That's what the original paper is kind of about. The Scientific American piece is a little misleading to me, because it implies the paper is about lack of evidence, when I think the original paper is a bit deeper than that, and about definitional issues and so forth. I think the original authors might say that it's about the need to figure out what exactly is working when it does, and what works and what doesn't.

Self-reports of well-being are important, because self-perceived well-being is so important to the construct. I'm not saying other perspectives aren't important, but think of it this way: if you thought you were severely lacking in well-being, but others thought you weren't, would you want others to ignore your own perspective on that?

Well I'm not an expert in the domain but I'd wager there's models for what happiness is, from which you can determine how to measure that. For example if you thinks it's a variable composed of self-appreciation and appreciation of others you can have a good estimate with a properly-designed study.

Why would that seem wrong ? Because it could miss cases where people don't know their own happiness ? I think a model based on actual observations is more right than a hypothetical transcendental happiness. Because people could lie ? That's where the proper design of the experiment steps in.

This is off the top of my head.