You missed the last part of my post where I suggested is might be the same sensor die as the 5D MkIII, but with a greater number of pixels failing to make the grade. These would then be averaged together as a single pixel. If you do that across a sensor you get very respectable performance but with a lower total MP count.
That wouldn't work out, you'd just have a ton more hot pixels and I've never heard of a manufacturer calling poorly made 22MP sensors that have 2 million dead pixels a 20MP sensor. And just think, they list the image dimensions, how do you get that change with random dead pixels here and there throughout?
In sensor production, it is not all or nothing when it comes to pixels. Yes, there are dead pixels and hot pixels, but that's not what I'm talking about. More common problems are noisy pixels, non-linear pixels and low ISO pixels. Very often these can be corrected using the average value of a group of adjacent pixels, rather than rely on just that one. Or by comparing them to nearby pixels and using an offset. Sometimes the only solution is to ignore the output. Typically, these problem pixels are not uniformly distributed - they tend to cluster, while the rest of the pixels may be perfect.
Since you can't use a sensor that is high res everywhere except the lower right corner where the problem pixels are, any solution must be applied uniformly to the whole sensors. Even if it means ignoring or averaging the output of perfectly good pixels.
Each sensor produced is tested and the response of each pixel measured. If a model of averaging, offsetting or ignoring problem pixels is found that will produce a lower resolution but perfectly performing sensor, the model is a applied and the sensor is used at the lower MP count. This is how many small, compact camera sensors are produced. I also think it is where the SLR AF sensor come from.
It is not necessary to apply the same correction model to every problem pixel on a given sensor as long as you end up with a uniform distribution of good pixel clusters. On a single sensor, there may be a mix of non-linear pixels, noisy pixels and pixels that only respond at high or low light levels. As long as the corrections gives results in a consistent and uniform pitch, say groups of 2, 3 or 4 pixels, it can be tailored to correct each problem separately. Various correction models can be tested mathematically to find the optimal solution - cleanest output, highest MP count. The results are stored in sensor correction array map that is ROM on the chip.
This method has limitations. Too many problem pixels in one area or across the sensor, and all correction models fail - the sensor is discarded. But is is quite possible to start with a sensor that won't pass at 24 MP but with correction, will work at 12 MP. Or start with 40 MP and get 20 MP. Not all the corrections are simple averaging, so the relationship between the starting point and ending point is not always a factor of 2 or 3.