perf(cpu): close Raspberry Pi 5 A76 BF16 GEMM gap against llama.cpp

Row

KERNEL-GEMM-CPU-ELEM-A76

Baseline

The landed Raspberry Pi 5 / Cortex-A76 lane (#65, #79) is correct but does not meet the same-file llama.cpp floor for Qwen3.5-2B Q8_K_XL on four A76 cores:

• prefill: 12.81 vs 27.77 tok/s (0.461x)
• decode: 2.55 vs 3.91 tok/s (0.653x)
• output-equivalent E2E: 2.46 vs 3.77 tok/s (0.653x)
• peak RSS: 2.841 vs 3.747 GiB (0.758x, already better)

The last clean profile attributes 57.76% of the model run to the BF16 Bt16Neon elementwise GEMM. This is the largest measured gap and the next lever named by the landed RPi5 campaign spec.

Scope

1. Reproduce the current main baseline and profile both vllm.cpp and llama.cpp on the identical workload.
2. Add a focused microbenchmark and PMU/disassembly evidence for the reached BF16/F16/F32 elementwise GEMM shapes.
3. Optimize the C++/NEON implementation first, recursively gating the kernel, enclosing model phase, and full end-to-end workload.
4. Do not write or select assembly until the C++ path is beyond llama.cpp on prefill, decode, and E2E without regressing correctness or the existing RSS win.
5. Once that C++ floor is closed, use compiler disassembly plus Pi PMU evidence to decide whether an AAPCS64 implementation has measurable remaining headroom.

Constraints

• Build and test AArch64 artifacts locally under QEMU; do not compile on the Pi.
• Use rich@rpi5fan.lan only for execution and physical-PMU measurements.
• Same GGUF bytes, prompt/token counts, four-core affinity, sampling, and llama.cpp pin as the existing binding evidence.
• Preserve token/output correctness and measure every recursive enclosing scope after each accepted optimization.
• No dependency or package installation and no large asset download.

Done when

• C++/NEON vllm.cpp is faster than llama.cpp on the binding prefill, decode, and E2E axes, with RSS no worse than the landed baseline and repeated idle same-binary evidence.
• Any later assembly has mutation-tested correctness, ABI/disassembly proof, PMU evidence, and a recursive model win over that already-parity C++ baseline.

Begin with the KERNEL-GEMM-CPU-ELEM-A76 lane and the Bt16Neon elementwise GEMM implementation. Reproduce the identical Qwen3.5-2B Q8_K_XL workload under QEMU, then profile vllm.cpp and llama.cpp and collect microbenchmark, disassembly, and PMU evidence. Done means the C++/NEON path beats llama.cpp on prefill, decode, and E2E without correctness or RSS regressions.