So only a fraction of the 32bit FP dynamic range is actually used. But its
still pretty large and the programmers don't have to worry about integer
overflows anymore.

I'm not sure about the actual number of bits used in this case but due to the nature of floats it's a lot bigger than a fraction, probably around 24 bits of bitdepth.

I assume you also better not run a signal above 0db through a ladspa plugin.
This is also alsa isn't it?

No, ladspa plugins are run internally in the application, alsa is mainly just the drivers for the hardware on a linux system.

But you are right, generally in a digital environment; stay away from 0db is a good rule of thumb. :)