"Turn your head" means turn your head 45 degrees, and look at the half of the squares you haven't colored green. You see a board of uncolored diamonds. Recursively checkerboard that. Like Huffman coding, green gets one bit, red and blue get two.

In math we can barely understand each other. It is often easier to reverse-engineer code than to read someone else's code. Go look at the picture on Wikipedia, ask yourself how you'd describe it to someone in an elevator, then read my description again?

Equally distributing RGB cells would be less efficient. Not equally distributing RGB cells would be more efficient. We're saying the same thing.

The copper works out easier, not using a honeycomb. In the 1980's I imagined that by now we'd have compilers that could generate a custom circuit for every programming task, and computers would be whiteboards capable of directly simulating any circuit. This would be the primary mode for all computation. On one hand, programmable gate arrays approximate this, but they're only used when one needs that flexibility, as they're not efficient enough to replace general purpose CPUs. Someone on a plane set me straight. Again, one could give an entire course explaining the answer, but the gist is the same: it's the copper. Fewer traces.

My Dad already had in mind post-processing algorithms, and computers in the 1970's were quite primitive. Ask someone to program on a honeycomb at a stressful tech interview, and they won't get the job. Plain arrays are easier, more efficient. Now computers are faster, but the processing has moved into the camera.

Our fixation on rectangles is a cultural conceit. If photographs were hexagons, we might prefer honeycomb sensors. Silicon wafers are round; I'm surprised that the largest telescope camera sensors aren't hexagons; they're piecing together photographs, not putting them up on a wall for display. Yet they piece together square sensors, use a square grid. Huh.

I coded up a Bayer sampler just now based on the Wikipedia page you linked, that was fun. One simple mapping I rediscovered for determining the relevant RGB/BGR channel of a given (x, y) pixel was, in C, (x & 1) + (y & 1).