Take for example Haskell. It's a functionnal, typed language. In Haskell, at compile time, the compiler analyzes all the types of all your functions and if they all match, it drops them completely. There is no type information at all left in a compiled Haskell program, because the compiler can know ahead of runtime if it is correct.

You could, but there isn't much benefit. The purpose of all that extra information is generally to make the program easier to understand for a human. The computer doesn't need any of it, that's why it's not preserved in compilation. So it is possible to automatically translate assembly to C++, but the resulting program would not be much (if any) easier for a human to understand and work with.

To give a bad analogy, imagine some driving directions: turn left at 9th street, enter the highway at ramp 36, go right when you're past the burger king, etc. These are translated into physical control inputs by the driver to actually take the car to its destination. Now we could look at the driver's physical inputs and turn that back into a written list of instructions: turn the wheel left 70 degrees, turn it right 70 degrees, push the gas for 10 seconds, and so on.

All the street name references are now gone. There are no abstracted instructions like "enter the highway" or even "take the second left." It would be quite difficult for a person to look at these instructions and figure out the trip's destination. Let alone make some alterations to it because there is roadwork along the way and a detour is needed.