@elizaos/app-core 2.0.0-beta.3 → 2.0.11-beta.6

package/scripts/aosp/compile-libllama.mjs

@@ -32,23 +32,26 @@
 //   Minimum tested: zig 0.13.0. Earlier versions ship older libc++ headers
 //   that miss <bit> / <span> shims llama.cpp's CMake feature checks rely on.
 //
-// llama.cpp pin (matches plugins/plugin-aosp-local-inference/src/aosp-llama-adapter.ts):
-//   fork:   https://github.com/milady-ai/llama.cpp
-//   tag:    v0.4.0-milady          (milady/integration HEAD)
+// llama.cpp pin (matches the fork the runtime loads via
+// plugins/plugin-aosp-local-inference/src/aosp-local-inference-bootstrap.ts):
+//   fork:   https://github.com/elizaOS/llama.cpp
+//   tag:    v1.0.0-eliza           (the kernel-complete v0.4.0-eliza tree,
+//                                   re-tagged on the elizaOS org rename)
 //   commit: 08032d57e15574f2a7ca19fc3f29510c8673d590
 //
-//   v0.4.0-milady adds W4-B CUDA QJL + PolarQuant Q4 + TBQ3_TCQ kernels
-//   on top of v0.3.0-milady. The CUDA paths only matter for the
-//   linux-x64-cuda host target (the AOSP arm64 path stays CPU-only),
-//   but the pin is shared so both AOSP and host build paths land on
-//   identical kernel sources.
+//   This tree adds the W4-B CUDA QJL + PolarQuant Q4 + TBQ3_TCQ kernels
+//   on top of the earlier eliza-lineage tags. The CUDA paths only matter
+//   for the linux-x64-cuda host target (the AOSP arm64 path stays
+//   CPU-only), but the pin is shared so both AOSP and host build paths
+//   land on identical kernel sources. A rebase onto a newer upstream is a
+//   deferred effort — see docs/porting/upstream-rebase-plan.md.
 //
-//   v0.2.0-milady (subset of this pin) added DFlash speculative decoding
-//   CLI surface (--spec-type dflash, --draft-min-prob alias, n_drafted_total
-//   / n_drafted_accepted_total Prometheus counters) on top of v0.1.0-milady.
+//   v0.2.0-eliza (subset of this pin) added MTP speculative decoding
+//   CLI surface (--spec-type mtp, --spec-draft-n-min/max, n_drafted_total
+//   / n_drafted_accepted_total Prometheus counters) on top of v0.1.0-eliza.
 //
 // Why this fork (not stock ggml-org/llama.cpp b8198):
-//   The Milady fork composes four techniques onto upstream b8198:
+//   The Eliza fork composes four techniques onto upstream b8198:
 //
 //     - TBQ3_0 (slot 43) + TBQ4_0 (slot 44) — 3-bit / 4-bit TurboQuant V-cache.
 //       Cherry-picked from apothic/llama.cpp-1bit-turboquant @ b2b5273.
@@ -62,7 +65,7 @@
 //       W1-B's Polar series. Bumped from upstream slot 45 to 47 because
 //       slot 46 is now QJL.
 //     - Metal kernel sources (.metal) for TBQ3_0/TBQ4_0/TBQ3_TCQ/QJL/Polar
-//       under ggml/src/ggml-metal/milady-kernels/. Source-only landing —
+//       under ggml/src/ggml-metal/eliza-kernels/. Source-only landing —
 //       dispatcher wiring is the next agent's job.
 //
 //   The CPU implementations of all four techniques (NEON for arm64, AVX2
@@ -82,23 +85,27 @@
 //   adapter never called it anyway).
 //
 // Output (per ABI):
-//   apps/app/android/app/src/main/assets/agent/{abi}/libllama.so
-//   apps/app/android/app/src/main/assets/agent/{abi}/libggml.so
-//   apps/app/android/app/src/main/assets/agent/{abi}/libggml-cpu.so
-//   apps/app/android/app/src/main/assets/agent/{abi}/libggml-base.so
-//   apps/app/android/app/src/main/assets/agent/{abi}/libeliza-llama-shim.so
-//   apps/app/android/app/src/main/assets/agent/{abi}/llama-server          (DFlash spec-decode HTTP server)
+//   packages/app/android/app/src/main/assets/agent/{abi}/libllama.so
+//   packages/app/android/app/src/main/assets/agent/{abi}/libggml.so
+//   packages/app/android/app/src/main/assets/agent/{abi}/libggml-cpu.so
+//   packages/app/android/app/src/main/assets/agent/{abi}/libggml-base.so
+//   packages/app/android/app/src/main/assets/agent/{abi}/llama-server          (MTP spec-decode HTTP server)
+//   (with --target *-fused: libelizainference.so — the fused FFI lib)
 //
 // libllama.so has NEEDED entries on the entire libggml family (see
 // `readelf -d`); the dynamic linker resolves them from the per-ABI asset
 // dir via the LD_LIBRARY_PATH ElizaAgentService.java sets at process
-// launch. ABIs: arm64-v8a (real phones) and x86_64 (cuttlefish + emulators).
+// launch. ABIs: arm64-v8a (real phones), x86_64 (cuttlefish + emulators),
+// and riscv64 (cf_riscv64_phone + future riscv64 hardware).
 //
-// libeliza-llama-shim.so is the bun:ffi struct-by-value workaround: a
-// thin C wrapper (llama-shim/eliza_llama_shim.c next to this script)
-// that converts llama.cpp's struct-by-value entry points into
-// pointer-style equivalents bun:ffi can speak. NEEDED-links
-// libllama.so; resolved from the same asset dir at runtime.
+// libllama.so + the libggml*.so family are NOT loaded directly by any TS
+// adapter anymore — the runtime loads the fused libelizainference.so. But
+// the fused lib `target_link_libraries(elizainference PUBLIC llama)` against
+// the SHARED libllama.so, so libllama.so + libggml*.so remain runtime
+// DT_NEEDED dependencies of libelizainference.so and MUST keep building +
+// staging. The old bun:ffi struct-by-value shims (libeliza-llama-shim.so +
+// libeliza-llama-speculative-shim.so), consumed by the now-deleted
+// aosp-llama-adapter.ts, have been retired from this builder.
 //
 // Approximate build cost on a modern Linux x86_64 builder (16 cores, NVMe):
 //   - llama.cpp clone:    ~30 s, ~150 MB working tree.
@@ -153,7 +160,7 @@
 //
 // Repo-root resolution:
 //   The script defaults `--assets-dir` to
-//   `<repoRoot>/apps/app/android/app/src/main/assets/agent` and
+//   `<repoRoot>/packages/app/android/app/src/main/assets/agent` and
 //   `--cache-dir` to `~/.cache/eliza-android-agent/llama-cpp-<tag>`.
 //   `<repoRoot>` is derived from this script's location: walk up from
 //   `eliza/packages/app-core/scripts/aosp/` to the host repo root by
@@ -167,6 +174,13 @@ import os from "node:os";
 import path from "node:path";
 import process from "node:process";
 import { fileURLToPath } from "node:url";
+import { androidArm64SimdCmakeFlags } from "../build-helpers/arm64-simd.mjs";
+import {
+  fusedCmakeBuildTargets,
+  fusedExtraCmakeFlags,
+} from "../build-helpers/omnivoice-merged.mjs";
+import { verifyFusedSymbols } from "../build-helpers/verify-fused-symbols.mjs";
+import { patchVulkanKernels } from "../kernel-patches/vulkan-kernels.mjs";
 import { resolveRepoRootFromImportMeta } from "../lib/repo-root.mjs";
 import { main as compileShimMain } from "./compile-shim.mjs";
 
@@ -177,11 +191,19 @@ const here = path.dirname(fileURLToPath(import.meta.url));
 // inside the elizaOS source checkout it's the elizaOS repo root.
 const repoRoot = resolveRepoRootFromImportMeta(import.meta.url);
 
-// milady-ai/llama.cpp @ v0.4.0-milady (milady/integration HEAD).
-// Composes TBQ (apothic) + QJL (W1-A) + Q4_POLAR (W1-B) + Metal sources
-// (W1-D) + DFlash spec-decode (W2) + W3-B fused CPU kernels + W4-B CUDA
-// QJL/Polar/TBQ3_TCQ kernels onto upstream b8198. See
-// docs/porting/unified-fork-strategy.md for the full migration story.
+// elizaOS/llama.cpp @ 33c888a7b. Composes TBQ (apothic) +
+// QJL (W1-A) + Q4_POLAR (W1-B) + Metal sources (W1-D) + MTP spec-decode
+// (W2) + W3-B fused CPU kernels + W4-B CUDA QJL/Polar/TBQ3_TCQ kernels onto
+// upstream b9213. See docs/porting/unified-fork-strategy.md for the full
+// migration story.
+//
+// The fork ships in-tree as the git submodule at
+// plugins/plugin-local-inference/native/llama.cpp (next to the MTP build at
+// scripts/build-llama-cpp-mtp.mjs — same pinned commit so both build paths
+// land on identical kernels). When that
+// submodule is initialized this path defaults to it (no clone needed); pass
+// `--src-dir` to point at another checkout, or `--cache-dir` to force a
+// standalone clone of `${LLAMA_CPP_REMOTE}` at `${LLAMA_CPP_TAG}`.
 //
 // Pre-2026-05-09 the AOSP path consumed apothic/llama.cpp-1bit-turboquant
 // directly and applied vendored QJL + PolarQuant patch series via
@@ -189,11 +211,44 @@ const repoRoot = resolveRepoRootFromImportMeta(import.meta.url);
 // flow is now replaced by a single canonical fork — the patches are
 // baked in. apply-patches.mjs is kept around for one release as a
 // rollback path; see scripts/aosp/llama-cpp-patches/README.md.
-export const LLAMA_CPP_TAG = "v0.4.0-milady";
-export const LLAMA_CPP_COMMIT = "08032d57e15574f2a7ca19fc3f29510c8673d590";
-export const LLAMA_CPP_REMOTE =
-  "https://github.com/milady-ai/llama.cpp.git";
+export const LLAMA_CPP_TAG = "v1.2.0-eliza";
+export const LLAMA_CPP_COMMIT = "33c888a7be0b0b8ffb54cd3f0e05b4bed20cc52e";
+export const LLAMA_CPP_REMOTE = "https://github.com/elizaOS/llama.cpp.git";
 export const MIN_ZIG_VERSION = "0.13.0";
+// Floor for the RVV-on riscv64 build. Zig 0.13's bundled LLVM rejects the
+// GCC-style `-march=rv64gcv*` ISA string the vendored llama.cpp's
+// ggml-cpu/CMakeLists hard-codes when GGML_RVV / GGML_RV_ZFH / etc. are ON;
+// zig 0.14+ accepts it. Below this floor we hard-disable RVV + Zfh + Zvfh +
+// Zicbop + Zihintpause + Zvfbfwma + XTheadVector + Zba + SpaceMit so the
+// MARCH_STR collapses to plain `rv64gc` (which Zig 0.13's argv filter then
+// strips, since the `riscv64-linux-musl` triple already implies rv64gc/lp64d).
+// At or above this floor we leave the upstream defaults ON, the filter
+// becomes a no-op, and the vendored RVV intrinsic kernels (q4_0/q4_1/q5_0/
+// q5_1/q8_0/q8_1/q4_K/q5_K/q6_K/q8_K/iq*/tq1_0/tq2_0/mxfp4 in
+// ggml/src/ggml-cpu/arch/riscv/quants.c) light up.
+export const MIN_ZIG_RVV_VERSION = "0.14.0";
+
+// The in-repo submodule checkout of the fork.
+// `repoRoot` resolves to the repo root that contains a top-level package.json.
+const LLAMA_CPP_SUBMODULE_DIR = path.join(
+  repoRoot,
+  "plugins",
+  "plugin-local-inference",
+  "native",
+  "llama.cpp",
+);
+// True when the submodule is checked out (has a worktree). When so, the AOSP
+// cross-compile defaults its source dir to it instead of cloning.
+export function llamaCppSubmodulePresent() {
+  try {
+    return (
+      fs.existsSync(path.join(LLAMA_CPP_SUBMODULE_DIR, ".git")) &&
+      fs.existsSync(path.join(LLAMA_CPP_SUBMODULE_DIR, "CMakeLists.txt"))
+    );
+  } catch {
+    return false;
+  }
+}
 
 export const ABI_TARGETS = [
   {
@@ -206,13 +261,227 @@ export const ABI_TARGETS = [
     zigTarget: "x86_64-linux-musl",
     cmakeProcessor: "x86_64",
   },
+  {
+    androidAbi: "riscv64",
+    zigTarget: "riscv64-linux-musl",
+    cmakeProcessor: "riscv64",
+  },
 ];
 
