I hope I've shown that I agree with you, that "the original response from the server ought to be informed by the user agent string in an environment like YouTub

I hope I've shown that I agree with you, that "the original response from the server ought to be informed by the user agent string in an environment like YouTube, mainly for latency reasons". But, differently from YouTube, it's possible to serve a latency-optimised, user agent-optimised response in such a way that it's just a performance heuristic, without changing the actual functionality, which is still determined by traditional feature detection to the extent that is possible to do. When it's implemented that way, it's quite safe to serve the "wrong" response sometimes. Because it's safe to serve the wrong response, that allows automatic latency optimsiation, instead of brittle, hard-coded rules that will be wrong sometimes. Inputs are user-agent, client IP, client cookie state, and feedback about whether other resource bundles were fetched as well, or other kinds of feedback such as detected features. Output is which bundle to serve. This may improve performance over a hard-coded approach, because it adapts continually to new user-agents out there. Yet it is still coded as traditional feature detection (for features where that's possible), and works the way feature detection has always worked.