First, a bit of terminology: There’s no gamut in ‘black’. Black is the absence of light energy and gamut is about the ability of rendition media to resolve colors of extreme energy, at the other end of the energy continuum, so to speak. What you’re probably concerned with are negative values, which are the product of manipulation.
The data right out of your image has no negative values; indeed, the minimum value of any channel is about 147. The interesting thing I found is in the NEF metadata, the MakerNote tag BlackLevel contains four values, all 600. If you subtract that from each pixel, you drive most of the meaningful data in the image below 0, given that the maximum value for this camera with it’s 12-bit sensor is 4095. Libraw delivers a more sensible value, 150, which contributes to a “more colorimetric” image.
Assuming negative values are clipped to 0 by black subtraction, I found a few subsequent operations that would drive values negative: some demosaic algorithms, lanczos resize, and convolution sharpen. The all affected only an insignificant number of pixels, my export conversion clips all such to 0 and the image looks fine.
Also, this image responded to a rather tame filmic curve. I can’t comment about the darktable filmic; I started with a simple normalization of the linear data to 0.0-1.0, and I only had to boost the B param of rawproc filmic to 26.2 (from 16.2) to get this, in rawproc:
I’m posting the 800x600 proof to demonstrated the value of rendition resizing in reducing noise…
This is a really good demonstration image for all of the raw-processing fundamentals; with your blessing I’d like to use it in a demonstration video…