+// `*-fused` android targets that are wired to real AOSP artifacts.
+// Membership in this set is the only way the fused (omnivoice-grafted) build
+// path activates from this script — there is no env-var shortcut and no
+// implicit upgrade from a non-fused `--abi` invocation.
+export const FUSED_ANDROID_TARGETS = Object.freeze([
+  "android-arm64-cpu-fused",
+  "android-x86_64-cpu-fused",
+  "android-riscv64-cpu-fused",
+]);
+
+// Extra cmake targets whose build failure must STILL abort the run. Only the
+// fused `elizainference` lib (libelizainference.so) is bundled into the APK;
+// every other extra target is a standalone CLI driver that ships nothing, so a
+// compile break in one of those (e.g. the fork's stale omnivoice-tts.cpp) must
+// not fail a build whose required libs are otherwise good.
+const CRITICAL_EXTRA_TARGETS = new Set(["elizainference"]);
+
+/**
+ * Parse one of the `android-<arch>-<backend>[-fused]` target strings used by
+ * the mtp build script into the pieces this script needs (the Android ABI
+ * + the fused/backend flags). Throws on unsupported triples — there is no
+ * implicit translation; the operator either asks for one of the known
+ * triples or gets a hard error.
+ *
+ * Exported for tests.
+ */
+export function parseAndroidTarget(target) {
+  if (typeof target !== "string" || target.length === 0) {
+    throw new Error(`[compile-libllama] target must be a non-empty string`);
+  }
+  const fused = target.endsWith("-fused");
+  const base = fused ? target.slice(0, -"-fused".length) : target;
+  const match = /^android-(arm64|x86_64|riscv64)-(cpu|vulkan)$/.exec(base);
+  if (!match) {
+    throw new Error(
+      `[compile-libllama] unsupported --target ${target}. ` +
+        `Supported: ${[
+          "android-arm64-cpu",
+          "android-arm64-cpu-fused",
+          "android-x86_64-cpu",
+          "android-x86_64-cpu-fused",
+          "android-riscv64-cpu",
+          "android-riscv64-cpu-fused",
+        ].join(", ")}`,
+    );
+  }
+  const [, arch, backend] = match;
+  // Android Vulkan is wired for arm64 only (the GPU device target). The
+  // GGML_VULKAN CMake flags + NDK glslc/headers + libggml-vulkan.so staging +
+  // the eliza-1 qjl/polar Vulkan kernel patches are applied in the build path
+  // when backend === "vulkan" (see resolveVulkanBuildConfig / the build fn).
+  if (backend === "vulkan" && arch !== "arm64") {
+    throw new Error(
+      `[compile-libllama] unsupported --target ${target}: Android Vulkan is ` +
+        `only wired for arm64 (the GPU device target). Use android-arm64-vulkan ` +
+        `or android-${arch}-cpu${fused ? "-fused" : ""}.`,
+    );
+  }
+  // Map the parsed arch token to the Android ABI directory name. arm64 →
+  // arm64-v8a (only Android ABI for aarch64); x86_64 and riscv64 share
+  // their name with the parsed token.
+  let androidAbi;
+  if (arch === "x86_64") androidAbi = "x86_64";
+  else if (arch === "riscv64") androidAbi = "riscv64";
+  else androidAbi = "arm64-v8a";
+  return { target, arch, backend, fused, androidAbi };
+}
+
+/**
+ * GGML_VULKAN CMake flags for the android-arm64 Vulkan backend. Points cmake at
+ * the NDK's host `glslc` (the shader compiler ggml-vulkan's codegen invokes),
+ * the Vulkan headers, and the aarch64 Vulkan loader stub. ggml-vulkan dlopen()s
+ * the device's real libvulkan.so at runtime; the NDK stub only satisfies the
+ * link. Throws (fail-closed) if any NDK Vulkan prerequisite is missing, so the
+ * build never silently falls back to CPU for a Vulkan target.
+ */
+export function resolveAndroidVulkanCmakeFlags({
+  androidApi = 31,
+  stagingDir,
+} = {}) {
+  const sdk = process.env.ANDROID_HOME || process.env.ANDROID_SDK_ROOT;
+  if (!sdk) {
+    throw new Error(
+      "[compile-libllama] ANDROID_HOME/ANDROID_SDK_ROOT is required for the " +
+        "android Vulkan build (NDK glslc + Vulkan headers + loader).",
+    );
+  }
+  const ndkRoot = path.join(sdk, "ndk");
+  const ndks = fs.existsSync(ndkRoot)
+    ? fs
+        .readdirSync(ndkRoot)
+        .filter((d) => /^\d+\./.test(d))
+        .sort(compareSemver)
+    : [];
+  if (ndks.length === 0) {
+    throw new Error(
+      `[compile-libllama] No NDK found under ${ndkRoot}; install one for the Vulkan build.`,
+    );
+  }
+  const ndk = path.join(ndkRoot, ndks[ndks.length - 1]);
+  const sysroot = path.join(
+    ndk,
+    "toolchains/llvm/prebuilt/linux-x86_64/sysroot",
+  );
+  const glslc = path.join(ndk, "shader-tools/linux-x86_64/glslc");
+  const libBase = path.join(sysroot, "usr/lib/aarch64-linux-android");
+  const apis = fs.existsSync(libBase)
+    ? fs
+        .readdirSync(libBase)
+        .filter((d) => /^\d+$/.test(d))
+        .map(Number)
+        .filter((n) => n >= androidApi)
+        .sort((a, b) => a - b)
+    : [];
+  if (apis.length === 0) {
+    throw new Error(
+      `[compile-libllama] No aarch64 libvulkan.so >= API ${androidApi} under ${libBase}.`,
+    );
+  }
+  const libVulkan = path.join(libBase, String(apis[0]), "libvulkan.so");
+
+  // ggml-vulkan.cpp includes <vulkan/vulkan.hpp> (the C++ Vulkan-Hpp bindings).
+  // The NDK sysroot ships only the C headers (vulkan.h / vulkan_core.h) + the
+  // libvulkan.so loader, NOT the C++ wrapper. Vulkan headers are pure API
+  // declarations (arch-independent); only the loader is arch-specific. So we
+  // stage a complete host Vulkan-Hpp header set (vulkan.hpp + the structs/enums/
+  // handles/funcs/raii partials it pulls in) into an ISOLATED include root and
+  // point Vulkan_INCLUDE_DIR there. Staging only the `vulkan/` subtree keeps the
+  // musl/zig cross-compile from ever seeing host glibc headers via a wide -I.
+  const headerCandidates = [
+    process.env.VULKAN_SDK && path.join(process.env.VULKAN_SDK, "include"),
+    "/usr/include",
+    "/usr/local/include",
+  ].filter(Boolean);
+  const hostVulkanDir = headerCandidates
+    .map((d) => path.join(d, "vulkan"))
+    .find((d) => fs.existsSync(path.join(d, "vulkan.hpp")));
+  if (!hostVulkanDir) {
+    throw new Error(
+      "[compile-libllama] vulkan/vulkan.hpp (C++ Vulkan-Hpp bindings) not found. " +
+        "ggml-vulkan needs it; install the Vulkan headers (e.g. `apt install " +
+        "libvulkan-dev`) or set VULKAN_SDK. Searched: " +
+        headerCandidates.join(", "),
+    );
+  }
+  const incRoot = stagingDir
+    ? path.resolve(stagingDir)
+    : path.join(os.tmpdir(), "eliza-vulkan-headers");
+  const stagedVulkan = path.join(incRoot, "vulkan");
+  fs.rmSync(stagedVulkan, { recursive: true, force: true });
+  fs.mkdirSync(incRoot, { recursive: true });
+  fs.cpSync(hostVulkanDir, stagedVulkan, { recursive: true });
+  // vulkan_core.h `#include <vk_video/...>` the video-codec extension headers,
+  // which live in a sibling `vk_video/` dir next to `vulkan/`. Stage it too so
+  // the single -I<incRoot> resolves both.
+  const hostVkVideoDir = path.join(path.dirname(hostVulkanDir), "vk_video");
+  if (fs.existsSync(hostVkVideoDir)) {
+    const stagedVkVideo = path.join(incRoot, "vk_video");
+    fs.rmSync(stagedVkVideo, { recursive: true, force: true });
+    fs.cpSync(hostVkVideoDir, stagedVkVideo, { recursive: true });
+  }
+
+  // ggml-vulkan.cpp also `#include <spirv/unified1/spirv.hpp>` (SPIR-V Headers,
+  // for shader reflection). The NDK bundles SPIRV-Headers under shaderc; prefer
+  // those (they match the glslc/shaderc toolchain version), else fall back to a
+  // host spirv-headers install. Stage the `spirv/` subtree into the same root.
+  const spirvCandidates = [
+    path.join(
+      ndk,
+      "sources/third_party/shaderc/third_party/spirv-tools/external/spirv-headers/include",
+    ),
+    process.env.VULKAN_SDK && path.join(process.env.VULKAN_SDK, "include"),
+    "/usr/include",
+    "/usr/local/include",
+  ].filter(Boolean);
+  const hostSpirvRoot = spirvCandidates.find((d) =>
+    fs.existsSync(path.join(d, "spirv/unified1/spirv.hpp")),
+  );
+  if (!hostSpirvRoot) {
+    throw new Error(
+      "[compile-libllama] spirv/unified1/spirv.hpp (SPIRV-Headers) not found. " +
+        "ggml-vulkan needs it; expected it under the NDK shaderc third_party " +
+        "tree or a host spirv-headers install. Searched: " +
+        spirvCandidates.join(", "),
+    );
+  }
+  const stagedSpirv = path.join(incRoot, "spirv");
+  fs.rmSync(stagedSpirv, { recursive: true, force: true });
+  fs.cpSync(path.join(hostSpirvRoot, "spirv"), stagedSpirv, { recursive: true });
+
+  for (const [name, p] of [
+    ["vulkan/vulkan.hpp", path.join(stagedVulkan, "vulkan.hpp")],
+    ["glslc", glslc],
+    ["libvulkan.so", libVulkan],
+  ]) {
+    if (!fs.existsSync(p)) {
+      throw new Error(
+        `[compile-libllama] android Vulkan build prerequisite ${name} missing: ${p}`,
+      );
+    }
+  }
+  return [
+    "-DGGML_VULKAN=ON",
+    `-DVulkan_INCLUDE_DIR=${incRoot}`,
+    `-DVulkan_GLSLC_EXECUTABLE=${glslc}`,
+    `-DVulkan_LIBRARY=${libVulkan}`,
+  ];
+}
+
 export function parseArgs(argv) {
   const args = {
     androidAssetsDir: path.join(
       repoRoot,
-      "apps",
+      "packages",
       "app",
       "android",
       "app",
@@ -228,9 +497,16 @@ export function parseArgs(argv) {
       `llama-cpp-${LLAMA_CPP_TAG}`,
     ),
     abis: ABI_TARGETS.map((t) => t.androidAbi),
+    // Optional explicit --target=android-<arch>-<backend>[-fused] triples
+    // (see parseAndroidTarget). When present, this list takes precedence
+    // over --abi (which is the legacy bulk-build entry point that produces
+    // CPU-only libllama.so for one or both ABIs, no fusion).
+    targets: [],
     skipIfPresent: false,
     jobs: Math.max(1, Math.min(os.cpus().length, 8)),
     srcDir: null,
+    cacheDirExplicit: false,
+    dryRun: false,
   };
 
