The retarget landed

RDNA2, gfx1030, 64-lane waves. Every suite green except one device test, described at the end.

The family rows came from the published reference - "RDNA 2" Instruction Set Architecture: Reference Guide, AMD document 70648, cited in REFERENCES.md. Four values had moved: the long-form vector format to 0xD4000000, scalar memory to 0xF4000000, export to 0xF8000000, and interpolation to 0xC8000000 - out of the slot the long form now occupies, which is why every long-form arithmetic instruction had been decoding as an interpolation.

Surprises

• The document is wrong about one field, and the solver caught it. Its field table puts the LDS opcode at bits [24:17]. Under that, ds_read_b32 decodes as 108 and ds_write_b32 as 26; the document's own opcode table, forty pages earlier, says 54 and 13 - which is what [25:18] gives. orbistoun-gen encodings had solved [25:18] from assembled bytes before the document was opened, and the disagreement is the only reason anyone checked the document against itself. The experiment corrected the reference, which is not the direction that pairing was built for.
• This generation's long form can carry a literal; the previous one's could not. All three of its sources sit in the second word, and the literal check only ever looked at the first, so every such instruction decoded four bytes short. OperandField now carries which word it lives in.
• A skipped fixture left the previous generation's bytes on disk. One source stopped assembling (the buffer formats were unified into one BUF_FMT_* enumeration), the generator reported the skip, and the stale .txt stayed - so the differential test compared this target's decoder against last target's output and reported the decoder as broken. The generator deletes the output now.
• Six tests hard-coded an encoding, and every one of them failed after the retarget while blaming the code under test rather than itself. A truncation test held a scalar load, a block-splitting test held one too, a pipeline test held an export, the self-consistency test held four words, and the whole agreement generator held every family and every opcode number. All of them ask the table now. A test that writes down what it is testing stops testing it exactly when it matters.
• Operand code 125 names a register here and was a hole before. It arrived as <unmapped:125> rather than as anything obviously wrong. Reads return zero and writes are discarded, and v_div_scale_f32 uses it for a result it does not want.

Outstanding

the_two_models_agree_on_generated_programs faults the process (STATUS_ACCESS_VIOLATION) partway through its run on the device. Not the retarget, and not a bad program: it stops after 43 programs from seed 1, and seeds 44-91 all pass when run first. So it is cumulative state across roughly 86 dispatches rather than anything a particular shader does - a resource the Vulkan runner is not releasing between dispatches is the obvious suspect. Deliberately not chased by re-running it; the next step is to count live allocations across dispatches rather than to bisect on a device.