I have never seen old CGA captures, so I don't know for sure about it.
I do actually have an old CGA and a capture card, so I'll have a go at creating some.
The captures are at http://www.reenigne.org/misc/old_cga_captures.zip . I used my chart.com program, so each screen shows a different graphics mode (the port 0x3d8 and port 0x3d9 values are in the top left) with different patterns in different columns and different palette register values on different rows.
* I adjusted the brightness and contrast so that "black" (i.e. CGA colour 0) is about 27,27,27 and "white" (CGA colour 15) is about 233,233,233. This way we can be sure that the colours aren't being clipped. Be sure to compensate for this when calibrating algorihms against these images. All other controls were left at default.
* My capture card is rather useless in the color-burst disabled modes. For one thing it doesn't have a color killer circuit, so it displays colour in these modes, but it uses the wrong color burst frequency (off by ~1.3KHz giving a pattern that repeats every 12 scanlines). It also adjusts the gain to attempt to normalize the (non-existent) color burst amplitude - i.e. turns the gain right up to maximum, causing white to be clipped, black to be much lighter than it should be and some of the colours to be wrapped. My TV has a similar extra gain in these modes but does turn off the colour. Also these modes have a vertical offset for some reason - perhaps the same gain adjustment is causing it to become confused about whether to trigger the vertical sync on the rising or falling edge.
* There's a bright scanline near the bottom in the color burst enabled images - I'm not sure why at the moment. The software isn't doing anything there.
* The capture card is set to 640x480 mode which captures two rows of pixels for each scanline. Because it uses the correct aspect ratio, that means that each CGA pixel corresponds to about 7/8 of a pixel horizontally in these images.
* There's some 7MHz signal that is not being filtered out (meaning that this capture card is using a notch filter rather than a bandpass filter for some reason). There's also some noise at about 1.7KHz giving a faint repeating pattern every 9 scanlines - not sure where that's coming from.
* Most of the colours seem to be pretty close to prediction except that the saturation is lower on the real hardware. The TV output seems a bit more saturated (compare the photos at http://www.reenigne.org/misc/cga_composite_calibration.zip), so this could be a difference between hardware and software decoders.
Unfortunately I won't be able to do any more tests on real hardware for a month or two (I'm in the process of moving and my XT is on a ship somewhere).