Thanks Joe. I'd already reached the conclusion that the second of your suggestions, doing a second HTTP GET on the img src attribute was the most expedient way forward for getting the test working. I guess the main thrust of my question (which wouldn't have been apparent from the way asked the question!) was to do with the underlying object graph that the webkit browser was building up. My reasoning was that it would have built a complete DOM, just like a non-headless browser would visiting the same page, and that I'd ultimately be able to get hold of an object with, say, a width attribute that would only be set by pulling in the actual image. However, despite spending some time with the source of both capybara and capybara-webkit, and using pry to inspect the object graph, I couldn't find a way to do that.
Tldr: I've solved the presenting problem to get the test working, but I'm still wondering quite how much of the rendered page is accessible to other tests I might think of writing in the future.
Thanks again,
Ian