Keep dense GGUF rows through GLU and RMSNorm fusions

[kvark]

Stacked on #193; ordinary command recording only.

• Keep dense F32/F16 GGUF projections in native [N,K] rows and use MatMulBT. Block formats retain their existing layout.
• Reuse the transposed GEMV shader for fused residual adds and preserve RMSNorm folding. Both forms enter the existing qualified workgroup-width/reduction search.
• Share gate/up packing across orientations, SwiGLU/GeGLU, and greedy/egglog rewrites. Row updates preserve current destination bytes, including unaligned f16 ranges.
• Use a portable two-dimensional dispatch grid for wide output vocabularies. This respects B570's 65,535-workgroup X limit without changing the common-case geometry.

No device/model thresholds, expanded tuning budget, precision changes, or command-buffer replay.

Matched ordinary-recording results

Source 69bbb48450fee8f9957b682ffec95733ca50c1f4; control 861d87b9a85f666a5a131f82d07c0b66d9131e16 includes attention tuning. Median of three fresh-process medians, with rotated engine order:

GPU	Phase	Control (ms)	This PR (ms)	llama.cpp (ms)
RTX 5070	Prefill	8.777	8.479	7.087
RTX 5070	Decode/token	3.002	2.514	1.385
Arc B570	Prefill	21.337	21.251	15.610
Arc B570	Decode/token	5.561	5.129	3.425

Decode improves in all three trials: approximately 16% on NVIDIA and 8% on Intel. NVIDIA's new process medians span 2.475–2.748 ms; Intel's span 5.107–5.133 ms. Decode remains 1.82×/1.50× llama.cpp. CPU/GPU clocks were not fixed.

This is one SmolLM2-135M sequence with 128-token prefill and 32 cached decode steps, including full-logit host copies, not batched-throughput evidence. Both sessions visit all seven eligible tuning classes, qualify all comparisons and finish within their existing budgets. All 18 saved logit sets are finite and retain 33/33 independent CPU-reference next-token choices; this is a fixed-token numerical check, not language-quality evidence.

Protocol, ranges, preparation times and limitations.

Verification

452 CPU tests pass. The full 22-test GGUF suite and 17-test GEMV group pass on both NVIDIA and Intel, including normalized tuning, both activations/storage types, source restaging, odd f16 dimensions and a wide-vocabulary grid. All-target/all-feature Clippy, formatting and rustdoc with warnings denied pass. No new test executable or relaxed tolerance. The known Naga Workgroup ArrayStride warning remains unchanged.

CI for the measured code: all six jobs passed, including Linux coverage, macOS, Windows and MSRV. Rust host line coverage is 83.6% overall (WGSL execution is not instrumented). The final follow-up is documentation only.

No raw data or binaries in Git. The submitted paper and frozen Inferena cohort are unchanged.

Rebased after #192 merged into main. The file tree is unchanged from the pre-rebase head; the measured revisions remain on experiment/llama-gemv-2026-09-19. Fresh post-rebase CI is running.