I've never built a weight loss goal, but my understanding is that it automagically calculates the width of the road based on observed variance in past datapoints.
That said, I generally find it's best to beemind things I can control directly (as with some of your supportive goals). I would probably approach this by 1) setting up the supportive goals with initial slopes that I believed would achieve my goals, 2) setting up a weight-loss type goal with a flat slope, for data tracking/averaging purposes only, and then 3) adjusting the slopes on the supportive goals as needed to achieve the change I wanted in the appropriate timeframe. In other words, I'd put the "sting" of Beeminder on the things that are actions, while still taking advantage of the tracking functionality for the outcome.
At some point it's a question of why you want to improve your pace. Is it simply a "do better" sort of goal, or is there a race you're trying to win? If you're committed to achieving that pace for race day, you could put a commitment on the weight-loss type goal, but as you said, that is subject to variations, and you might end up derailing due to factors you can't entirely control. But there will be things you can't control on race day either, so maybe that sort of hard commitment is what you need to achieve the real goal - winning the race.