   const readFlagValue = (flag, index) => {
@@ -248,6 +524,7 @@ export function parseArgs(argv) {
       i += 1;
     } else if (arg === "--cache-dir") {
       args.cacheDir = path.resolve(readFlagValue(arg, i));
+      args.cacheDirExplicit = true;
       i += 1;
     } else if (arg === "--src-dir") {
       args.srcDir = path.resolve(readFlagValue(arg, i));
@@ -262,6 +539,15 @@ export function parseArgs(argv) {
       }
       args.abis = [value];
       i += 1;
+    } else if (arg === "--target") {
+      const value = readFlagValue(arg, i);
+      // Validates the triple and records it. Resolved further below.
+      args.targets.push(parseAndroidTarget(value));
+      i += 1;
+    } else if (arg.startsWith("--target=")) {
+      args.targets.push(parseAndroidTarget(arg.slice("--target=".length)));
+    } else if (arg === "--dry-run") {
+      args.dryRun = true;
     } else if (arg === "--jobs" || arg === "-j") {
       const value = Number.parseInt(readFlagValue(arg, i), 10);
       if (!Number.isFinite(value) || value <= 0) {
@@ -275,12 +561,28 @@ export function parseArgs(argv) {
       console.log(
         "Usage: node eliza/packages/app-core/scripts/aosp/compile-libllama.mjs " +
           "[--assets-dir <PATH>] [--cache-dir <PATH>] [--src-dir <PATH>] " +
-          "[--abi <arm64-v8a|x86_64>] [--jobs <N>] [--skip-if-present]\n" +
-          "  --src-dir <PATH>  Use an existing llama.cpp / buun-llama-cpp checkout\n" +
-          "                    instead of cloning. The directory's HEAD is used as-is;\n" +
-          "                    the pinned LLAMA_CPP_TAG/COMMIT in this script is ignored.\n" +
-          "                    Use this to build the spiritbuun/buun-llama-cpp fork\n" +
-          "                    (TurboQuant KV-cache + DFlash kernels).",
+          "[--abi <arm64-v8a|x86_64|riscv64>] [--target <android-<arch>-<backend>[-fused]>] " +
+          "[--jobs <N>] [--skip-if-present] [--dry-run]\n" +
+          "  --target <TRIPLE>  Build a single target: android-{arm64,x86_64,riscv64}-cpu[-fused].\n" +
+          "                    riscv64 requires NDK r27+ (first stable NDK with a real\n" +
+          "                    riscv64-linux-android sysroot); older NDKs will fail the\n" +
+          "                    compiler probe before any TU compiles.\n" +
+          "                    Android Vulkan targets fail closed until GGML_VULKAN\n" +
+          "                    flags and Vulkan backend artifact staging are wired.\n" +
+          "                    -fused enables the omnivoice graft (same as mtp's\n" +
+          "                    *-fused desktop targets) — one binary serving text +\n" +
+          "                    POST /v1/audio/speech.\n" +
+          "  --dry-run         Print the cmake invocation + graft steps + expected\n" +
+          "                    output layout WITHOUT running cmake/ndk. Honored for\n" +
+          "                    every --target.\n" +
+          "  --src-dir <PATH>  Use an existing llama.cpp checkout instead of the\n" +
+          "                    in-repo submodule / a fresh clone. The directory's HEAD\n" +
+          "                    is used as-is; the pinned LLAMA_CPP_TAG/COMMIT is ignored.\n" +
+          `  Default source:   the git submodule plugins/plugin-local-inference/native/llama.cpp\n` +
+          `                    (elizaOS/llama.cpp @ ${LLAMA_CPP_TAG}) when initialized;\n` +
+          `                    otherwise a standalone clone under --cache-dir.\n` +
+          "  --cache-dir <PATH>  Force the standalone-clone path even when the submodule\n" +
+          "                    is present.",
       );
       process.exit(0);
     } else {
@@ -288,6 +590,14 @@ export function parseArgs(argv) {
     }
   }
 
+  // Default the source dir to the in-repo submodule when it is initialized and
+  // the caller did not point us elsewhere (--src-dir) or force a standalone
+  // clone (--cache-dir). Keeps both build paths (mtp + AOSP) on the exact
+  // same pinned commit.
+  if (!args.srcDir && !args.cacheDirExplicit && llamaCppSubmodulePresent()) {
+    args.srcDir = LLAMA_CPP_SUBMODULE_DIR;
+  }
+
   return args;
 }
 
@@ -354,6 +664,126 @@ export function probeZig({
   return version;
 }
 
+/**
+ * Decide the riscv64 build plan based on the detected Zig version + env knobs.
+ * Pure: takes the version string + env, returns a structured plan. No side
+ * effects.
+ *
+ * Returns one of:
+ *   { rvv: false, allVariants: false, zigVersion, reason: "zig-too-old" }
+ *     — Zig < MIN_ZIG_RVV_VERSION. Scalar parity; the riscv64ArgFilter in
+ *     ensureZigDrivers strips `-march=rv64gc*` and the resulting binary is
+ *     a plain rv64gc/lp64d build.
+ *
+ *   { rvv: true, allVariants: false, zigVersion, reason: "zig-supports-rvv" }
+ *     — Zig >= MIN_ZIG_RVV_VERSION. RVV + Zfh + Zvfh + Zicbop + Zihintpause
+ *     ON (the upstream llama.cpp defaults). Resulting libggml-cpu.so requires
+ *     RVV-capable hardware at runtime; SIGILLs on a scalar core.
+ *
+ *   { rvv: true, allVariants: true, zigVersion, reason: "all-variants-opt-in" }
+ *     — Zig >= MIN_ZIG_RVV_VERSION AND env ELIZA_GGML_CPU_ALL_VARIANTS=1.
+ *     Builds GGML_BACKEND_DL + GGML_CPU_ALL_VARIANTS so two
+ *     libggml-cpu-riscv64_{0,v}.so variants ship; runtime picks via
+ *     riscv_hwprobe (`ggml-cpu/arch/riscv/cpu-feats.cpp`). Opt-in until the
+ *     Android loader story for the DL-backend dispatch is verified end-to-end
+ *     across arm64/x86_64 — flipping that on by default would change the
+ *     artifact list for non-riscv64 ABIs too.
+ *
+ *   { rvv: false, allVariants: false, zigVersion: null, reason: "zig-not-detected" }
+ *     — Probe failed (used in dry-run mode where zig may legitimately not be
+ *     on PATH). Falls back to scalar so the dry-run plan reflects the
+ *     conservative default.
+ *
+ * Exported for tests.
+ */
+export function resolveRiscv64BuildPlan({
+  zigVersion = null,
+  probe = probeZig,
+  env = process.env,
+  isDryRun = false,
+} = {}) {
+  let version = zigVersion;
+  if (version === null) {
+    if (isDryRun) {
+      try {
+        version = probe();
+      } catch {
+        return {
+          rvv: false,
+          allVariants: false,
+          zigVersion: null,
+          reason: "zig-not-detected",
+        };
+      }
+    } else {
+      version = probe();
+    }
+  }
+  if (compareSemver(version, MIN_ZIG_RVV_VERSION) < 0) {
+    return {
+      rvv: false,
+      allVariants: false,
+      zigVersion: version,
+      reason: "zig-too-old",
+    };
+  }
+  const allVariantsOptIn = env.ELIZA_GGML_CPU_ALL_VARIANTS === "1";
+  return {
+    rvv: true,
+    allVariants: allVariantsOptIn,
+    zigVersion: version,
+    reason: allVariantsOptIn ? "all-variants-opt-in" : "zig-supports-rvv",
+  };
+}
+
+/**
+ * Map a riscv64 build plan to the cmake -D flags that select the right
+ * GGML_RVV / GGML_RV_ZFH / etc. combination. Returns an empty array for
+ * non-riscv64 ABIs.
+ *
+ * Exported for tests.
+ */
+export function riscv64CmakeFlagsForPlan({ abi, plan }) {
+  if (abi !== "riscv64") return [];
+  if (plan.rvv === false) {
+    return [
+      "-DGGML_RVV=OFF",
+      "-DGGML_RV_ZFH=OFF",
+      "-DGGML_RV_ZVFH=OFF",
+      "-DGGML_RV_ZICBOP=OFF",
+      "-DGGML_RV_ZIHINTPAUSE=OFF",
+      "-DGGML_RV_ZVFBFWMA=OFF",
+      "-DGGML_XTHEADVECTOR=OFF",
+      "-DGGML_RV_ZBA=OFF",
+      "-DGGML_CPU_RISCV64_SPACEMIT=OFF",
+    ];
+  }
+  // RVV-on. Leave the vendored llama.cpp defaults (ON) for RVV / Zfh / Zvfh /
+  // Zicbop / Zihintpause. Keep Zvfbfwma / XTheadVector / Zba / SpaceMit off
+  // unless explicitly opted in — they're hardware-specific extensions that
+  // SIGILL on generic RVV cores.
+  const flags = [
+    "-DGGML_RVV=ON",
+    "-DGGML_RV_ZFH=ON",
+    "-DGGML_RV_ZVFH=ON",
+    "-DGGML_RV_ZICBOP=ON",
+    "-DGGML_RV_ZIHINTPAUSE=ON",
+    "-DGGML_RV_ZVFBFWMA=OFF",
+    "-DGGML_XTHEADVECTOR=OFF",
+    "-DGGML_RV_ZBA=OFF",
+    "-DGGML_CPU_RISCV64_SPACEMIT=OFF",
+  ];
+  if (plan.allVariants) {
+    // GGML_CPU_ALL_VARIANTS implies GGML_BACKEND_DL and builds per-variant
+    // libggml-cpu-riscv64_{0,v}.so; the loader picks via riscv_hwprobe.
+    // GGML_NATIVE is incompatible with GGML_BACKEND_DL in this mode (see
+    // ggml/src/ggml-cpu/CMakeLists.txt:491-494), so the existing
+    // `-DGGML_NATIVE=OFF` we pass must stay OFF (it does).
+    flags.push("-DGGML_BACKEND_DL=ON", "-DGGML_CPU_ALL_VARIANTS=ON");
+  }
+  return flags;
+}
+
 function run(command, args, { cwd, env = process.env } = {}) {
   const result = spawnSync(command, args, {
     cwd,
@@ -397,6 +827,7 @@ export function ensureLlamaCppCheckout({
     log(`[compile-libllama] Reusing cached llama.cpp checkout at ${cacheDir}`);
     patchLlamaCppSourceForMusl({ srcDir: cacheDir, log });
     applyVendoredPatches({ srcDir: cacheDir, log });
+    assertSwaSpecDecodeFallback({ srcDir: cacheDir });
     return cacheDir;
   }
   if (!fs.existsSync(path.join(cacheDir, ".git"))) {
@@ -441,7 +872,7 @@ export function ensureLlamaCppCheckout({
  * and across partial-failure re-runs.
  *
  * Patches under `llama-cpp-patches/qjl/` add `GGML_TYPE_QJL1_256` (=46),
- * the QJL kernel sources vendored from `packages/native-plugins/qjl-cpu/`,
+ * the QJL kernel sources vendored from `packages/native/plugins/qjl-cpu/`,
  * the type-traits + op-dispatch wiring, and the `tests/test-qjl-cache.cpp`
  * synthetic-graph test.
  *
@@ -449,7 +880,7 @@ export function ensureLlamaCppCheckout({
  * `polarquant` series under the same directory exists but conflicts with
  * `qjl` over the GGML_TYPE_COUNT tag (PolarQuant claims id 45, QJL
  * claims 46) and is owned by a separate landing. When that series is
- * unified with QJL, append it here.
+ * merged with QJL, append it here.
  *
  * Order is:
  *   1. checkout -> 2. patchLlamaCppSourceForMusl -> 3. applyVendoredPatches.
@@ -478,6 +909,30 @@ export function applyVendoredPatches({
   spawn("node", [applierPath, "--repo", srcDir, "--series", "qjl"], {});
 }
 
