Maxwell Render has a feature called multilight, where one can ray-trace a scene with multiple light sources, and change the brightness of the individual light sources afterwards. Luxrender has a similar feature called light groups.
I image that behind the scenes, these renderers keep track of the contributions of the different light sources separately, and then composing a final image (given brightnesses of light sources) just amounts to mixing these per-light contributions with appropriate coefficients. Effectively, instead of keeping track of the color of a ray as a tuple (R,G,B), one keeps track of more numbers (R1, G1, B1, R2, G2, B2, ..., Rk, Gk, Bk) if there are k light sources (let's assume no spectral rendering, just for ease of notation -- it's conceptually identical).
From a mathematical point of view, the final image is a linear combination of the images resulting from the individual light sources.
Now it seems interesting to have a similar feature that allows changing the colors of objects in the scene after rendering. This can be done with an entirely similar technique: suppose that we have two objects whose color we want to be able to change after rendering. The final image is a polynomial in the color of these objects, with two variables (one for each object) and of degree n, where n is the maximum number of bounces for a ray.
Specifically, instead of keeping track of tuples (R,G,B), the integrator would keep track of expressions of the form (R00,G00,B00) + (R10,G10,B10)x + (R01,G01,B01)y + (R11,G11,B11)xy + (R21,G21,B21)x²y + ..., for a total of k^n terms, where k is the number of colors you want to be able to change, and n the maximum number of bounces. For modest k and n, this certainly seems tractable. Color calculations become more expensive, as does storing the resulting "polynomial-image", but all the geometric calculations are identical.
Is this feature implemented in any widely available renderers?
If so, which ones?
If not, why not? Is this not something that's useful in practice? Is it not computationally feasible?
I hope I've made myself clear in my explanation. I don't know much about computer graphics (mathematician by training), so my notation may be unusual. More general that just changing the colors of objects, one could change more properties of materials after rendering. For example, the transparency of objects could be changed afterwards (though assuming importance sampling is used by the integrator, this comes at the cost of noise, I guess).