I wonder if thinking about "tweaking things until the histogram looks right" is the right approach -- in some ways, the idealization is that there are 3 separate R,G, and B images. If there is a lot of leakage between the channels, it's actually information lost, no? But perhaps that's OK.

Custom white balance adjusts the relative exposure between those 3 color channels, and I'd hope that those sliders you're using do the same, in perhaps a constrained way. If we have a standard way to recompute the channel balance (which is to say, a standard deviation between channels... is that even possible without a reference?) we could add that as an automatic step in the processing, and it'd be more consistent, i hope.

But in images where there really isn't any photosynthesis going on, that might artificially inflate NDVI values, no?

"Tweaking things until the histogram looks right," and especially "tweaking things until the NDVI looks right," are exactly the approaches that must be avoided. Such tweaking will have to be part of establishing a system (of hardware and a processing protocol) that produces biologically meaningful NDVI images. If that system is then used consistently, it could provide useful information about plant health.

Overlap among channels probably produces artifacts that hide the desired signal, and in ways that are not easy (for me) to predict or identify. So that's a nasty issue.

Recreating what a camera custom white balance does is a mystery to me at the moment. My instinct is that if a bunch of different cameras take infrablue photos with a bunch of different white balance settings (not to mention exposure, filter, light source), it will be hard to automatically tweak those photos so that the NDVI derived from them will be comparable. Although if all the photos had the same (multicolored?) white balance card in them, it might be doable. But quantitative (or even relative) comparison of NDVI of different plants under different light from different cameras at different times is probably not an attainable goal for the Kickstarter camera.

The first goal should be a system that allows one infrablue photo with two different plants in it to produce NDVI that consistently represents a health difference between the two plants. The second goal is a system and protocol that allows NDVI from two separate photos from the same camera to be compared. Comparing NDVI from two different kinds of camera will require some careful control of variables.