Passive sign convention means choosing an arbitraty current direction, then assuming that current always enters the positive terminal and flows out of the negative terminal of an element. Then the following will hold true: $$ U_R = IR \qquad U_C = \frac{q}{C} \qquad U_L = L \frac{dI}{dt} $$
when we apply the KVL
And the sign of voltage in KVL is determined by the sign of the first encountered terminal.
Under which conditions does this convention work and why does it work under them?
Why do I doubt this always holds true?
First of all, because I haven't seen any comlete general rigorous reasoning that this holds true in an arbitrary circuit.
And also it at least doesn't seem to hold for cases when we have a single loop with elements in series and thus single current, yet the polarities of the elements' terminals may not always be "+ first, then -" or vise versa for all elements as we follow along the loop.
When we apply KVL, for example, to a circuit obtained by connecting an uncharged capacitor to a charged capacitor through an inductor, the system will become something like that:
And the top-plate capacitor terminals will always be positive. Thus we cannot choose the signs according to passive sign convention like that: