Since header lengths are not limited, and a single TCP packet's payload is quite limited, long headers can cause very measurable latency difference. Additionally, while I agree the generation / parsing overhead is probably quite small, saving it for every HTTP request is still a boon.
I'm also curious where you are reading raw HTTP from?
For me it's primarily in two situations, reading a packet capture from WireShark, or in the browser's debugger. In both cases, the tool will end up translating the request for me.
Just because I use Charles or Wireshark doesn't mean that I only want to use those specialized tools. I have definitely been in situations where I'm doing something like running nc as a proxy and looking at raw HTTP. I wouldn't choose to throw that away and revert to the bad old days without a big win.
But now nc supports SSL natively, making it super easy. Just as it will support binary HTTP natively, making it super easy. And everyone will finally stop caring about ASCII.
I'm questioning the measurability of this, though. Smells like premature optimization. You could be right, but I'd like to at least measure it before we go about changing one of the fundamental protocols on the internet.
Last time I read raw HTTP was when writing a script to automate some stuff on a web page. I specifically did not want the browser's headers and behaviors. I had a bug which only happened from my script, and raw HTTP helped me track it down. I could have used wireshark, but I am much faster in vim for a simple task like that.
As a comparison to your scripting story, you would use Wget or Curl or LWP::UserAgent or a thousand other things to automate HTTP requests. One function call to do what you did manually. To find bugs you would use an HTTP fuzzer like Skipfish to automate the process. If you think somehow your manual process was faster, I say to you, teach a man to fish...
(I automate things in web pages for a living, and I only use tools like Firebug and LWP)
I can't tell you how many times I've manually read HTTP. To be sure, it's insignificant compared to how many HTTP headers have passed through my computer unseen by me.
ASCII may be inefficient, but computers are really fast, people are not. I don't have any measurements, but in making a browser, HTTP header parsing/writing was never near a performance issue. Bandwidth wise it also tiny compared to that image file you'd inevitably download for every page visit.
And sometimes you don't have tools. Sometimes you don't want tools. Sometimes you want to use tools that work with text to analyse your problem.
(You other arguments might still stand, though :) )