As said by John Rennie, it has to do with the shadows' fuzzyness. However, that alone doesn't quite explain it.

Let's do this properly:

![Fuzzy overlapping shadows][1]

I've simulated shadow by blurring each shape and multiplying the brighness values<sup>1</sup>. [Here's the GIMP file, so you can see how exactly and move the shapes around yourself.](http://ubuntuone.com/6Z4kPONPxZ7cGmqGwLw57h)

I don't think you'd say there's any bending going on, at least to me the book's edge still look perfectly straight.

So what's happening in your experiment, then?

_Nonlinear response_ is the answer. In particular in your video, the directly-sunlit wall is overexposed, i.e. regardless of the "exact brightness", the pixel-value is pure white. For dark shades, the camera's noise surpression clips the values to black. We can simulate this for the above picture:

![Nonlinear response of overlapping shadows][2]

Now that looks a lot like your video, doesn't it?

With bare eyes, you'll _normally_ not notice this, because our eyes are kind of trained to compensate for the effect, which is why nothing looks bent in the unprocessed picture. This only fails at rather extreme light conditions: probably, most of your room is dark, with a rather narrow beam of light making for a very large luminocity range. Then, the eyes also behave too non-linear, and the brain cannot reconstruct how the shapes would have looked without the fuzzyness anymore.

Actually of course, the brightness topography is always the same, as seen by quantising the colour palette:

![Brightness isobares][3]

<hr>

<sup>1</sup>To simulate shadows _really_ properly, you need to use [convolution](http://en.wikipedia.org/wiki/Convolution) of the whole aperture, with the sun's shape as a kernel. But for the purpose of answering this question, an ordinary Gaussian blur was quite sufficient.


  [1]: https://i.sstatic.net/JPahR.png
  [2]: https://i.sstatic.net/TjvcS.png
  [3]: https://i.sstatic.net/fOZM5.png