} +
e^{i(\omega_1+\omega_2)}) \tag 9 $$
Notice the difference of the frequencies on the left, and the sum on the right. Each group is comprised of positive and negative variations of the same frequency, which by (7) we know simplifies to just a real-valued sinusoid. So (9) further simplifies (again neglecting that factor of 2) to:
$$ \cos((\omega_1-\omega_2) t) + \cos((\omega_1+\omega_2) t) \tag {10} $$
And there you have your common real-valued function heterodyning equation.
Thus, any real-valued function has both positive and negative frequencies in it, but the negative frequencies are just a "mirror" of the positive ones. It's because of those negative frequencies that LSB demodulation can "flip" the spectrum, and it's the negative frequencies that cause the difference term when heterodyning real-valued functions.
