fix: harden Lunar Lake Gemma 4 attention and quantized CPU copies

[marksverdhei]

Hybrid Gemma 4 attention could route a transposed quantized V cache to a CPU copy that wrote beyond its output allocation. Reject unsupported quantized copy layouts before scheduling or direct/planned CPU execution, and correct row-byte counts for valid reshaped copies.

Restrict the hybrid experiment to Mesa Vulkan on the validated Lunar Lake PCI device (8086:64a0), full layer/KV offload, F16 K/V, one sequence, and context up to 2048. Only the exact environment value 1 enables eligibility; 0 disables it, and FA auto-probing retains its fused graph. Limit the added Q4_0 MMVQ exception to the same device/driver. Add pinned model artifacts, bounded serving presets, CPU-logit/KV-reuse tests, interleaved depth benchmarks, and a one-hour server soak runner.

Validation completed:

• CPU regression tests and ASan/UBSan pass, including rejection through direct/planned execution, one-block views, and valid reshaped copies.
• Vulkan: 280 copy/contiguous comparisons, 29 Q4_0 matmul comparisons, and 247 Gemma attention comparisons pass.
• Local CPU CI (GG_BUILD_LOW_PERF=1): Debug 45/45 and Release 47/47 pass. LLAMA_FATAL_WARNINGS=OFF was needed for a GCC 16 warning in unchanged vocabulary code.

Full-model preset checks pass on both targets: FA-off, bounded hybrid, automatic hybrid, and strict KV reuse. The complete 26B sweep still fails one plain-FA-on control (KL 0.00991866 at 16 continuation tokens, limit 0.005); that failure remains visible. Plain FA-on is not an offered serving preset. Full-model metrics now compare token distributions with explicit bounds; raw NMSE remains reported and KV reuse retains its original strict threshold. Tests use the GGUF chat template and check answer content.

An isolated original-branch build produces byte-identical raw-prompt logits for both models in FA-off and hybrid, confirming the initial raw-prompt discrepancies predate these fixes. The validation collector now records transient firmware telemetry errors and preserves interrupted-run evidence; three infrastructure regression tests pass.

All 11 corrected lifecycle profiles pass, including tools, thinking, cache reuse, cancellation after generated content, and MTP target-token agreement. The final-budget one-hour soak passed all four phases (887 requests), including four concurrent 26B slots, with flat GPU residency and no recorded validation or telemetry errors. Serving presets use soft budgets of 1 GiB for prompt cache and 512 MiB per slot for checkpoints; measured memory use is recorded separately.

Draft: five-run depth benchmarks are running. See docs/backend/VULKAN-GEMMA4-INTEL-XE2-VALIDATION.md for measured results and limitations. No fresh speedup or deployment-readiness claim yet. The historical Xe2 artifact hashes were missing; the manifest matches documented filenames and establishes a checksummed fresh baseline.