a. I tried to compress with higher lever and got a same result
b. If I have few alphabet symbols (less then 64 (6 bits Huffman length) - I took only 2!!) - Huffman code for every symbol takes less the 8 bits even when
I have a uniform probability each of them. so, for big enough literal stream we get a better compression then raw (even we add a tree).

Originally posted by @shulib in #3203 (comment)