+function sourceContainsSwaSpecDecodeFallback(srcDir) {
+  const serverContextPath = path.join(
+    srcDir,
+    "tools",
+    "server",
+    "server-context.cpp",
+  );
+  if (!fs.existsSync(serverContextPath)) return false;
+  const source = fs.readFileSync(serverContextPath, "utf8");
+  return (
+    source.includes("seq_rm probe failed but model declares SWA") &&
+    source.includes("llama_model_n_swa(model_tgt) > 0") &&
+    source.includes("ctx_tgt_seq_rm_type = COMMON_CONTEXT_SEQ_RM_TYPE_FULL")
+  );
+}
+
+function assertSwaSpecDecodeFallback({ srcDir }) {
+  if (sourceContainsSwaSpecDecodeFallback(srcDir)) return;
+  throw new Error(
+    `[compile-libllama] checkout ${srcDir} lacks the SWA-aware seq_rm fallback ` +
+      `required for --spec-type mtp on SWA target bodies (elizaOS/eliza#7635).`,
+  );
+}
+
 /**
  * Ensure `ggml/src/ggml.c` has the `<execinfo.h>` include gated on
  * `__GLIBC__`. musl libc does not ship `execinfo.h`, so a bare `__linux__`
@@ -598,7 +1053,12 @@ export function patchLlamaCppSourceForMusl({ srcDir, log = console.log }) {
  *
  * Exported for unit testing.
  */
-export function ensureZigDrivers({ cacheDir, abi, zigBin = "zig" }) {
+export function ensureZigDrivers({
+  cacheDir,
+  abi,
+  zigBin = "zig",
+  riscv64MarchPassthrough = false,
+}) {
   const target = ABI_TARGETS.find((t) => t.androidAbi === abi);
   if (!target) {
     throw new Error(`[compile-libllama] Unknown ABI: ${abi}`);
@@ -607,18 +1067,99 @@ export function ensureZigDrivers({ cacheDir, abi, zigBin = "zig" }) {
   fs.mkdirSync(driverDir, { recursive: true });
   const ccPath = path.join(driverDir, "zig-cc");
   const cxxPath = path.join(driverDir, "zig-cxx");
+
+  // riscv64 needs an extra arg-filtering step on Zig 0.13. The vendored
+  // llama.cpp's ggml-cpu CMakeLists hardcodes `-march=rv64gc -mabi=lp64d`
+  // (and adds extension suffixes when GGML_RVV / GGML_RV_ZFH / etc. are ON).
+  // Zig 0.13's bundled LLVM doesn't accept `-march=rv64gc` as a GCC-style
+  // ISA string — it tries to translate it to `-mcpu=` and bails with
+  // "unknown CPU: 'rv64gc'". The triple `riscv64-linux-musl` already
+  // selects the rv64gc/lp64d baseline as Zig's triple-derived CPU, so
+  // stripping these flags is byte-for-byte equivalent to the intended
+  // build when RVV is OFF.
+  //
+  // On Zig 0.14+ (`riscv64MarchPassthrough=true`) we leave every
+  // `-march=` / `-mabi=` flag alone: Zig 0.14's LLVM accepts the
+  // GCC-style ISA string with the full `_zfh_zvfh_zicbop_zihintpause`
+  // extension suffix, which is exactly what flips the RVV intrinsic
+  // codepaths on in ggml/src/ggml-cpu/arch/riscv/quants.c.
+  //
+  // Filter logic: walk the argv via the POSIX `set --` idiom (no eval —
+  // CMake escapes embedded quotes in -DGGML_VERSION=\"0.12.0\", and a
+  // naive `eval exec "..."` collapses them and the C preprocessor sees
+  // `0.12.0` as a malformed numeric literal). Each non-stripped arg is
+  // re-pushed onto $@ in place; the final `exec "$zig" cc ... "$@"`
+  // forwards the whole array with every quote and space preserved.
+  const riscv64ArgFilter =
+    abi === "riscv64" && !riscv64MarchPassthrough
+      ? '_n=$#\n' +
+        'i=0\n' +
+        'while [ $i -lt $_n ]; do\n' +
+        '  arg=$1\n' +
+        '  shift\n' +
+        '  i=$((i+1))\n' +
+        '  case "$arg" in\n' +
+        '    -march=rv64gc|-march=rv64gc_*) ;;\n' +
+        '    -mabi=lp64d|-mabi=lp64) ;;\n' +
+        '    *) set -- "$@" "$arg" ;;\n' +
+        '  esac\n' +
+        'done\n'
+      : null;
+
+  // arm64: the ggml-cpu CMakeLists emits the GCC-style ISA string
+  // `-march=armv8.2-a+dotprod+fp16+i8mm` (from GGML_CPU_ARM_ARCH). Zig 0.13's
+  // bundled LLVM rejects that for the aarch64 target — it tries to translate
+  // `armv8.2-a` to a `-mcpu=` value and dies with "unknown CPU: 'armv8.2'"
+  // (the same class of breakage the riscv64 filter handles). Zig instead
+  // speaks `-mcpu=<cpu>+<feature>` with its OWN feature names. Rewrite the
+  // GCC `-march=armv8.x-a+...` into the equivalent zig `-mcpu=generic+...`:
+  // dotprod→dotprod, i8mm→i8mm, fp16→fullfp16. This sets exactly the same
+  // __ARM_FEATURE_DOTPROD / __ARM_FEATURE_MATMUL_INT8 /
+  // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC macros (verified), so the live QJL
+  // NEON-dotprod / i8mm / fp16 kernel bodies survive preprocessing and the
+  // ggml ARM-feature configure probes pass. Any other `-march=` is passed
+  // through untouched (there shouldn't be one for arm64).
+  const arm64ArgFilter =
+    abi === "arm64-v8a"
+      ? '_n=$#\n' +
+        'i=0\n' +
+        'while [ $i -lt $_n ]; do\n' +
+        '  arg=$1\n' +
+        '  shift\n' +
+        '  i=$((i+1))\n' +
+        '  case "$arg" in\n' +
+        '    -march=armv8.*-a+*)\n' +
+        '      _feats=""\n' +
+        '      case "$arg" in *+dotprod*) _feats="${_feats}+dotprod" ;; esac\n' +
+        '      case "$arg" in *+i8mm*) _feats="${_feats}+i8mm" ;; esac\n' +
+        '      case "$arg" in *+fp16*) _feats="${_feats}+fullfp16" ;; esac\n' +
+        '      set -- "$@" "-mcpu=generic${_feats}" ;;\n' +
+        '    *) set -- "$@" "$arg" ;;\n' +
+        '  esac\n' +
+        'done\n'
+      : null;
+
+  const argFilter = riscv64ArgFilter ?? arm64ArgFilter;
+  const exec =
+    argFilter !== null
+      ? (subcmd) =>
+          argFilter +
+          `exec "${zigBin}" ${subcmd} --target=${target.zigTarget} "$@"\n`
+      : (subcmd) =>
+          `exec "${zigBin}" ${subcmd} --target=${target.zigTarget} "$@"\n`;
+
   // Quote zigBin so a path with spaces still works. The driver runs under
   // /bin/sh which is POSIX-portable across Linux, macOS, Alpine.
   const ccBody =
     "#!/bin/sh\n" +
     "# Auto-generated by eliza/packages/app-core/scripts/aosp/compile-libllama.mjs.\n" +
     "# Do not edit — regenerated on every build.\n" +
-    `exec "${zigBin}" cc --target=${target.zigTarget} "$@"\n`;
+    exec("cc");
   const cxxBody =
     "#!/bin/sh\n" +
     "# Auto-generated by eliza/packages/app-core/scripts/aosp/compile-libllama.mjs.\n" +
     "# Do not edit — regenerated on every build.\n" +
-    `exec "${zigBin}" c++ --target=${target.zigTarget} "$@"\n`;
+    exec("c++");
   fs.writeFileSync(ccPath, ccBody, "utf8");
   fs.writeFileSync(cxxPath, cxxBody, "utf8");
   fs.chmodSync(ccPath, 0o755);
@@ -653,6 +1194,12 @@ export function buildLibllamaForAbi({
   zigBin = "zig",
   log = console.log,
   spawn = run,
+  // Optional pass-through hooks used by the explicit-triple path
+  // (`mainTargets`) to layer in the fused omnivoice flags + targets without
+  // forking this helper. The non-fused bulk --abi path defaults both to
+  // empty so its behavior stays byte-for-byte identical.
+  extraCmakeFlags = [],
+  extraBuildTargets = [],
 }) {
   const target = ABI_TARGETS.find((t) => t.androidAbi === abi);
   if (!target) {
@@ -661,10 +1208,84 @@ export function buildLibllamaForAbi({
   const buildDir = path.join(srcDir, `build-${abi}`);
   fs.mkdirSync(buildDir, { recursive: true });
 
