I'm not saying it's a good idea, I'm saying that tptacek is overstating the risk compared to the far easier option of just spying on the keyboard.
And c'mon, no common security camera could pick up a usable copy of that little graph thing. Half the time they're not even colour. It would take specialised equipment. I was thinking more like thieves going to some rich person's house, finding a computer screen near a window, attaching a high-res camera and waiting for them to log on to internet banking, or something. Sure, in that situation, very bad risk. But geeze, if you're sitting in the Chinese embassy and logging on to the CIA web site using your fingers on a keyboard, tiny coloured graphs of your password hash are the least of your problems.
It is an interesting idea though. I don't like how much information is in the graph, but if it was reduced to just, say, a block of 16-bit colour, that might still be a useful hint to users. Probably more useful actually; the detailed graph is just confusing. And I don't think even tptacek would claim 16 bits of colour is much of a hint brute-forcing a 256 bit hash.
Taking that idea and running with it a little - you could even "salt" the colour randomly for each user, feeding the salt value into the JS. In fact, implementing a per-user salt for the system as proposed would eliminate all of tptacek's concerns. It wouldn't be the same across sites, but it would be consistent per site. I might suggest this as an improvement.
update: duh, the system wouldn't know which salt to load before the user logged in .. forget that idea. Unless it was stored in a cookie of course. Or could be loaded via AJAX when focus shifted out of the username field.