The non-identity problem, in its cleanest form, runs like this. Suppose policy A leads to a future population of people who live decent but constrained lives, and policy B leads to a different future population whose lives are flourishing. The same individuals do not exist in both futures — the policies change who comes into being. On any plausible person-affecting view, we cannot say A harms anyone (the people under A are glad to exist; the people under B never existed to be deprived). And yet B looks better than A. Almost everyone agrees B is better. And almost no one can explain why.
The fixes that don’t quite work
Wide person-affecting views let us compare lives across populations, but they require us to weigh the welfare of merely possible people, which most of us find vertiginous. Impersonal totalist views handle the case but inherit the repugnant conclusion. Asymmetry views say we are obligated to avoid creating bad lives but not obligated to create good ones — elegant on paper, but they generate strange recommendations in mixed cases.
The non-identity problem is one of those rare philosophical results where every escape route turns out to lead to a conclusion the escapee themselves does not want to defend.
Where I have landed
I have stopped looking for a single principle that delivers all our intuitions and started looking for a story about why our person-affecting intuitions are so robust here. My best guess is that they encode, fairly badly, a much older intuition: that obligations require a possible relationship. We owe future people because, if they come to exist, they will be in a relationship with the policies we set. The non-identity puzzle is what that intuition looks like when you push it past its design parameters.
I agree that the puzzle is best read as showing the limits of person-affecting reasoning, not as something to be solved within it.