Nice trick, in fact with 4 instructions it's as efficient as extract/insert and it works for all ADD/SUB/OR/XOR/CMP instructions (not for AND), except if the source is a high-byte register. However it's not really a problem if code generation is not great in this case: compilers in practice will not generate accesses to these registers, and while old 16-bit assembly code has lots of such accesses it's designed to run on processors that ran at 4-20 MHz.

Flag computation and conditional jumps is where the big optimization opportunities lie. Box64 uses a multi-pass decoder that computes liveness information for flags and then computes flags one by one. QEMU instead tries to store the original operands and computes flags lazily. Both approaches have advantages and disadvantages...