[drakestapleton]

The combined 128-row/group4 setting preserved every output-score bit and next-token choice in all six frozen 27B reading cases. The repeated 5,017-token read took 18.817 seconds with the 64-row/group0 comparison and 14.348 seconds with the selected combination. The 128-row/group5 alternative took 15.379 seconds. These are shared-machine diagnostics, not quiet-machine qualification; production defaults remain unchanged.

Omega owns the bounded launch grid and complete coordinate-remapping instruction plan. The existing emitter interprets those primitive instructions. This restores the work-order mechanism examined in historical Omega 86211a399 without changing arithmetic or the 45-cycle matrix wait. Runtime API 0.10.0 adds an experimental setter. Stacked on #257; its cache-budget correction is included.

Validation: the initial work-order source passed all 78 host targets and the API version gate, 137 independent plan cases, five Omega source mutants, matrix simulation, and default-code byte identity. After merging the cache correction, targeted cache admission, matrix simulation, Omega fault tests and API version checks passed again. The complete combined host rerun passed all 78 targets and the API version gate; see om/prefill-shape/order-combined-full-host-validation.json. Full-model scores match for the tested 64-row group0/3/4/5 and 128-row group0/4/5 settings.

Source/binary identities and complete comparisons are in om/prefill-shape/order-combined-validation.json and order-validation.json. Per-model serving integration, full serving qualification and five quiet alternating pairs remain. Coordinated Ribbon evidence is in aien-sovereign-core-mv#898. No deployment or performance qualification is claimed.

Apache-2.0 WITH LLVM-exception notices retained; no new production dependency.

[aien-spark]

Merge desk (AI), 2026-11-02. Landed on main as part of the reviewed Ribbon stack through #260 (merge 183fa4f). This head (a73ab65) is contained in main. Closing as landed.