That's right, it is fun to make a per-channel alpha blender and play with it (you can simulate in OpenGL by doing one pass per channel with a channel mask). However, it is just a weird trick that doesn't get you closer to physics, and doesn't do anything particularly intuitive with respect to computed color spaces.
Because our digital color systems encode perceived color and brightness rather than spectral energy distribution, you cannot model things like selective absorption or re-emission, the basis of real-world pigments and filters. While you might think it works after thinking about the simplistic "red glass" filter, it doesn't really work for different lights and glasses we would perceive as red but which are really mixtures of different wavelengths.
Think of each pixel as a 2D curve of energy over wavelength at this point in the image plane, representing the entire cone of space behind it. Instead of a multi-channel blending function, you would have a transfer function that changes the arriving 2D curve into another 2D curve. However, some light sources and pigments often have very narrow emission lines or notch-filtering behaviors, best represented as some sort of variable set of discrete wavelength intervals, while other black-body kinds of radiator have broad spectra best represented as a smoothed curve. Any lossy compression scheme would likely destroy the physical properties you are trying to simulate, unless tailored for the specific lights/materials/effects at play along the light path.
A more likely physical simulation approach would be something like monte carlo sampled ray-tracing or photon-mapping, reevaluating the scene with a large number of discrete wavelength-energy bundles, giving you some dithered and accumulated final result in the image plane. As an added benefit, things like refraction would really work right, so you could shine a white light through a prism and get a rainbow out the other side, based on the different wavelength-specific interactions of each sample with the materials in the scene.
The conversion from spectral energy into perceived color would have to be delayed until the final accumulated spectral energy plot is available at the image plane. You could imagine either a framebuffer with a 2D curve at each pixel, or perhaps a stack of hundreds or thousands of monochromatic layers each representing one sampled wavelength. Things like absorption and re-emission at different wavelengths would add complication, since a particular "ray" could change wavelength as it traverses the scene.
Of course, given all that, we might still wonder why we cannot simulate polarizing filters, iridescence of nanostructures, or constructive/destructive interference of light... that would require an even more complex framebuffer!