[chris-beckman]

The available Intel Arc A310 has 4 GB VRAM, below the E4B runner's current admission plan. Add an explicit E2B mode to the generation reference benchmark so it can exercise the shared Vulkan backend on that hardware. Inputs and results record the selected size; production E4B validation remains unchanged.

An explicit portable CPU build supports diagnostic VMs that mask AVX instructions. The default retains AVX2. Build receipts bind the CMake configuration as well as the executable and sources, so changing compiler options requires a rebuilt measurement artifact.

Depends on #1734 → #1811 → #1812 → #1735. All other open prerequisite work lands first. Keep draft under the user's hold.

Verification: 20 reference harness fixtures pass. The initial physical Arc probe detected the GPU but stopped before model load on an AVX instruction; GDB confirmed that the VM exposes a generic CPU without AVX. The portable build and physical benchmarks now pass: warm 512-token prefill 353–355 tokens/s and decode about 23 tokens/s; 4096-token prefill 135–136 tokens/s and decode 18.72–18.84 tokens/s. All repeated synthetic trials produce identical tokens, with 1.70–1.72 GB reported GPU buffers and mapped PLE. This does not establish natural-language quality or concurrent-model operation. The official E2B QAT artifact is pinned to revision 675cff42a74c774d6cb76f76d8eacb49b48c9b93 with verified SHA-256. This is diagnostic coverage, not E4B performance or application qualification.

[chris-beckman]

Delivered on main through #1987 (ad462d8629) from this exact source head 4a9ecb089c23. The source branch is retained.