+  // riscv64: gate the RVV-on path on the detected Zig version. The vendored
+  // llama.cpp defaults GGML_RVV / GGML_RV_ZFH / GGML_RV_ZVFH / GGML_RV_ZICBOP
+  // / GGML_RV_ZIHINTPAUSE to ON, which builds the `-march=rv64gcv_zfh_zvfh_
+  // zicbop_zihintpause` ISA string. Zig 0.13's bundled LLVM doesn't accept
+  // that as a -march value (it tries to translate it to -mcpu= and fails
+  // with "unknown CPU"); Zig 0.14+ does. resolveRiscv64BuildPlan() reads the
+  // probed version and tells us which lane to take:
+  //   - rvv=false (Zig < 0.14)  -> scalar parity. Force every RVV / Zfh /
+  //     Zvfh / Zicbop / Zihintpause / Zvfbfwma / XTheadVector option OFF so
+  //     MARCH_STR collapses to plain `rv64gc`. The driver-script argv
+  //     filter then strips `-march=rv64gc -mabi=lp64d` (Zig already implies
+  //     them via the triple).
+  //   - rvv=true  (Zig >= 0.14) -> upstream defaults. The full
+  //     `-march=rv64gcv_zfh_zvfh_zicbop_zihintpause -mabi=lp64d` string
+  //     passes straight through to Zig 0.14's LLVM; quants.c's intrinsic
+  //     codepaths light up.
+  //   - allVariants (env ELIZA_GGML_CPU_ALL_VARIANTS=1) -> additionally
+  //     enable GGML_BACKEND_DL + GGML_CPU_ALL_VARIANTS so the build emits
+  //     libggml-cpu-riscv64_{0,v}.so siblings and the loader picks via
+  //     riscv_hwprobe at runtime. Opt-in until the Android DL-loader
+  //     plumbing for arm64/x86_64 is also verified.
+  const riscv64Plan =
+    abi === "riscv64"
+      ? resolveRiscv64BuildPlan({ env: process.env })
+      : { rvv: false, allVariants: false, zigVersion: null, reason: "n/a" };
+  if (abi === "riscv64") {
+    log(
+      `[compile-libllama] riscv64 plan: zig=${riscv64Plan.zigVersion ?? "unknown"} ` +
+        `rvv=${riscv64Plan.rvv ? "ON" : "OFF"} ` +
+        `all-variants=${riscv64Plan.allVariants ? "ON" : "OFF"} ` +
+        `reason=${riscv64Plan.reason}`,
+    );
+  }
+  const riscv64BuildFlags = riscv64CmakeFlagsForPlan({ abi, plan: riscv64Plan });
+
+  // x86_64: the mobile x86_64 ABI only ever runs on cuttlefish / the Android
+  // x86_64 emulator (both KVM-backed by an AVX2-class host, and our emulator
+  // recipe boots with `-cpu host`). GGML_NATIVE=OFF leaves the build at the
+  // baseline x86_64 ISA, which has two problems: (1) ggml's own AVX2 kernels
+  // stay off, and (2) — fatal — the vendored QJL kernels gate their AVX2
+  // implementations on `__AVX2__` while `qjl_dispatch.c` references
+  // `qjl_quantize_rows_avx2` (and the score/projection AVX2 entry points)
+  // unconditionally, so a baseline build links with an UNDEFINED symbol and
+  // `dlopen(libllama.so)` fails at runtime with
+  // `Error relocating libggml-cpu.so.0: qjl_quantize_rows_avx2: symbol not
+  // found`. Turning on the standard ggml AVX2/FMA/F16C/AVX feature flags
+  // defines `__AVX2__` for the ggml-cpu translation units (QJL included) so
+  // those entry points are actually compiled. Runtime CPU dispatch still picks
+  // scalar vs AVX2 per-call, but the symbols now exist.
+  const x86_64BuildFlags =
+    abi === "x86_64"
+      ? ["-DGGML_AVX=ON", "-DGGML_AVX2=ON", "-DGGML_FMA=ON", "-DGGML_F16C=ON"]
+      : [];
+
+  // arm64-v8a: GGML_NATIVE=OFF leaves the cross-build at the bare armv8-a
+  // baseline, which keeps ggml's dotprod/i8mm/fp16 NEON kernels AND the eliza
+  // QJL NEON-dotprod kernel dead. Pin the armv8.2-a+dotprod+fp16+i8mm floor and
+  // flip the QJL dispatch define so the Pixel-class Tensor G4 actually runs the
+  // accelerated paths. See build-helpers/arm64-simd.mjs for the full rationale.
+  const arm64BuildFlags = androidArm64SimdCmakeFlags(abi);
+  if (arm64BuildFlags.length > 0) {
+    log(
+      `[compile-libllama] arm64 SIMD floor: ${arm64BuildFlags.join(" ")} ` +
+        `(dotprod/i8mm/fp16 + QJL NEON-dotprod dispatch)`,
+    );
+  }
+
   // Per-ABI driver scripts that wrap `zig cc --target=<triple>` so cmake's
   // single-binary compiler probe works. See ensureZigDrivers() for why
-  // passing `--target=` via CMAKE_C_FLAGS doesn't work on its own.
-  const { ccPath, cxxPath } = ensureZigDrivers({ cacheDir, abi, zigBin });
+  // passing `--target=` via CMAKE_C_FLAGS doesn't work on its own. When
+  // RVV is on, the riscv64 driver passes `-march=rv64gc*` through to Zig
+  // 0.14+ instead of filtering it out.
+  const { ccPath, cxxPath } = ensureZigDrivers({
+    cacheDir,
+    abi,
+    zigBin,
+    riscv64MarchPassthrough: riscv64Plan.rvv,
+  });
 
   log(
     `[compile-libllama] Configuring llama.cpp for ${abi} (${target.zigTarget}) in ${buildDir}`,
@@ -680,9 +1301,10 @@ export function buildLibllamaForAbi({
       "-DBUILD_SHARED_LIBS=ON",
       "-DLLAMA_BUILD_EXAMPLES=OFF",
       "-DLLAMA_BUILD_TESTS=OFF",
-      // llama-server is required for the AOSP DFlash speculative-decode path
-      // (target + drafter share one process; aosp-llama-adapter.ts spawns this
-      // binary and routes inference over the OpenAI-compatible HTTP API). The
+      // llama-server is required for the AOSP MTP speculative-decode path
+      // (target + drafter share one process; the AOSP local-inference
+      // bootstrap spawns this binary and routes inference over the
+      // OpenAI-compatible HTTP API). The
       // server target also pulls in the JSON/HTTP common-lib pieces, but adds
       // ~1.5 MB stripped per ABI; small price relative to the spec-decode
       // throughput win.
@@ -699,6 +1321,9 @@ export function buildLibllamaForAbi({
       // device of the target ABI. The default tunes for the build host's
       // native cpu, which is wrong for a cross-build.
       "-DGGML_NATIVE=OFF",
+      ...riscv64BuildFlags,
+      ...x86_64BuildFlags,
+      ...arm64BuildFlags,
       // Don't bake in an absolute RUNPATH to the build tree. The default
       // CMAKE_BUILD_RPATH points at the per-ABI build dir, which is a
       // path-leak in shipped APKs and adds dead lookup entries at runtime.
@@ -709,6 +1334,11 @@ export function buildLibllamaForAbi({
       "-DCMAKE_SKIP_INSTALL_RPATH=TRUE",
       "-DCMAKE_BUILD_WITH_INSTALL_RPATH=TRUE",
       "-DCMAKE_INSTALL_RPATH=",
+      // `extraCmakeFlags` carries the omnivoice fused-build flags
+      // (-DELIZA_FUSE_OMNIVOICE=ON, etc.) when the explicit-triple
+      // path asked for a fused build. Empty for the non-fused bulk
+      // --abi path.
+      ...extraCmakeFlags,
     ],
     {},
   );
@@ -720,6 +1350,72 @@ export function buildLibllamaForAbi({
     {},
   );
 
+  // Build any extra cmake targets the caller asked for — for fused builds
+  // this is omnivoice-core + libelizainference + llama-omnivoice-server +
+  // the bench/completion drivers (see fusedCmakeBuildTargets()). We filter
+  // out `llama` + `llama-server` upstream (the dedicated build steps below
+  // already handle those), so the extra-target invocation only adds NEW
+  // CMake target names. The non-fused path passes an empty list.
+  //
+  // Targets are filtered against what the configured build tree actually
+  // exposes. The eliza llama.cpp fork's target set drifts from the script's
+  // pinned expectations — e.g. `llama-speculative-simple` is an upstream
+  // example the fork drops in favour of MTP spec-decode. A
+  // requested-but-absent *auxiliary* target must not abort the whole
+  // libllama build: the libllama.so + libggml*.so family is the critical
+  // output and is fully built by the `llama` target above. We warn on the
+  // gap and continue. A target that *exists* but fails to build still
+  // hard-errors via `spawn()`.
+  if (extraBuildTargets.length > 0) {
+    const helpProbe = spawnSync(
+      "cmake",
+      ["--build", buildDir, "--target", "help"],
+      { encoding: "utf8" },
+    );
+    const availableTargets = new Set(
+      (helpProbe.stdout || "")
+        .split("\n")
+        .map((line) => line.replace(/^\.\.\.\s*/, "").trim())
+        .filter(Boolean),
+    );
+    for (const extraTarget of extraBuildTargets) {
+      if (availableTargets.size > 0 && !availableTargets.has(extraTarget)) {
+        log(
+          `[compile-libllama] Skipping extra cmake target ${extraTarget} for ${abi} — ` +
+            `not defined in this llama.cpp checkout (auxiliary target; libllama.so is unaffected).`,
+        );
+        continue;
+      }
+      log(
+        `[compile-libllama] Building extra cmake target ${extraTarget} for ${abi}`,
+      );
+      try {
+        spawn(
+          "cmake",
+          ["--build", buildDir, "--target", extraTarget, "-j", String(jobs)],
+          {},
+        );
+      } catch (err) {
+        // The fused libelizainference.so (`elizainference` target) is the only
+        // extra target the APK actually bundles, and `verifyFusedSymbols`
+        // enforces it after this loop — so it stays fatal. The rest are
+        // standalone CLI drivers (omnivoice-tts / omnivoice-codec / llama-cli /
+        // llama-bench / llama-completion / llama-mtmd-cli) that ship nothing
+        // into the APK. The pinned fork's `omnivoice-tts.cpp` currently calls a
+        // removed `backend_init("LM")` overload (backend.h only exposes
+        // `backend_init_auto()`), so that driver fails to compile while
+        // libelizainference.so builds fine. Don't let a broken auxiliary CLI
+        // abort the build that produces the lib we need; warn and continue.
+        if (CRITICAL_EXTRA_TARGETS.has(extraTarget)) throw err;
+        log(
+          `[compile-libllama] WARN: auxiliary cmake target ${extraTarget} failed to build for ${abi}; ` +
+            `continuing — it bundles nothing into the APK and libllama.so/libelizainference.so are unaffected. ` +
+            `Cause: ${err instanceof Error ? err.message : String(err)}`,
+        );
+      }
+    }
+  }
+
   // llama-server target. Built in a second --target invocation so a future
   // operator can disable it via a flag without touching the libllama target.
   // The target name is `llama-server` on the apothic fork (verified against
@@ -741,53 +1437,73 @@ export function buildLibllamaForAbi({
   // libllama.so so the dynamic linker resolves the whole graph from the
   // per-ABI asset dir at runtime (LD_LIBRARY_PATH set by
   // ElizaAgentService.java).
-  const builtLlama = locateBuiltLib(buildDir, "libllama.so");
-  if (!builtLlama) {
-    throw new Error(
-      `[compile-libllama] Could not locate built libllama.so anywhere under ${buildDir}.`,
-    );
-  }
-  const builtGgmlLibs = locateBuiltGgmlLibs(buildDir);
-  if (builtGgmlLibs.length === 0) {
-    throw new Error(
-      `[compile-libllama] Could not locate any libggml*.so under ${buildDir}. ` +
-        `libllama.so has NEEDED entries for the ggml family; without co-copying ` +
-        `them the runtime dlopen will fail. Check that BUILD_SHARED_LIBS=ON took effect.`,
-    );
-  }
+  // Static-fuse (BUILD_SHARED_LIBS=OFF — the `*-fused` targets): llama/ggml/
+  // mtmd build as STATIC `.a` archives folded into a self-contained
+  // libelizainference.so (no DT_NEEDED on libllama.so/libggml*.so). So the
+  // legacy co-copy of the shared libllama.so + ggml family is both impossible
+  // (.so don't exist) and unnecessary (nothing dlopens them) — skip it. The
+  // non-fused (BUILD_SHARED_LIBS=ON) bulk `--abi` path keeps staging the shared
+  // family verbatim for its libllama.so-loading consumers.
+  const isStaticFused = extraCmakeFlags.some((f) =>
+    /BUILD_SHARED_LIBS\s*=\s*OFF/i.test(String(f)),
+  );
 
