Biased sample if those who flee the long-replies-to-downvoted-comments threads have already left

Valid point. How can we test this?

At the point where LW starts being unfun for me to read, I panic.

Being concerned about the signal to noise ratio is reasonable, but yes this sounds like panicking. Deciding that there's a problem is not the same thing as deciding that a specific course of action is a good solution to the problem. (I shouldn't need to tell you that.)