My opinion is that this is a machine vision problem in which you should be controlling the lighting and have a good idea of the maximum brightness of a non-glare pixel brightness in the image. Defect detection is generally a machine vision problem rather than a computer vision problem.
What we see as a result of lighting is an addition of specular and diffuse reflections of light (plus some emittance but its negligible here).
The specular component is the glare, on shiny surface like this apple, it is much more than the diffuse reflection (>10x)
This means that if you setup your lighting, gain and exposure prior to this, on a diffuse surface, you can be sure that nothing will be even close to saturated. So using a fixed threshold is actually the preferred solution here, as long as you've proven with enough data that "no pixels not containing glare" would be above the threshold. In essence you are setting up the lighting conditions, and camera parameters such that classification of a pixel becomes trivial, in this case performed by a simple threshold, rather than a more complex machine learned function of pixels around it.
I like "vini"'s approach, no real need to show the RGB planes. Just a simple grayscale threshold would actually work here.
1- you design the lighting conditions, not ambient
2- make the classification job extremely trivial (thresholding)
3- measure the feature
4- compare to tolerance