Either performance is so critical that a few kb to do feature detection is too much, or line performance has improved so much that 2MB of JavaScript for a text box and two buttons is "acceptable".
You can't have it both ways.
Your argument goes against empirical evidence in this instance. You can have it "both ways" when client-side feature detection is the slower choice on high bandwidth connections and you want to consistently render the UI within 200ms.
Performance goes beyond raw bandwidth, and as with all things engineering, involves tradeoffs: client-side feature detection has higher latency (server-client-server round trip and network connection overheads) and is therefore unsuitable for logic that executes before the first render of above-the-fold content. All of this is pragmatic, well-known and not controversial among people who work on optimizing FE performance. Your no-serverside-detection absolutism is disproved by the many instances of UA-string parsing in our present reality.
By your logic, because McDonalds is popular, it must be a healthy choice then.
This was back when a large fraction of search users were on 56k modems. Advances in broadband connectivity, caching, browser rendering, resource loading scheduling, and front-end engineering practices may result in the non-intuitive scenario where the 2MB Google homepage in 2024 has the same (or better!) 99-percentile First-Meaningful-Paint time as a stripped-down 2kb homepage in 2006.
The homepage size is no longer that important because how much time do you save by shrinking a page from 2MB to 300kb on a 50mbps connection with a warm cache?Browser cache sizes are much larger than they were 10 years ago (thanks to growth in client storage). After all, page weight is mostly used as a proxy for loading time.