Here is a plot of CIE 1931 xy chromaticity for the input “scene-referred” colorimetry on the left, the picture formed by ACES 2.0 Picture Formation in the middle, and the colorimetry of the formed picture on the right.
As you can see, candle light as captured by the RED Camera Observer tends towards the camera observer spectral locus, and is placed quite far “outside” the human observer spectral locus by the colorimetry formation matrix.
All of the Picture Formation approaches you pictured except for Arri Reveal and ACES 2.0 seem to handle this pretty well though, so what’s going on here? It turns out there is a similar strange behavior in reddish-orange hues as there is in blue hues.
If we take a hue sweep of the same hue angle as the skin, we can see a strange region where the hue changes suddenly from reddish orange to yellow and then back to reddish orange:
Here’s a full screen image of the hue sweep so you can see it better:
And here is a plot of the hue as it varies from left to right. The increase 3/4 of the way across the plot is the hue distorting towards yellow and then suddenly distorting back towards red.
In 3 dimensions, we can see the same hue distortion. Note that the behavior changes over the intensity range. Higher intensity actually has the inverse behavior.
All this leads to a perplexing and unintuitive behavior of the Picture Formation, where if you increase the saturation of the input image data, the perceived saturation of the image is reduced due to the hue shift towards yellow.
Perhaps this is one aspect of what Daniele was warning about when the ODT VWG first started to go down the path of using a human observer color appearance model in their approach?
It’s all fine though, this image is “pathological” so we don’t need to worry about it 
Here is the nuke script used to generate the above images if anyone wants to take a look for themselves.
devils-in-the-details.nk (596.6 KB)