The above halo problem is a pretty standard problem. Even the approach I did has that same problem.
The custom node version can fix this problem though by comparing depths and rejecting a sample if it came from a depth that is greater than the window. But then you have to do something like use an old color or use no color, which means you still have a halo of some kind, like a halo of no blur at all. Clever use of temporalAA may help solve it. I haven’t tried that yet but it might be fun to mess around with.