Hi Andy
Yes, we are aware of this issue... and in fact we have found to this date anything to "fix" it. Let me explain the problem. HDRWT works in a multiscale way, modifying not only the weight of the wavelet channels, but changing the contents of each wavelet layer with respect with other wavelet layers and residuals. So, if you have a preview that covers only a partial zone of the image there will be different contributions to the whole computation, and the result may (well, almost surelly) differ from appying the process to the main image. Something related happens if you try the process on a subsample version of the image...
So, we have to find a way to make previews to give an accurate result, without the need of applying the process to the whole image, and it is not easy. Meanwhile, I suggest you to try as a first approximation a 2x downsampled version, and then fine tune the result with a preview with the same contents/size as the main image.
Thanks for the feedback!