The function you are thinking about (both statement 1 and statement 2) is the same function. Allow me to use your framework to describe it.
The domain of the neural network is the range of $X_1$ and there exists some distribution function $F_{X1}(x)$ that describes how the values of $X_1$ are generated. The same is true for $X_2$ and $F_{X2}(x)$. Then there exists $f$ which maps:
$$
f: \mathbf{range}(X_1) \rightarrow \mathbf{range}(X_2)
$$
Note that this is not $\dot{f}: X_1 \rightarrow X_2$, (not sure if we can this easily define a function whose domain and range are spaces with different densities, an interesting idea to explore but off-topic for us here).
Now. Here I'll diverge from your framework. We will assume that we in fact constructed and trained a neural network which estimates (with good confidence) $f: \mathbf{range}(X_1) \rightarrow \mathbf{range}(X_2)$. What we can say about this neural network and about $f$, is:
$f$ can be used to predict, given a value from the range of $X_1$ a corresponding value in the range of $X_2$. This is by our understanding that our model is a good predictor.
The model models $F_{X2}(x)$. In other words $f$ is $F_{X2}(x)$; and it models the distribution of $P(X_2)$.
Why is that?
In order for $f$ to be a model of $F_{X2}(x)$ then we must have:
$$
f: \mathbf{domain}(X_2) \rightarrow \mathbf{range}(X_2) \Longleftrightarrow
f: \mathbf{range}(X_1) \rightarrow \mathbf{range}(X_2)
$$
Which is true by the way how we designed our neural network. We said that the network receives the range of $X_1$ as its input, which is the domain of the neural network. We also said that the output of the network is the range of $X_2$, therefore the network itself if the model of $P(X_2)$.
Common Sense Reasoning
The neural network receives values taken from $F_{X1}(x)$ as the input but compares its output against values in $F_{X2}(x)$, and it can only change (train) itself against values it can compare something to (its output). In other words, the network does not attempt to understand the distribution of its input. If we have the input as just a single sample from $F_{X1}(x)$ we should still be able to evaluate $f$ correctly.
Reality may sometimes seem different though, and the distribution of inputs actually may affect the training of a neural network in practice. For example, classes/regions with very little support would impact the score of the network (and consequently its estimate of $P(X_2)$. Yet, the idea of little support is not relevant for us since we are thinking of $X_1$, which is a random variable and its expected value would be from a big enough number of samples.
Extra Note
The constrain that $X_2$ must be from the same sample space ($\Omega$) as $X_1$ is not needed, although possible. This constrain would affect how the network is trained (it is a constrain after all), yet it will be the same constrain on how $P(X_2)$ looks.
