Opcodes as shown on original opcode page seem to require quite a bit of complexity in decoding stage, as well as synchronous register stores and some extra complexity for dealing with multi-byte instructions (immediate values). I think I found a way to remove them altogether, without sacrificing too much code density, at a cost of some opcode space.
There can be just one instruction type accepting immediates: MOV reg, imm. Normally this would
be one two-byte opcode: [MOV|reg][imm]. However, we can do this in the same two bytes, but
consisting of two separate instructions: XXX imm4, 0, XXX imm4, 1. Opcode name is not yet decided,
hence temporary XXX. imm4 in the above is a nibble of immediate to be moved, and the other
argument can be one of 2 values: 0, 1. Every other execution of this instruction (as controlled
by a single bit flag) will do a different thing. The first one will store both arguments to B
register (using 6 of its bits). The second one will do the actual store to the target register.
The immediate to be stored is just a combination of the value in B and current argument.
The target register in turn can be calculated from the 0, 1 part. Two of these give
4 combinations, enough to encode 4 targets.
This will require 32 of 256 opcodes.