I am skeptical that "looks pretty" drives conversions. I think customers pretty much don't care, based on "ugly" versions winning in many, many of my tests. Remember: the customer doesn't get to see the pretty alternative laid out side by side with the ugly alternative.
Example: form submit buttons in your Internet browser of choice look fairly pretty, right? (Or at least comfortingly familiar.) I would not describe http://www.bingocardcreator.com/images/bcc.net/next-step.png as pretty -- it is what I could throw together in ten minutes. But it clocks "pretty" buttons with the same text on it.
You're welcome to try it yourself if you think pixel-perfect alignment will make a difference. Please post results.
Not that it matters -- the data matter -- but I can certainly rationalize why such a button would be more effective than a standard form submit button: My peripheral vision has seen a hundred thousand form submit buttons by now. I'm not quite as blind to them as I am to standard-sized non-animated web ads, but it's close.
And it seems to me that the problem with being a designer is that it's physically impossible to design a web form in less time than the average user will spend looking at it. You will necessarily spend orders of magnitude more time looking at your forms than a visitor will: Visitors don't look at forms for fun. They click through them fast, or they click away from them even faster. If I had to stare at that button for an hour, as opposed to simply finding it and clicking it as fast as I could, I'd design it differently. Which might very well screw it up.
In the Huffduffer example the keywords are exposed, the fields look clear and spacing is just the way it feels nice. There's also a rhythm - keyword, field, keyword, field... even though it's not a standard form - you just know what to do with it. LukeW's implementation is not as pleasing, but still looks good, because there's some whitespace and no clutter.
I don't think you actually compared good standard form to good madlibs form. You compared it to a hacky, badly implemented madlibs and wrote off the whole idea ("Anyhow: test ended, not touching the madlibs idea again").