The target generation was never checked

Every table in this thread is derived against gfx900 - the fifth generation of the architecture - because that is what MCPU is set to in the fixture and probe generators. The encoding families, all sixty-two operand layouts, the 126-instruction differential corpus and every mnemonic come from assembling and disassembling for that target.

Nobody checked whether it is the right one. Measured against a later generation:

Probe instructions assembling on gfx900	322
Rejected outright on gfx1030	69 (21%)
Encodings differing in scalar.s alone	62 of 120 (52%)

v_add_f32_e64 is 0xD1010000 on one and 0xD5030000 on the other - a different family value and a different opcode number. v_cmp_lt_f32_e32 vcc, … does not assemble at all on the later target, because its thirty-two-lane mode makes a comparison destination thirty-two bits wide.

This is the deepest version of the pattern this session keeps finding. A table that is plausible, internally consistent, and differentially verified - against the wrong reference. The differential test would pass forever, because it compares against the same disassembler the tables were derived from. Self-consistency is not evidence, and neither is agreement with the thing you copied.

The cost is far below what 52% suggests, and for a reason worth naming. The tables are generated, not transcribed. MCPU is one constant; re-running the generators produces new tables and the differential test re-verifies them. D097 chose to solve rather than transcribe on the grounds that a solved table can be checked; the larger payoff turns out to be that a solved table can be retargeted.

What would not carry over. The translator keys on (family, opcode number) across about sixty match arms. Opcode numbers move between generations; mnemonics do not. Keying on the mnemonic makes the generation a configuration choice rather than something baked into code - principle 5 applied to the one place it was not.

Not resolved here. Which generation the target actually is, is a question for the side of the project holding the binaries. It also decides the wavefront width, which is the central parameter of the subgroup fidelity model - so that model waits rather than being built against an assumption that has just been shown to be load-bearing and unchecked.