Yes, I wasn't precise. As I understand it, "white balance" data has two roles in the Adobe image pipeline. The first role is the usual one, which is to set the white balance of the image, and happens after demosaic, the same as for a "normal" image pipeline. So yes, the white balance operation occurs after demosaic.
However, white balance data may be used before demosaic as well: The second role that white balance has is in an Adobe pipeline is to avoid metamerism, etc, where Adobe use two color matrixes and interpolate them based on the color temperature. That part may/probably(?) occurs before/during demosaic in order to get an accurate measure of distance in color space for use in the demosaic process itself. Assuming they're using a demosaicing technique that needs that. (Big assumption - at the time I wrote the response above, in 2009, I was fairly sure they did. Circa 2012, I'm not nearly as sure.)
interpolation between 2 color matrices shall depend on the particular WB that we select in UI, right ?
change in WB by you in UI shall lead to reinterpolation between 2 color matrices in ACR/LR, it can't be a fixed interpolation - can it ?
and demosaick as said by Eric happens before WB, "the usual one, which is to set the white balance of the image", then technically, you need to do the whole stuff again and again when WB sliders are changed in UI...
you need to do matrix multiplication for the whole raw data using a reinterpolated matrix, demosaick and regular WB operation after that for each slider move...
at the same time we have linear DNG (demosaick happened and for example done by genuine ACR algorithm, not by crippled /as Eric said/ DNG converter code) produced by ACR and WB in ACR on that file (linear DNG) gives exactly the same image as WB in ACR on the source raw file... so either there are no reinterpolation or everything is done on demosaicked data.
please tell me where I am wrong, because I probably am.