Both approaches are correct. At least as I understand what I think you are trying to say. I'm not sure, though, what you mean by, "(4 time bigger)".
Raw luminance values are monochromatic in the sense that there is only one intensity value for each sensel (what we call a pixel well). But every one of those monochromatic luminance values is the result of filtering for one of three colors of light. If a pixel is filtered for green light, then a full well value indicates the presence of intense light that contains green. If a pixel well is filtered for red light, then a full well value indicates the presence of intense light that includes a lot of red wavelengths. If a pixel well is filtered for blue light, then a full well value indicates the presence of a lot of blue light.
If the light striking a particular area of the sensor is white, then all the adjacent pixel wells should (at least after the data has been calibrated) return roughly the same value. But when the light striking a particular area of a sensor is predominately one color, such as blue, then the pixel wells filtered for blue light will have luminance values much higher than the surrounding red and green filtered pixel wells.
For more on various demosaicing approaches, please see What are the pros and cons of different Bayer demosaicing algorithms? and Would a demosaic algorithm for black and white be useful?
For what happens when demosaicing goes wrong, usually due to a hot or stuck pixel, please see: What is this bright squarish pattern that I'm getting in long exposure images?