I don't have any specific knowledge of Apple's actual implementation, but I've always assumed it used the full-resolution still image as a keyframe (I-frame), and the video clip as B-frames before and afterwards, compensating for changes over time. Choosing a new key photo may be full resolution, but is likely to be degraded in sub-perceptual ways.

It does pull multiple full frames off the sensor in rapid exposure tho, and chooses whichever one it thinks best (blur, ML, whatever) for the initial key photo.