No, it takes multiple photos with the same settings and just averages out the noise in the shadows. That wouldn't keep the shadows from lacking detail though. You'd just get a smoother black!
You're correct that it's taking multiple photos with the same settings, but the trick is that it can choose different settings than it would have for a single photo. So instead of a single medium-ISO shot that loses all shadow detail but minimizes noise, it could take multiple high-ISO shots to capture the shadow detail and then average them to minimize noise.
So if you think of your dynamic range as being limited by the noise floor and the overexposure ceiling, this technique lowers the noise floor while keeping the ceiling right where it is, giving you a better effective dynamic range. ISO bracketing extends both the floor and ceiling by a much larger margin, of course, but its has issues that complicate its use in automated HDR shots, as described in the article.
If the pixels were literally black that would be the case, but even on very underexposed pictures I don't think that ever happens. There's still photons on the sensor, it's just that they are swamped by the noise. Like, most DSLRs nowadays capture 12-bit values, which suggests you would have to underexpose by something on the order of 12EV to truncate the output to zero.
Yes, and that is very, very possible. In any case, doing that would give you some relatively painful quantisation error, which is why the HDR approach works better.
With a tripod to stabilise a camera, is it possible to get more detail through multiple shots? Just averaging the noise would smooth things out, but I wonder if there is any more intelligent processing that could improve things. More shots could (in theory) provide more data.
A similar problem might be to try and improve the resolution of a single hand-held photo, shot in normal lighting conditions, by taking many pictures in rapid succession and then processing them. Each picture would be in a very slightly different position and so (again, in theory) might provide more data?
No, you can only get cleaner data and maximize the pixels you have, not more effective pixels. The way you could get "more" data is to simulate, say, a 28mm equivalent lens with, say, multiple shots from an 135mm equivalent lens at different locations and stitch them together.
Thats what they do when you see those 3GP+ shots you can zoom in on.
You an absolutely gwt subpixel detail by averaging multiple photos taken with the same setting with almost exactly the same frame. I had software in 2004 that did this, converting short video to high res photo
So if you think of your dynamic range as being limited by the noise floor and the overexposure ceiling, this technique lowers the noise floor while keeping the ceiling right where it is, giving you a better effective dynamic range. ISO bracketing extends both the floor and ceiling by a much larger margin, of course, but its has issues that complicate its use in automated HDR shots, as described in the article.