I think I figured it out: it makes physically more sense to express the XYZ primaries in terms of the primaries of the CFA. The best you can then do in the colors from the CFA to XYZ, is to minimize the error made. This is what is implemented in the pseudoinverse() function, as described here:
http://www.sci.utah.edu/~gerig/CS6640-F2012/Materials/pseudoinverse-cis61009sl10.pdf

This is done by minimizing the squared error of an overconditioned system of equations.

if you do not want any color profile, you need to set params.output_color=0

With this setting image to be white balanced, demosaiced and scaled to fill 16-bit range.

noise reduction and highlight recovery are not enabled by default

Here is the sample of AHD maze: https://picturecode.cachefly.net/photoninja/images/demosaic_before.jpg

Your blue dot artifacts are very different from that.

Found it. Mea culpa.

I've added the path to the old libraw 0.17 include files. Removed it and it now works.

But awsome how fast you did reply! :)

Sorry, no MSVC 2015 on hands, just tried with MSVC 2013:
call "C:\Program Files (x86)\Microsoft Visual Studio 12.0\VC\vcvarsall.bat" x64
nmake -f Makefile.msvc

After that:
mem-image -4 -6 filename.cr2

filename.cr2.ppm (16 bit, linear) was produced, looks OK

http://www.libraw.org/docs should be updated then to remove the LibRaw license.

WB settings does not change raw data, so you may use any in-camera setting.

And if data is scaled to 8 bit somewhere, your result looks correct.

coloir.maximum is in RAW data domain, so for 14-bit camera it is about 16000.
imgdata.image[], after processing, is in 0..64k range.
your luminocity calculation, in first assumption, is in same range as imdata.image, so 64k.

So your jpeg generation code will show upper two stops of image as 'overflow'

The data array is set in my first example. It is just an approximation of the luminosity of each pixel. (2*red+3*green_blue)/6.

This link should show you the output of the above as jpeg files. You can see the integer overflow in the overexposed clouds https://1drv.ms/f/s!AgLCmsxNxVV_jelOf_MIR7MeGcsXvQ

Hmm, It seems to me that processed data are not being scaled to the 0-65535 range. In fact the data that are not saturated fall into the range 0 - colour.maximum, only saturated values are larger than color.maximum.

Is there a way I can upload an image to the forum? I have generated a greyscale jpeg from the data vector in my code above using

std::vector<unsigned char> jpegData(data.size());
for(size_t i=0; i<data.size(); ++i)
	jpegData[i] = unsigned char((data[i]*255)/iProcessor.imgdata.color.maximum);

The ouput shows the buffer overrun from conversion to unsigned char very clearly.

Phil

Thank-you Alex
I have just loaded up an image and I see the WB_Coeffs array. Elements 1, 3, 4, 10, 11 and 14 are filled, which presumably correspond to the 6 options of my EOS 40D (daylight, cloudy, shade, tungsten, fluorescent, flash). That's everything I needed.

Thanks for the excellent info.

Phil