I think what you’re saying has merit, but you have to remember the reason why we want to unwind those H1 optimizations for H2: Caching. If any part of a bundled asset (be it CSS, JS or an image sprite) is changed, the entire bundle is invalidated and has to be downloaded again, even the parts that didn’t change.
We unwind these H1 optimizations in H2 environments because connections are cheap. In fact, they’re not even connections anymore. They’re multiplexed streams encapsulated by one TCP connection.
Now to the merit of what you’re saying: Yes, compression is a factor in this case. But you also need to gauge your audience. How many of them are return visitors? First impressions are certainly important, but if you have a high portion of your audience that is returning (and you’re deploying new code/assets frequently!) caching effectiveness is important, too.
I think the only sensible answer to this what you said: “Go the smart route and never forget to test!” I would amend it to say, “Go the smart route, find out who your audience is and how they behave, and never forget to test!” In this ever-changing field of web performance, there are often multiple “correct” answers.
Thanks for reading and offering your input. :) I appreciate it!