Arguably the resource is the dataset of stock exchanges, and the CSV representation is forced to omit all the metadata but the HTML representation isn't.

I understand what people are getting at, as it's not really that big of a logical leap. I think the fact that it is somewhat of a stretch, but still "in the lines," is exactly why it is a "trick": it isn't doing anything particularly invalid or hacky, it's just not necessarily what you'd imagine when reading the RFC. Content negotiation to me is more about serving the optimal content to a given agent, not really about selecting modalities for different use cases based on different types of user agents.

I think both cases are "valid" although I think it is inherently less tricky if the document talking about and previewing the dataset is referenced via a separate URL from the dataset itself. (Which, of course, entirely mitigates problems like Apache Spark having HTML in the Accept header.)

Well, neither a picture is really "the same" if it's encoded in different formats, say, JPEG or PNG.

Actually, that's a pretty good point. The thing is that the URL itself refers to some conceptual resource, and the response is ideally a representation of that resource, potentially one of multiple. If you take the same source image and encode it multiple ways, although the two resulting images are different from eachother, they are representations of the same underlying image/resource. But if you were to provide a different image, or alter the image in other ways, I think this would be pretty tricky actually, even if the modification was something trivial. You can imagine a simple use case like a WebP image with the text "Your browser supports WebP" and a PNG image with the text "Your browser does not support WebP." The point is that content negotiation being used to present different logical resources that are not necessarily interchangable representations of the same data feels tricky. I think that is still compatible with the fact that it's all within specifications.