the point of calibration is to fix up the per-pixel artifacts that a given sensor has. subtracting bias/darks removes bias/thermal signal. each pixel is slightly different, having different response to thermal signal. if you register a frame, you rotate and translate the pixels around such that a given pixel in the registered sub no longer falls on top of the same pixel in the corresponding calibration frame. further, some interpolation is done which mixes pixels together.
same goes for flats - they correct for minute differences in quantum efficiency and for optical train issues like vignetting. if you shift the pixels in the lights around by registering, they no longer fall on the right pixels of the flat.
rob