The retarget was not finished, and cargo had been hiding it

Reported the retarget green. It was not: cargo test stops at the first failing test binary, agreement was failing ahead of execute, and execute.rs had not run at all. Sixty failures were behind it.

All sixty were the same thing as the six found earlier - fifty-two encodings written down in a test file - and all of them now go through one head(name) helper that asks the table. The whole GPU side is green: 210 tests.

Surprises

• The lesson had a bill attached and I paid it twice. The six tests fixed during the retarget were the visible ones. execute.rs is the largest device test file in the project and it never compiled its way into a run, so the same class of bug sat there looking fixed. Read which binaries actually ran, not just the last line of output.
• the_supported_list_and_the_translator_agree carried its own epitaph. Its hand-written encoding list had a comment saying it had already failed to catch the drift it exists to catch, because the list of encodings it checks is also written by hand. It is derived from the table now, and it covers every name the target has rather than the dozen someone thought of.
• A supported instruction can be refused, and that is not drift. Synthetic encodings with zeroed operand fields make the carry arithmetic name an ordinary register pair, which the translator legitimately refuses. The test asks why it was refused now, and the reason is a shared constant rather than matched prose.
• The decoder was stopping at the wrong end of the shader. decode_program stopped at the first s_endpgm, but a shader ends its wave once per exit path and the compiled control-flow fixture has two - so it was truncated in the middle, a branch then pointed at an instruction that was no longer decoded, and the report blamed the branch. What actually ends the code is s_code_end padding, which the reference describes as an illegal instruction placed past the end so a prefetch faults rather than executing memory. The decoder recognises padding now, in both modes.
• PROGRAM_END_OPCODE: u32 = 1. The same hard-coded-opcode bug, in the decoder itself, surviving the retarget only because s_endpgm happens to be opcode 1 on both generations. By name now.
• s_clause is real and does nothing. The reference: define a clause of instructions which are executed together. It groups the following instructions for issue and computes nothing, so it translates to nothing - a translation, not a shortcut, and cited as such.