You are correct that there can be more interactions on the frontend, but frontends were both faster and easier to deal with when they did not do all of this. For many internal tools, bouncing back the validation from the backend is also easier to understand than the validation that the website is providing.
I suspect it is a bit of a curve. Zero interaction on the front is not pleasant. All of the state replicated on the front is also not useful and likely to be more problematic.
And I know it is more than a straw man. There really were bad sites like that. We have swung to another side of bad code, though.
Not that I like the idea of several MBs of JS being sent over the wire either... I know that is also a thing, not to mention poor state mgt, which is also a regular occurrence (most angular apps have many state bugs).
Personally, I'm not too bothered by client rendered/driven apps... They can often be split accordingly, and it's reasonable to stay under 500kb JS payload for a moderately sized web application. Not that everyone has as much awareness. I think of the browser as a thick-ish client toolkit as much as it is a rendering platform. That doesn't mean SPA is the right way for everything. I wouldn't do it for a mostly text content site (news, blog, etc). And most of my work is web based applications, not text driven content sites.
There was this really horrific one I did experience, come to find out the entire company is relying on this angular app that connects to an ancient mac mini that IT is responsible for but mostly forgot about. Wasn't a large company, but they weren't stalling in growth!
I can't really cite examples, as I'm not at the old job. But silly things like our "news" homepage was getting so that it lazy loaded all of the news. HR things would load in parts, too. And issue tracking seemed to constantly be a mess.
The tool my team controlled had bloated to excessive points. Some of that was almost certainly my fault. I had designed a somewhat granular backend. That said, I remember things were certainly faster before we had a ton of redux based code to do things.
And yeah, for public sites, tracking almost certainly dominates how slow things are. I was thinking of internal tooling sites. I remember how fast many of the tools were when I started at the last job. I also remember how slow all of them were getting as I was leaving said job. It is more than a touch insane.