   fs.mkdirSync(abiAssetDir, { recursive: true });
-  const llamaOut = path.join(abiAssetDir, "libllama.so");
-  fs.copyFileSync(builtLlama, llamaOut);
-  const ggmlOuts = builtGgmlLibs.map((src) => {
-    const dst = path.join(abiAssetDir, path.basename(src));
-    fs.copyFileSync(src, dst);
-    return dst;
-  });
-
-  // The apothic fork builds with SONAME chains: libllama.so has
-  // SONAME=libllama.so.0 and NEEDED entries pointing at SONAME (e.g.
-  // "libggml.so.0"), not at the unversioned filename. The dynamic linker
-  // matches NEEDED against on-disk SONAME, so we must ship a copy at
-  // libfoo.so.0 (or the linker fails to resolve and dlopen returns NULL).
-  // We do NOT ship the .so.X.Y.Z versioned tail — only the SONAME alias
-  // that NEEDED references.
-  //
-  // Cost: ~5MB per ABI of duplicated content (the .so and .so.0 are
-  // identical). APK build dedupes identical files automatically; even
-  // without dedup this is well under the per-ABI .so budget.
-  const sonameAliases = [];
-  for (const out of [llamaOut, ...ggmlOuts]) {
-    const soname = readSoname(out);
-    if (soname && soname !== path.basename(out)) {
-      const aliasPath = path.join(abiAssetDir, soname);
-      fs.copyFileSync(out, aliasPath);
-      sonameAliases.push(aliasPath);
-      log(
-        `[compile-libllama] Copied ${path.basename(out)} -> ${soname} ` +
-          `(NEEDED-resolution alias for ${abi}).`,
+  let llamaOut = null;
+  let ggmlOuts = [];
+  let runtimeSiblingOuts = [];
+  let sonameAliases = [];
+  if (!isStaticFused) {
+    const builtLlama = locateBuiltLib(buildDir, "libllama.so");
+    if (!builtLlama) {
+      throw new Error(
+        `[compile-libllama] Could not locate built libllama.so anywhere under ${buildDir}.`,
+      );
+    }
+    const builtGgmlLibs = locateBuiltGgmlLibs(buildDir);
+    if (builtGgmlLibs.length === 0) {
+      throw new Error(
+        `[compile-libllama] Could not locate any libggml*.so under ${buildDir}. ` +
+          `libllama.so has NEEDED entries for the ggml family; without co-copying ` +
+          `them the runtime dlopen will fail. Check that BUILD_SHARED_LIBS=ON took effect.`,
       );
     }
+    const builtRuntimeSiblingLibs = ["libllama-common.so", "libmtmd.so"]
+      .map((name) => locateBuiltLib(buildDir, name))
+      .filter(Boolean);
+
+    llamaOut = path.join(abiAssetDir, "libllama.so");
+    fs.copyFileSync(builtLlama, llamaOut);
+    ggmlOuts = builtGgmlLibs.map((src) => {
+      const dst = path.join(abiAssetDir, path.basename(src));
+      fs.copyFileSync(src, dst);
+      return dst;
+    });
+    runtimeSiblingOuts = builtRuntimeSiblingLibs.map((src) => {
+      const dst = path.join(abiAssetDir, path.basename(src));
+      fs.copyFileSync(src, dst);
+      return dst;
+    });
+
+    // The apothic fork builds with SONAME chains: libllama.so has
+    // SONAME=libllama.so.0 and NEEDED entries pointing at SONAME (e.g.
+    // "libggml.so.0"), not at the unversioned filename. The dynamic linker
+    // matches NEEDED against on-disk SONAME, so we must ship a copy at
+    // libfoo.so.0 (or the linker fails to resolve and dlopen returns NULL).
+    // We do NOT ship the .so.X.Y.Z versioned tail — only the SONAME alias
+    // that NEEDED references.
+    for (const out of [llamaOut, ...ggmlOuts, ...runtimeSiblingOuts]) {
+      const soname = readSoname(out);
+      if (soname && soname !== path.basename(out)) {
+        const aliasPath = path.join(abiAssetDir, soname);
+        fs.copyFileSync(out, aliasPath);
+        sonameAliases.push(aliasPath);
+        log(
+          `[compile-libllama] Copied ${path.basename(out)} -> ${soname} ` +
+            `(NEEDED-resolution alias for ${abi}).`,
+        );
+      }
+    }
   }
 
   // Locate + stage the llama-server binary. cmake puts it under
@@ -810,12 +1526,66 @@ export function buildLibllamaForAbi({
   } else {
     log(
       `[compile-libllama] WARN: llama-server binary not found under ${buildDir}/bin/ or ${buildDir}/. ` +
-        `DFlash speculative decode on AOSP requires it; rebuild with -DLLAMA_BUILD_SERVER=ON.`,
+        `MTP speculative decode on AOSP requires it; rebuild with -DLLAMA_BUILD_SERVER=ON.`,
+    );
+  }
+
+  // Stage the fused-build artifacts when they are present: libelizainference.so
+  // (the SHARED target the cmake graft declares) plus the legacy CLI smoke
+  // target llama-omnivoice-server. We do NOT throw when these are missing —
+  // a non-fused build (extraBuildTargets empty) won't produce them, and the
+  // caller is responsible for invoking `verifyFusedSymbols` only on fused
+  // targets. Mirrors the mtp install-loop's conditional copy of the same
+  // pair.
+  const fusedLibSrcCandidates = [
+    path.join(buildDir, "libelizainference.so"),
+    path.join(buildDir, "src", "libelizainference.so"),
+    path.join(buildDir, "bin", "libelizainference.so"),
+  ];
+  const fusedLibSrc =
+    fusedLibSrcCandidates.find((c) => fs.existsSync(c)) ??
+    locateBuiltLib(buildDir, "libelizainference.so");
+  let fusedLibOut = null;
+  if (fusedLibSrc) {
+    fusedLibOut = path.join(abiAssetDir, "libelizainference.so");
+    fs.copyFileSync(fusedLibSrc, fusedLibOut);
+    log(
+      `[compile-libllama] Copied libelizainference.so for ${abi} (${(fs.statSync(fusedLibOut).size / (1024 * 1024)).toFixed(2)} MB).`,
+    );
+  }
+  // Under static-fuse the self-contained libelizainference.so is the ONLY
+  // shipped native lib (no libllama.so/libggml*.so), so its absence is fatal.
+  if (isStaticFused && !fusedLibOut) {
+    throw new Error(
+      `[compile-libllama] static-fuse build for ${abi} produced no libelizainference.so ` +
+        `under ${buildDir}. The fused self-contained lib is the only artifact this ` +
+        `target ships — verify the elizainference cmake target built.`,
+    );
+  }
+  const fusedServerSrcCandidates = [
+    path.join(buildDir, "bin", "llama-omnivoice-server"),
+    path.join(buildDir, "llama-omnivoice-server"),
+  ];
+  const fusedServerSrc = fusedServerSrcCandidates.find((c) => fs.existsSync(c));
+  let fusedServerOut = null;
+  if (fusedServerSrc) {
+    fusedServerOut = path.join(abiAssetDir, "llama-omnivoice-server");
+    fs.copyFileSync(fusedServerSrc, fusedServerOut);
+    fs.chmodSync(fusedServerOut, 0o755);
+    log(
+      `[compile-libllama] Copied llama-omnivoice-server for ${abi} (${(fs.statSync(fusedServerOut).size / (1024 * 1024)).toFixed(2)} MB).`,
     );
   }
 
-  const stripTargets = [...ggmlOuts, llamaOut, ...sonameAliases];
+  const stripTargets = [
+    ...ggmlOuts,
+    ...runtimeSiblingOuts,
+    llamaOut,
+    ...sonameAliases,
+  ].filter(Boolean); // llamaOut is null under static-fuse (no shared libllama.so)
   if (llamaServerOut) stripTargets.push(llamaServerOut);
+  if (fusedLibOut) stripTargets.push(fusedLibOut);
+  if (fusedServerOut) stripTargets.push(fusedServerOut);
   for (const out of stripTargets) {
     const sizeBefore = fs.statSync(out).size;
     const stripped = stripBinary({ filePath: out, zigBin, log });
@@ -833,131 +1603,18 @@ export function buildLibllamaForAbi({
       );
     }
   }
-  // Re-chmod llama-server after strip — system strip may reset perms.
+  // Re-chmod executables after strip — system strip may reset perms.
   if (llamaServerOut) fs.chmodSync(llamaServerOut, 0o755);
-  return { llama: llamaOut, ggml: ggmlOuts, llamaServer: llamaServerOut };
+  if (fusedServerOut) fs.chmodSync(fusedServerOut, 0o755);
+  return {
+    llama: llamaOut,
+    ggml: ggmlOuts,
+    llamaServer: llamaServerOut,
+    elizainference: fusedLibOut,
+    omnivoiceServer: fusedServerOut,
+  };
 }
 
