If we wanted a huge DTE table and a "DTE-for-DTE" table, would you suggest using a switch byte indicator (to tell the game to start reading from a different table of the next letter isn't part of the current table) utilizing multiple character/DTE tables?
No. It's much simpler. In your case where 50 non DTE character are used, there would be two 205 character tables, one for the left character and one for the right. It's exactly like regular DTE. The only difference is that those tables are allowed to reference themselves, hence my term for "recursive" DTE.
Isn't that basically just a wacky-coded dictionary compression?
More people need to learn Huffman.
My claim that it is the most efficient does not come unsourced. I have written a tool
that performs many different compression algorithm to the same data, and I just noticed that for text, most of the time recursive DTE was the most efficient.
I really wish my great efforts to implement more advanced algorithm like Huffman and LZ77 weren't wasted, but really, recursive DTE is more effective. Huffman often comes 2nd for text data. The large size of the Huffman table, as well as Huffman's inability to compress words that comes really often, is probably why it doesn't perform as good as expected. Huffman can only encode on symbol probability, however in text, the order of letters appearing is as important than the letters themselves when it comes to redundancy.
That and Huffman is much more complex to code, needs more RAM, etc, etc....