2026-08-19 - The worklist starts moving

Done. Closed the loop the whole project is built around, then walked one step down it.

• predicated::SUPPORTED - the instructions the translator handles - is public, and instruction consults it before dispatching, so the list is authoritative rather than a second opinion.
• orbistoun-cli shaders asks the translator what it supports instead of assuming nothing.
• Three more instructions, each a TDD cycle: s_waitcnt, a scalar register file, and s_mov_b32.

Verified. Gate green, device tests executed. Seven execution tests pass on hardware. The worklist moved from 0 to 45 of 106 instructions translatable on four instructions, and now ranks real work:

      5    6  global_store_dword  (FLAT:0x1c)
      4    5  s_load_dwordx4      (SMEM:0x2)
      3    3  v_add_f32_e32       (VOP2:0x1)

Surprises.

• Four instructions covered 42% of the corpus. s_endpgm, s_waitcnt, s_mov_b32 and v_mov_b32 are apparently what shaders are mostly made of. Worth remembering when the remaining count looks discouraging: instruction kinds are distributed nothing like instruction counts, which is the entire argument for ranking by shaders blocked rather than by frequency (D086), arriving from the other direction.

• The agreement test caught an error in its own fixture. 0x7E00_0080 was written meaning v_mov_b32 and is opcode zero - the opcode bits sit at 9 and were left out. The test reported that the translator refused something the list claimed, which was true and for the opposite reason to the one suspected.

• Two instructions that legitimately emit nothing tripped match_same_arms. Merged into one arm with both reasons written out, rather than split to satisfy the lint - splitting would have implied the bodies differ.

Not done. Still no arithmetic, no memory access, no control flow. The next three blockers are memory - global_store_dword and the scalar loads - which need guest memory modelled as a second buffer, and float arithmetic, which needs the register file bitcast rather than treated as u32.

And it remains unmasked: register writes ignore the execution mask, because nothing diverges yet. That is the largest unbuilt part of D098 and it wants doing together with control flow.