I'm not sure what your professional experience making websites is, but mine has shown me time and time again that making good looking sites is an iterative process full of workarounds for real-world considerations.
The method you're suggesting would make it excruciating to converge on a good design with any real designer. Look at the markup any good looking and well-implemented web page (especially one with dynamic content), and you'll see a huge amount of extra tag work to accommodate the real world of cross-and-backwards-compatibility between browsers, CSS limitations, browser quirks, and other concerns that people making professional sites require.
If every one of these changes and workarounds required changes to the code, which would require my time instead of the designer's time?
The Rails way is just copying a method that nearly every web toolkit uses, and they use it for a reason. Arc's way may be cleaner, but without a templating language it will be at a disadvantage.
- If you get the layout, including both how it should look and how things should work, from an external source, or if there is substantial external input, use templates. It will save time and headache.
- If you are just about the only one responsible for the layout but support skins (for lack of a better word; the CSS part), use the HTML tag functions and macros embedded in Arc.
Since Arc is at this point targeted for exploratory programming, the former is an unlikely case.
What I would like to see, though, is a templating engine akin to HTML::Template  in Perl 5. It has the minimum of control structures and extra syntax on top of standard HTML, just enough to make it possible to use the template for dynamic content. All templating frameworks suffer from the same fault, though: they define a new but severely restricted embedded language that's mixed in an unpleasant way with the HTML source. The language embedded in Arc is a much cleaner way to do things.