The only issue with your second point is that it can eventually become a quagmire when you do need to upgrade it.
I work for a very old company who held to that philosophy for many years. And while any individual component could be looked at and seen as running fine, when they did finally decide it was time to upgrade they were faced with needing to upgrade everything simultaneously.
All of the tech was too old, so no current tech had the sort of backwards compatible bridge that helps you move forward. It’s like figuring out how to get your telegram system to also work on your WiFi network, nobody makes any interfaces for that.
Instead of slowly and gradually replacing components over time, they’re faced with a single major overhaul that’s put the entire company at risk because they have to completely shut down for over a month.
Yep the flaw is assuming that humans would actually select for constructive comments. It’s a case where humans claim that’s what they want, but human actions do not reflect this. We’d eventually build yet another ‘algorithm that picks what immediately appeals to most users’ rather than ‘constructive’. You’d also see the algorithm splinter along ideological lines as people tend to view even constructive comments from ideologies they disagree with unfavorably