[aien-spark]

The corrected 128-row candidate preserves every output-score bit and next-token choice in all six frozen 27B reading cases. Its repeated 5,017-token read took 14.759 seconds versus 18.817 seconds in the 64-row comparison. These are shared-machine observations, not quiet-machine qualification. The 64-row default stays unchanged.

Checked Omega admission now uses the exact final store-offset bound and permits the experimental 128-row shape. The kernel opts into the existing extended register budget. The first full-model trial exposed an older cache-admission cap that rejected this declaration before launch; commit 572a73b makes the cache honor the explicit budget while retaining the default cap and hardware checks. The real cache regression fails against the old library and passes the correction, including missing-opt-in and over-budget refusals.

Validation: the earlier candidate passed all 77 host targets, matrix simulation and nine Omega source mutants. The corrected cache path passes the targeted suite and full-model exact-score comparison; the corrected full host rerun passed all 77 targets. The command and log digest are in om/prefill-shape/cache-full-host-validation.json. Original 64-row kernel bytes and resources were unchanged in ten representative cases. Evidence is in om/prefill-shape/large-shape-gpu-validation.json, cache-admission-regression.json and the earlier host receipts. Stacked on #252; work-order combination is #258. No merge, deployment or quiet 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 (2089a92) is contained in main. Closing as landed.