With a text field you have a tap to edit, a tap to switch to the numerical keyboard (unless you can start there with the SDK), a tap for each # (say 3 on average), and a tap to save the record and send the keyboard away. That's 5-6 taps with a fair bit of distance between them on the screen, assuming that you don't have to drag the largely keyboard-obscured screen around to see other data ("I know I it's a few miles further than the post office... What did I enter in for that?").
OTOH, I just tried to simulate the odometer with the alarm clock setting controls on the iPhone (which is similar) and the flick-tap that it allows is pretty damn fast with zero major screen changes.
I find the iPhone keyboard to be a PITA and typos abound, even after practice. Editing a text fiend would bring up the keyboard, obscuring a LOT of info.
I'm not disagreeing, per se-- just saying that it's not so cut and dry that I wouldn't test the two side by side for speed/learning curve.