>In the HTTP2 world, extra requests don't add any additional latency.

They don't add any additional TCP connection latency, sure. I think it's a bit unrealistic to claim there won't be any congestion of the multiplexed stream which causes additional latency.

>I think I'd heard that it even uses a common compression dictionary for all requests, which means that boilerplate code that's duplicated between components will get compressed away.

I may be misremembering, but I thought that was just for request headers?