I would say yes, your understanding is wrong. My guess is that they are not using OCR for finding the alignment marks or the bubbles. The alignment marks, I assume, are used to figure out the correct region and location for the bubbles. OCR on the other hand is used to turn pixel images (images which contain text) into text. For instance, a scanned text based document.

So if you scanned a ballot and sent the scanned image to an OCR engine it would return the text that it found on the ballot image. All non-text would just be skipped over. Even some of the text might not be recognized depending on the quality of the image, for instance multiple times being faxed, and how much it might be skewed.

You'll notice on the text based captcha's they degrade the characters in hopes to make it such that a human can figure it out but an OCR engine could not.

My guess is that the text on the ballot is just to aid the person filling out the ballot. The ballot reader will most likely just be trying to find the two columns of bubbles and determining which ones are filled in. Let's say the left column is democrat and the right column is republican. The first row is presidential, second row first senate race, etc.