You have to read a thread enough to know you want to collapse it, which incurs exactly the same "scroll-pain" as actually scrolling. It's not a physical tax from the action required to scroll, but rather a mental stress from making a decision to skip something. Enough of that stress building up at once, and you decide that the comments page itself is probably skippable.
If uninteresting comment threads could start collapsed, that'd be great. But they won't, because humor-fluff and other such things are superstimuli for upvotes, so you can't use anything about the vote tally to determine collapsed-ness. (If there was a secondary voting system--like, say, if enough people collapsing a thread would make it start collapsed for others--this might work. But then you'd have to take into account the people who collapse everything as they read it, to mark their place...)
The problem is, your definitions of what is uninteresting, useless and fluff are subjective, and no more relevant than anyone else's, which leads to a conflict of interests within the community about the bounds of what Hacker News content should and shouldn't be.
I think the models which would satisfy the most people are opt-in, in this case, choosing to collapse threads and ignore users rather than expecting the hivemind to do it for you.
Perhaps if users had a custom set of filters which automatically collapsed threads for them based on their own criteria, that would solve part of the problem. But I don't think it's too much to ask of people to actually take the minor effort to form an opinion on what they read, or curate their own account, if they expect content they don't like to be hidden from themselves and potentially from others.
Disagree on your first point, the top comment is often obviously spawning a huge argument that goes nowhere because it touched some hot-button political issue that was a minor part of the story.
If uninteresting comment threads could start collapsed, that'd be great. But they won't, because humor-fluff and other such things are superstimuli for upvotes, so you can't use anything about the vote tally to determine collapsed-ness. (If there was a secondary voting system--like, say, if enough people collapsing a thread would make it start collapsed for others--this might work. But then you'd have to take into account the people who collapse everything as they read it, to mark their place...)