I don't mind breaking compatibility for the keying nodes. The files there usually don't have an alpha channel to begin with, so it's doesn't matter.

For the Set Alpha node I'm less sure how it is used in practice. We could add an enum with options Replace and Multiple on that node? Maybe that's useful for other purposes anyway.

I'm starting to have doubts here.

Having an option for compatibility in Set Alpha node seems fine.

But what is expected behavior for the "new" "proper" behavior, what is the mental model? It is simple multiplication, or is it something more elaborate like: output is properly premultiplied RGBA, with A channel exactly matched to what was passed to the Alpha socket?

I think the mental model for Multiply is applying an alpha mask to an image. The alpha channel is not necessarily the same as either the RGBA image or the alpha mask, it's a combination. We could call it "Apply" if that is more clear? Not sure if there is a better term.

@Brecht Van Lommel (brecht), sure mental model of "Multiply" is "apply a mask". But I was more wondering about what is the proper mental model of the "Set Alpha" node. In other words, what is the expected behavior? Replace the alpha (keeping the premultiplication of color) or "apply the mask" behavior?

Not sure I understand the question then. Multiplying I think is by far the most useful behavior and should be the default, but not sure if that's what you mean by it being the mental model of the node.

I do think that set alpha node always did what it says it did, but with this change the name doesn't match with what the user wants or the node does. We could also see it an opportunity to rename the node to Apply Mask Node and change the modes to Color and Alpha or Alpha. But before that we should validate with @Sebastian Koenig (sebastian_k) and Andy.

I'm am sorry for being so late with commenting on this.
I think it's good having that enum in that node!
I am a bit unsure of this terminology. "Replace Alpha" makes sense to me. "Apply Mask" however seems weird to me. Why not simply call it "Multiply Alpha"? Isn't that what the node is doing anyway? "Apply mask" implies that a mask is used, which in compositing is quite common indeed, and we also do have actual masks in the compositor. But more often the Set Alpha node is not used to only apply the mask from a mask node to the image, but the result of some operations with the Alpha channel of a render or an image, sometimes involving masks, sometimes not.
Also, there seem to be just 2 options in the node, which are opposites of eachother, so if the terms are not made up of the same logic and use different words for the same thing ("Alpha", "Mask"), I think it is confusing toi the user.
That's why I would suggest to use "Multiply Alpha" instead. The tooltip then makes it clear what's happening.