11. These sub-channel ISAs are themselves designed to be as space-efficient as possible while still being able to be directly executed without any kind of pre-transformation. They're often variable-length, with most instructions being single-byte. Opcodes are hand-placed into the same kind of bit-level Huffman trie you'd expect a DEFLATE-like algorithm to design if it were tasked with compressing a large corpus of fixed-length bytecode. Very common instructions (e.g. a brush to draw a horizontal line of a particular metatile across a page up to the page boundary) might be assigned a very short prefix code (e.g. `11`), allowing the other six bits in that instruction byte to to select a metatile to paint with from a per-tilemap metatile palette table. Rarer instructions, meanwhile, might take 2 bytes to express, because they need to "get out of the way of" all the common prefixes. (You could think of these opcodes as being filed under a chain of "Misc -> Misc -> Etc -> ..." prefixes.)
IMHO, these are all (so far) things that could be studied as generalizable data-compression techniques.
But here are two more techniques that are much more specific to game-dev, where you can change and constrain the data (i.e. redesign the level!) to fit the compressor:
12. Certain ISAs have opcodes that decode to entirely-distinct instructions, depending on the current states of some modal variables! (My guess is that this came about either due to more level features being added late in development after the ISAs has mostly been finalized; or due to wanting to further optimize data size and so seeing an opportunity to "collapse" certain instructions together.) This mostly applies to "brush" opcodes. The actual brush logic they invoke can depend on what the decoder currently sees as the value of the "level type" variable. In one level type, opcode X is an Nx[floor distance] hill; while in another level type, opcode X is a whale, complete with water spout! (In theory, they could have had an opcode to switch level type mid-level. Nothing in this part of the design would have prevented that; it is instead only impractical for other reasons that are out-of-scope here, to do with graphics memory / tileset loading.)
13. And, even weirder: certain opcodes decode to entirely-distinct instructions depending on the current value of the 'page' or 'column' register, or even the precise "instruction pointer" register (i.e. the current 'row' within the 'column'). In other words, if you picture yourself using a level editor tool, and dragging some particular object/brush type across the screen, then it might either "snap" to / only allow placement upon metatiles where the top-left metatile of the object lands on a metatile at a position that is e.g. X%4==1 within its page; or it might "rotate" the thing being dragged between being one of four different objects as you slide it across the different X positions of the metatile grid. (This one's my favorite, because you can see the fingerprint of it in much of the level design of the game. For example: the end of every stage returns the floor height to 2, so that "ground level" is at Y=13. Why? Because flagpole and castle objects are only flagpole and castle objects when placed at Y=13!)