-/**
- * Compile `llama-shim/eliza_llama_shim.c` (next to this script) into
- * `<abiAssetDir>/libeliza-llama-shim.so`. The shim provides
- * pointer-style wrappers around llama.cpp's struct-by-value entry
- * points that bun:ffi cannot call directly. See the file's header for
- * the full rationale.
- *
- * Linkage:
- *   - Compiled with the same per-ABI zig driver used for llama.cpp
- *     (musl-linked, matches the bun-on-Android runtime ABI).
- *   - NEEDED-links libllama.so via `-L<abiAssetDir> -lllama`. Runtime
- *     resolution comes through the per-ABI LD_LIBRARY_PATH that the
- *     privileged AOSP system app's agent service sets — same mechanism
- *     libllama.so uses to find libggml*.so.
- *   - RUNPATH stripped (`-Wl,--disable-new-dtags` + no -rpath) so we don't
- *     bake in a build-host path.
- *
- * Output: `<abiAssetDir>/libeliza-llama-shim.so`, stripped to ~10-30 KB.
- *
- * Exported for tests so we can assert the compile invocation arguments
- * without running zig end-to-end.
- */
-export function buildShimForAbi({
-  cacheDir,
-  abi,
-  abiAssetDir,
-  shimSourcePath = path.join(here, "llama-shim", "eliza_llama_shim.c"),
-  llamaIncludeDir,
-  zigBin = "zig",
-  log = console.log,
-  spawn = run,
-}) {
-  if (!fs.existsSync(shimSourcePath)) {
-    throw new Error(
-      `[compile-libllama] Shim source not found at ${shimSourcePath}. ` +
-        `Restore eliza/packages/app-core/scripts/aosp/llama-shim/eliza_llama_shim.c.`,
-    );
-  }
-  if (!fs.existsSync(llamaIncludeDir)) {
-    throw new Error(
-      `[compile-libllama] llama.h include dir missing at ${llamaIncludeDir}. ` +
-        `Did the llama.cpp checkout fail?`,
-    );
-  }
-  const llamaSo = path.join(abiAssetDir, "libllama.so");
-  if (!fs.existsSync(llamaSo)) {
-    throw new Error(
-      `[compile-libllama] Cannot link shim: ${llamaSo} is missing. ` +
-        `Run buildLibllamaForAbi() before buildShimForAbi().`,
-    );
-  }
-
-  const { ccPath } = ensureZigDrivers({ cacheDir, abi, zigBin });
-  const shimOut = path.join(abiAssetDir, "libeliza-llama-shim.so");
-
-  // llama.h transitively includes ggml.h, which lives under ggml/include/
-  // in the llama.cpp tree (separate from the llama include dir). We pass
-  // both -I flags so the compiler resolves the full header chain.
-  const ggmlIncludeDir = path.resolve(llamaIncludeDir, "..", "ggml", "include");
-  if (!fs.existsSync(path.join(ggmlIncludeDir, "ggml.h"))) {
-    throw new Error(
-      `[compile-libllama] ggml.h missing under ${ggmlIncludeDir}. ` +
-        `llama.h transitively includes it; the layout of the cached ` +
-        `llama.cpp checkout may have changed.`,
-    );
-  }
-
-  log(
-    `[compile-libllama] Compiling libeliza-llama-shim.so for ${abi} (NEEDED libllama.so)`,
-  );
-  // -fPIC + -shared: build a position-independent shared object.
-  // -O2: matches llama.cpp's release optimization level.
-  // -I<include>: pick up llama.h from the cached llama.cpp checkout, and
-  //   ggml.h from the ggml/include sibling.
-  // -L<abiAssetDir> -lllama: resolve libllama.so for the link step. The
-  //   resulting .so has NEEDED libllama.so; runtime resolution is via
-  //   LD_LIBRARY_PATH set by ElizaAgentService.java.
-  // -Wl,--disable-new-dtags + no -rpath: don't bake a RUNPATH that points
-  //   at the build-host cache dir.
-  spawn(
-    ccPath,
-    [
-      "-shared",
-      "-fPIC",
-      "-O2",
-      `-I${llamaIncludeDir}`,
-      `-I${ggmlIncludeDir}`,
-      `-L${abiAssetDir}`,
-      "-Wl,--disable-new-dtags",
-      "-o",
-      shimOut,
-      shimSourcePath,
-      "-lllama",
-    ],
-    {},
-  );
-
-  if (!fs.existsSync(shimOut)) {
-    throw new Error(
-      `[compile-libllama] Shim compile reported success but ${shimOut} is missing.`,
-    );
-  }
-  const sizeBefore = fs.statSync(shimOut).size;
-  const stripped = stripBinary({ filePath: shimOut, zigBin, log });
-  if (stripped) {
-    const sizeAfter = fs.statSync(shimOut).size;
-    if (sizeAfter === 0) {
-      throw new Error(
-        `[compile-libllama] Strip produced an empty libeliza-llama-shim.so ` +
-          `(was ${sizeBefore} bytes). This is the zig objcopy in-place ` +
-          `truncation bug — the script is supposed to strip out-of-place.`,
-      );
-    }
-    log(
-      `[compile-libllama] Stripped libeliza-llama-shim.so for ${abi} ` +
-        `(${sizeBefore} -> ${sizeAfter} bytes).`,
-    );
-  }
-  return shimOut;
-}
 
 /**
  * Find every `libggml*.so` under the build tree. b4500 shipped plain .so
@@ -1175,6 +1832,60 @@ function locateBuiltLib(buildDir, soName) {
  * Falls back to system `strip --strip-all <file>` (in-place safe on
  * GNU coreutils) if `zig objcopy` is missing or errors.
  */
+// Cache of the resolved llvm-strip path (from the Android NDK toolchain).
+// Set once on first call so we don't re-walk the NDK dir for every artifact.
+let _ndkLlvmStripPathCache = undefined;
+function locateNdkLlvmStrip() {
+  if (_ndkLlvmStripPathCache !== undefined) return _ndkLlvmStripPathCache;
+  // Honor the same env-var ladder as build-llama-cpp-mtp's resolveAndroidNdk()
+  // so operators with a custom NDK location get a consistent answer in both
+  // scripts.
+  const envRoots = [
+    process.env.ANDROID_NDK_HOME,
+    process.env.ANDROID_NDK_ROOT,
+    process.env.ANDROID_NDK,
+  ].filter((v) => typeof v === "string" && v.length > 0);
+  const sdkRoots = [
+    process.env.ANDROID_HOME,
+    process.env.ANDROID_SDK_ROOT,
+    path.join(os.homedir(), "Library", "Android", "sdk"),
+    path.join(os.homedir(), "Android", "Sdk"),
+  ].filter((v) => typeof v === "string" && v.length > 0);
+  const candidateNdks = [...envRoots];
+  for (const sdk of sdkRoots) {
+    const ndkDir = path.join(sdk, "ndk");
+    if (!fs.existsSync(ndkDir)) continue;
+    const versions = fs
+      .readdirSync(ndkDir, { withFileTypes: true })
+      .filter((e) => e.isDirectory())
+      .map((e) => e.name)
+      .sort();
+    if (versions.length > 0) {
+      candidateNdks.push(path.join(ndkDir, versions[versions.length - 1]));
+    }
+  }
+  const hosts = ["linux-x86_64", "darwin-arm64", "darwin-x86_64"];
+  for (const ndk of candidateNdks) {
+    for (const host of hosts) {
+      const cand = path.join(
+        ndk,
+        "toolchains",
+        "llvm",
+        "prebuilt",
+        host,
+        "bin",
+        "llvm-strip",
+      );
+      if (fs.existsSync(cand)) {
+        _ndkLlvmStripPathCache = cand;
+        return cand;
+      }
+    }
+  }
+  _ndkLlvmStripPathCache = null;
+  return null;
+}
+
 function stripBinary({ filePath, zigBin, log }) {
   const tmpPath = `${filePath}.stripped`;
   const zigStripResult = spawnSync(
@@ -1209,38 +1920,190 @@ function stripBinary({ filePath, zigBin, log }) {
         `error=${zigStripResult.error?.message ?? "none"}); falling back to system strip.`,
     );
   }
-  // Fallback: system strip. GNU coreutils strip is in-place safe.
+  // Fallback 1: system strip. GNU coreutils strip is in-place safe.
+  // x86_64-binutils doesn't grok riscv64 ELF or aarch64 ELF, so on a
+  // mismatched host this returns non-zero — fall through to llvm-strip.
   const systemStripResult = spawnSync("strip", ["--strip-all", filePath], {
     stdio: ["ignore", "pipe", "pipe"],
   });
   if (systemStripResult.status === 0) return true;
+  // Fallback 2: NDK's llvm-strip handles every Android ABI (arm64-v8a,
+  // x86_64, riscv64) including cross-host. This is the production path
+  // for riscv64 until Zig 0.14's objcopy lands across all build hosts.
+  const llvmStrip = locateNdkLlvmStrip();
+  if (llvmStrip) {
+    const llvmStripResult = spawnSync(
+      llvmStrip,
+      ["--strip-all", filePath],
+      { stdio: ["ignore", "pipe", "pipe"] },
+    );
+    if (llvmStripResult.status === 0) return true;
+    log(
+      `[compile-libllama] DEBUG: NDK llvm-strip (${llvmStrip}) failed ` +
+        `(status=${llvmStripResult.status}, error=${llvmStripResult.error?.message ?? "none"}).`,
+    );
+  }
   log(
     `[compile-libllama] WARN: could not strip ${filePath}; shipping with debug symbols.`,
   );
   return false;
 }
 
