If a lower resolution end image was acceptable, wouldn't it be possible to halve the width and halve the height of a low-light image and take the sum of the light in each block of 2x2 pixels when creating the new (4x brighter) image? After all, if the CCD sensors were 4x as big, they'd surely capture 4x as many photons? I'd expect that if this was done with raw RGB channels it might be as simple as addition, but with a common compressed image format like JPEG the aggregation expression might be different (to accommodate color spaces, etc.) Has this been tried or seen before?
Yes there have been plenty of cameras that have aggregated pixels in order to reduce noise, from my ancient Fuji Finepix s602z, which would output one megapixel images above ISO400 to the very latest PhaseOne medium format backs which offer higher ISO settings in "sensor+" mode but at one-quarter of the number of pixels (half the width and height).
It's important to note that whilst theoretically you get as much light by summing the values of four small pixels, compared to one big pixel, in reality there is usually a small amount of read noise when getting the value so you get 4 times the read noise with smaller pixels. Some sensors perform the addition (known as binning) right on chip prior to readout and digitization in order to combat this.
Fuji has been doing this for years. Matt mentioned the older S602Z but they have actually been working and perfecting this.
Since Fuji introduced their first SuperCCD EXR, they have being doing this with additional circuitry to do this earlier in the pipeline. Fuji uses a special color-filter array which is rotated 45° so that there are always two adjacent sub-pixels of the same color. The latest Fuji cameras use EXR on CMOS.
The design lets them solve two problems, one is low-light performance as you suggest. The other is extra dynamic range because they can read half of each pair of adjacent photosite partway through the exposure, before the entire pixel saturates. That way they get two half-resolution representation of the scene which get blended together to preserve highlight details, up to 4 stops (1600% as Fuji calls it).
If you resize an image (JPEG or other) to a quarter of its size, you are doing a similar thing but when the camera does it, it gets better results from what I've seen. However, the difference is minimal from what I've seen.
In theory the Foveon based cameras (currently only sold by Sigma) do this. The most recent cameras (with the Merrill sensor) have enough resolution to make it practical to half and quarter the resolution while still producing a image with a reasonable number of pixels.
It is a typical approach to reducing noise and it is called binning. In machine vision cameras it is a built in setting and it is done before the ADC, as to improve dynamic range. This way it simulates a larger pixel size. In post it is more normal to reduce noise by having multiple snapshots of the same static scene and averaging them. But it must be a static scene. Binning in post might have its merits and it cannot be achieved through resizing to 50% so it needs its own function.
Here is before and after post-binning in a dark image viewed at the same size and brightness, then everyone can make their own conclusions whether it is useful enough: