ggml-cpu: add opt-in Q2_0 VNNI64 four-row decode

docs/tbkern/phase9-four-row-vnni64.md

@@ -0,0 +1,13 @@
+# Phase 9: four-row VNNI64 decode tile
+
+Phase 9 adds a separately opt-in single-token Q2_0 decode route:
+
+```text
+GGML_TBKERN_Q2_0_VNNI64_4R=1
+```
+
+The VNNI64 route previously rebuilt the identical 64-byte Q8 activation vector for each output row. The four-row tile constructs that vector once for each 64-value subgroup, then applies it to four independent packed-weight rows. Each output row retains the original `b=0,1` and group accumulation order, per-32 Q8 scales, and precomputed Q8-sum correction. A thread's final one to three rows use the unchanged Phase 8 route.
+
+This is intentionally opt-in and does not change native Prism, Phase 7 VNNI, or Phase 8 VNNI64 selection. It requires AVX-512F and AVX-512VNNI. It changes no GGUF storage, repack allocation, or generic-ggml fallback behavior.
+
+Validation gate: exact native/Phase 8/Phase 9 logits and token parity on the real 27B Q2_0 model, then matched CPU decode and PPL measurements. Retain it only if matched decode improvement exceeds run-to-run noise.

docs/tbkern/phase9-prism-evidence.md

@@ -0,0 +1,63 @@
+# Phase 9: Prism 27B minimal-PR evidence
+
+This document records the evidence for the opt-in four-row CPU decode path.
+It is a decode measurement, not a claim about prompt processing or GPU
+throughput.
+
+## Scope
+
+The minimal branch is based on Prism's prism branch at 7529fdaaf. It keeps only
+the CPU backend integration, the Q2_0 cache packer, the parity comparator, and
+this evidence. The native Prism Q2_0 path remains the default unless the
+selector is set:
+
+GGML_TBKERN_Q2_0_VNNI64_4R=1
+
+## Configuration
+
+- Model: Ternary-Bonsai-27B-Q2_0.gguf, 7,165,121,600 bytes,
+  SHA-256 868c11714cf8fe47f5ec9eeb2be0ab1a337112886f92ee0ede6b855c4fa31757.
+- Host: Google Cloud C3 highmem-8, Intel Xeon Platinum 8481C, 4 physical
+  cores plus SMT siblings, one NUMA node, AVX-512F/VNNI/VBMI.
+- Decode command: llama-bench -m MODEL -p 0 -n 16 -t 8 -ngl 0 -b 32 -ub 32.
+- Baseline: unmodified Prism prism branch, same model, host, command, and
+  eight ggml threads, with no TBKERN selector.
+
+## Correctness
+
+llama-debug with prompt Hello, deterministic generation, and the same runtime
+settings produced bit-identical files relative to the native path:
+
+- logits SHA-256:
+  65d581397ae42288fe1115e7fa434589700cacd98edc67a750c0fba1765062f2
+- generated-token SHA-256:
+  7a7748eacf971049271242b9d921628019d6c44698574e9301da9b8c88026381
+- NMSE: 0
+- top-1 mismatches: 0
+- one-chunk PPL: 1.0645 +/- 0.03006 for both routes
+
+## Matched decode triplicate
+
+The recorded three-run matched set used the exact command above:
+
+| Route | Run 1 | Run 2 | Run 3 |
+| --- | ---: | ---: | ---: |
+| Native Prism | 1.09 tok/s | 1.09 tok/s | 1.09 tok/s |
+| Phase 9 four-row VNNI64 | 1.52 tok/s | 1.53 tok/s | 1.53 tok/s |
+
+The Phase 9 route is approximately 40% faster for target decode on this host.
+These are historical C3 measurements from the equivalent Phase 9 implementation
+before this cleanup. This cleanup removes unused files and narrows the packing
+header; it does not intentionally change the arithmetic or dispatch.
+
+## Reproduction status
+
+The current Windows workspace has no CMake, GCC, Clang, or llama-bench, so the
+cleaned branch could not be rebuilt and rerun here. Before marking the PR ready
+for review, rebuild this exact commit on the Linux C3 host and repeat native and
+Phase 9 llama-bench three times, then rerun the deterministic parity check.
+
+The route remains opt-in because the recorded perplexity workload was slower
+even though the PPL value matched. Attention, norms, RoPE, softmax, GDN
+recurrence/convolution, KV cache, sampling, scheduling, and GPU execution
+remain in Prism's native runtime.

ggml/src/ggml-cpu/CMakeLists.txt

@@ -1,3 +1,7 @@
+# Keep this at directory scope: CMAKE_CURRENT_FUNCTION_LIST_DIR was added
+# after the project's CMake 3.14 minimum and is not available there.
+set(TBKERN_ROOT "${CMAKE_SOURCE_DIR}/third_party/tbkern")
+
 function(ggml_add_cpu_backend_features cpu_name arch)
     # The feature detection code is compiled as a separate target so that
     # it can be built without the architecture flags
@@ -56,7 +60,14 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
 
     target_compile_features(${GGML_CPU_NAME} PRIVATE c_std_11 cxx_std_17)
     target_include_directories(${GGML_CPU_NAME} PRIVATE . ggml-cpu)
-
+    # Resolve from the project root inside this function to avoid caller-scope shadowing.
+    set(TBKERN_ROOT "${CMAKE_SOURCE_DIR}/third_party/tbkern")
+    # Q2_0 cache packing is opt-in. Scalar and AVX2 routes are selected by env;
+    # Prism's native Q2_0 vec-dot remains the fallback.
+    target_sources(${GGML_CPU_NAME} PRIVATE
+        "${TBKERN_ROOT}/src/q2_pack.c"
+    )
+    target_include_directories(${GGML_CPU_NAME} PRIVATE "${TBKERN_ROOT}/include")
     if (APPLE AND GGML_ACCELERATE)
         find_library(ACCELERATE_FRAMEWORK Accelerate)
         if (ACCELERATE_FRAMEWORK)