Here's the scenario that this caveat in the documentation worries about:
Suppose you had three modulii:
$M_1 = p \times q$
$M_2 = p \times r$
$M_3 = q \times s$
Then, when the program outputs the GCD of $M_1$, it'll output $p \times q = M_1$. It'd do this even though the above has enough information to factor $M_1$ efficiently.
It does this because it actually outputs $GCD( M_1, M_2M_3)$; because both factors of $M_1$ appear in $M_2M_3$, both factors appear in the output.
This is certainly possible; as for how likely it is, well, that rather depends on the details within exact RSA generation algorithm. I personally would be surprised, though; most RSA generation algorithms pick one prime, and then pick the other (and run into problems because entropy was lousy when they picked the first prime, and then got better when it picked the second). The primes it picks depends on the internal state (and the entropy). For the first primes to match, we need to assume that the initial entropy was lousy (however, it got better when picking the second prime); this is plausible. However, in order to share the second prime, then two runs must have distinct states when picking the first prime, and then somehow fall into exactly the same state when picking the second. That seems rather implausible (although not impossible) to me.