+/**
+ * No-op shim retained for backward compatibility with callers that still
+ * import `applyOmnivoiceGraft`. H2.c collapsed the W3-3 deprecation runway:
+ * the legacy clone-and-graft path is gone; OmniVoice is built exclusively
+ * from the merged in-fork tree at `tools/omnivoice/`.
+ *
+ * Returns a minimal info object so log lines that surface fields like
+ * `commit` or `sourceCount` keep their shape; the merged tree's source
+ * count is fixed at build time and recorded by `verifyFusedSymbols`.
+ */
+export function applyOmnivoiceGraft({ srcDir: _srcDir, log = console.log }) {
+  log(
+    "[compile-libllama] omnivoice: merged in-fork path (legacy graft removed)",
+  );
+  return { mode: "merged", source: "tools/omnivoice" };
+}
+
+/**
+ * Print the dry-run plan for one `android-<arch>-<backend>[-fused]` target:
+ * the cmake invocation, the post-cmake build target list, the graft steps
+ * (for fused targets), the expected output file layout, and the post-build
+ * verify step (for fused targets). Mirrors the structure of the mtp build
+ * script's --dry-run output so the two paths read the same.
+ *
+ * Exported for tests so the dry-run rendering can be asserted without going
+ * through the CLI entry point.
+ */
+export function describeAndroidTargetDryRun({
+  target,
+  srcDir,
+  cacheDir,
+  abiAssetDir,
+  jobs,
+  log = console.log,
+}) {
+  const parsed = parseAndroidTarget(target.target ?? target);
+  const abiTarget = ABI_TARGETS.find((t) => t.androidAbi === parsed.androidAbi);
+  if (!abiTarget) {
+    throw new Error(
+      `[compile-libllama] No ABI mapping for ${parsed.androidAbi}`,
+    );
+  }
+  const buildDir = path.join(srcDir, `build-${parsed.androidAbi}`);
+  const ccPath = path.join(cacheDir, "zig-driver", parsed.androidAbi, "zig-cc");
+  const cxxPath = path.join(
+    cacheDir,
+    "zig-driver",
+    parsed.androidAbi,
+    "zig-cxx",
+  );
+  log(`[compile-libllama] (dry-run) target=${parsed.target}`);
+  log(`  zig-target=${abiTarget.zigTarget} android-abi=${parsed.androidAbi}`);
+  log(`  src=${srcDir}`);
+  log(`  build=${buildDir}`);
+  log(`  install=${abiAssetDir}`);
+  if (parsed.fused) {
+    log(`  omnivoice: merged in-fork path (tools/omnivoice/)`);
+  }
+  const cmakeFlags = [
+    "-S",
+    srcDir,
+    "-B",
+    buildDir,
+    "-DCMAKE_BUILD_TYPE=Release",
+    "-DBUILD_SHARED_LIBS=ON",
+    "-DLLAMA_BUILD_EXAMPLES=OFF",
+    "-DLLAMA_BUILD_TESTS=OFF",
+    "-DLLAMA_BUILD_SERVER=ON",
+    "-DLLAMA_CURL=OFF",
+    `-DCMAKE_C_COMPILER=${ccPath}`,
+    `-DCMAKE_CXX_COMPILER=${cxxPath}`,
+    "-DCMAKE_C_COMPILER_LAUNCHER=",
+    "-DCMAKE_CXX_COMPILER_LAUNCHER=",
+    "-DCMAKE_SYSTEM_NAME=Linux",
+    `-DCMAKE_SYSTEM_PROCESSOR=${abiTarget.cmakeProcessor}`,
+    "-DGGML_NATIVE=OFF",
+  ];
+  // riscv64 build flags must show up in dry-run output too — the real
+  // buildLibllamaForAbi() resolves these from the detected Zig version and
+  // operators reading the dry-run plan should see the byte-exact list that
+  // will be passed. resolveRiscv64BuildPlan() falls back to scalar when zig
+  // is not installed (dry-run is allowed on toolchain-less boxes), so the
+  // plan reported here mirrors what a build invocation would actually emit
+  // on the same host.
+  if (parsed.androidAbi === "riscv64") {
+    const plan = resolveRiscv64BuildPlan({ env: process.env, isDryRun: true });
+    log(
+      `  riscv64 plan: zig=${plan.zigVersion ?? "unknown"} ` +
+        `rvv=${plan.rvv ? "ON" : "OFF"} ` +
+        `all-variants=${plan.allVariants ? "ON" : "OFF"} ` +
+        `reason=${plan.reason}`,
+    );
+    cmakeFlags.push(
+      ...riscv64CmakeFlagsForPlan({ abi: parsed.androidAbi, plan }),
+    );
+  }
+  // arm64-v8a SIMD floor (dotprod/i8mm/fp16 + QJL NEON-dotprod dispatch) — the
+  // real buildLibllamaForAbi() emits these too; surface them in the dry-run.
+  cmakeFlags.push(...androidArm64SimdCmakeFlags(parsed.androidAbi));
+  cmakeFlags.push(
+    "-DCMAKE_SKIP_BUILD_RPATH=TRUE",
+    "-DCMAKE_SKIP_INSTALL_RPATH=TRUE",
+    "-DCMAKE_BUILD_WITH_INSTALL_RPATH=TRUE",
+    "-DCMAKE_INSTALL_RPATH=",
+  );
+  if (parsed.fused) {
+    cmakeFlags.push(...fusedExtraCmakeFlags());
+  }
+  log(`  cmake ${cmakeFlags.join(" ")}`);
+  const buildTargets = parsed.fused
+    ? fusedCmakeBuildTargets()
+    : ["llama", "llama-server"];
+  log(
+    `  cmake --build ${buildDir} --target ${buildTargets.join(" ")} -j ${jobs}`,
+  );
+  log(`  expected output layout under ${abiAssetDir}:`);
+  log(`    libllama.so libggml*.so llama-server`);
+  if (parsed.fused) {
+    log(
+      `    libelizainference.so omnivoice-tts omnivoice-codec (merged-tree artifacts)`,
+    );
+    log(
+      `  verifyFusedSymbols outDir=${abiAssetDir} target=${parsed.target} (post-build)`,
+    );
+  }
+}
+
 export async function main(argv = process.argv.slice(2)) {
   const args = parseArgs(argv);
 
-  // Probe toolchain first so we fail loudly before doing any work.
-  const zigVersion = probeZig();
-  console.log(`[compile-libllama] Found zig ${zigVersion}`);
+  // If --target was passed, the caller is asking for the mtp-style
+  // explicit-triple build path. --abi still drives the legacy bulk-build
+  // (cpu only, no fusion) entry point so existing callers keep working.
+  if (args.targets.length > 0) {
+    return mainTargets(args);
+  }
+
+  // Probe toolchain first so we fail loudly before doing any work. Skip in
+  // dry-run mode — operators on a box without zig still want to inspect what
+  // the build WOULD do.
+  if (!args.dryRun) {
+    const zigVersion = probeZig();
+    console.log(`[compile-libllama] Found zig ${zigVersion}`);
+  } else {
+    console.log(`[compile-libllama] (dry-run) skipping zig toolchain probe`);
+  }
 
   let allPresent = true;
   for (const abi of args.abis) {
     const llama = path.join(args.androidAssetsDir, abi, "libllama.so");
     const ggml = path.join(args.androidAssetsDir, abi, "libggml.so");
-    const shim = path.join(
-      args.androidAssetsDir,
-      abi,
-      "libeliza-llama-shim.so",
-    );
     const llamaServer = path.join(args.androidAssetsDir, abi, "llama-server");
     if (
       !fs.existsSync(llama) ||
       !fs.existsSync(ggml) ||
-      !fs.existsSync(shim) ||
       !fs.existsSync(llamaServer)
     ) {
       allPresent = false;
@@ -1253,6 +2116,31 @@ export async function main(argv = process.argv.slice(2)) {
     );
     return;
   }
+  if (args.dryRun) {
+    console.log(
+      "[compile-libllama] (dry-run) bulk --abi mode requested; emit dry-run for each ABI as a non-fused android-<arch>-cpu target",
+    );
+    const srcDirForDry =
+      args.srcDir ??
+      (llamaCppSubmodulePresent() ? LLAMA_CPP_SUBMODULE_DIR : args.cacheDir);
+    for (const abi of args.abis) {
+      // arm64-v8a is the only ANDROID_ABI that doesn't share its name with
+      // the `android-<arch>-cpu` triple's <arch> token; x86_64 and riscv64
+      // map 1:1 to the same string in both spellings. Keep this in sync
+      // with parseAndroidTarget()'s arch→ABI mapping.
+      const arch = abi === "arm64-v8a" ? "arm64" : abi;
+      const target = `android-${arch}-cpu`;
+      const abiAssetDir = path.join(args.androidAssetsDir, abi);
+      describeAndroidTargetDryRun({
+        target,
+        srcDir: srcDirForDry,
+        cacheDir: args.cacheDir,
+        abiAssetDir,
+        jobs: args.jobs,
+      });
+    }
+    return;
+  }
 
   let srcDir;
   let srcDescription;
@@ -1260,19 +2148,42 @@ export async function main(argv = process.argv.slice(2)) {
     if (!fs.existsSync(path.join(args.srcDir, "CMakeLists.txt"))) {
       throw new Error(
         `[compile-libllama] --src-dir ${args.srcDir} does not contain a CMakeLists.txt; ` +
-          `expected a llama.cpp / buun-llama-cpp checkout.`,
+          `expected a llama.cpp checkout.`,
       );
     }
     srcDir = args.srcDir;
+    const isSubmodule =
+      path.resolve(srcDir) === path.resolve(LLAMA_CPP_SUBMODULE_DIR);
     let headRef = "(unknown)";
     try {
-      headRef = fs.readFileSync(path.join(srcDir, ".git", "HEAD"), "utf8").trim();
+      // A submodule's `.git` is a file (`gitdir: ...`), not a dir, so resolve
+      // HEAD via `git rev-parse` rather than reading `.git/HEAD` directly.
+      const out = spawnSync("git", ["-C", srcDir, "rev-parse", "HEAD"], {
+        encoding: "utf8",
+      });
+      if (out.status === 0) headRef = out.stdout.trim();
     } catch {}
-    console.log(
-      `[compile-libllama] Using --src-dir ${srcDir} (HEAD: ${headRef}); ` +
-        `pinned tag ${LLAMA_CPP_TAG} ignored.`,
-    );
-    srcDescription = `external src-dir ${srcDir}`;
+    if (isSubmodule) {
+      // The in-repo submodule is pinned by the eliza repo's gitlink. Discard
+      // the source patches a prior build left behind (tracked + untracked)
+      // before re-applying them, so a fresh artifact starts from the pristine
+      // submodule tree. Never detach/fetch — `bun install` keeps it pinned.
+      console.log(
+        `[compile-libllama] Using the in-repo llama.cpp submodule ${srcDir} ` +
+          `(HEAD: ${headRef}); resetting prior source patches.`,
+      );
+      run("git", ["-C", srcDir, "checkout", "--", "."], {});
+      run("git", ["-C", srcDir, "clean", "-fdx"], {});
+      assertSwaSpecDecodeFallback({ srcDir });
+      srcDescription = `submodule plugins/plugin-local-inference/native/llama.cpp @ ${headRef.slice(0, 12)}`;
+    } else {
+      console.log(
+        `[compile-libllama] Using --src-dir ${srcDir} (HEAD: ${headRef}); ` +
+          `pinned tag ${LLAMA_CPP_TAG} ignored.`,
+      );
+      assertSwaSpecDecodeFallback({ srcDir });
+      srcDescription = `external src-dir ${srcDir}`;
+    }
   } else {
     srcDir = ensureLlamaCppCheckout({
       cacheDir: args.cacheDir,
@@ -1284,6 +2195,12 @@ export async function main(argv = process.argv.slice(2)) {
 
   for (const abi of args.abis) {
     const abiAssetDir = path.join(args.androidAssetsDir, abi);
+    // Builds + stages libllama.so + the libggml*.so family + llama-server.
+    // libllama.so + libggml*.so are runtime DT_NEEDED dependencies of the
+    // fused libelizainference.so (the fork's `elizainference` target does
+    // `target_link_libraries(elizainference PUBLIC llama)` with
+    // BUILD_SHARED_LIBS=ON), so they stay required even though no TS adapter
+    // dlopens libllama.so directly anymore.
     buildLibllamaForAbi({
       srcDir,
       cacheDir: args.cacheDir,
@@ -1293,17 +2210,6 @@ export async function main(argv = process.argv.slice(2)) {
       log: console.log,
       spawn: run,
     });
-    // Compile the bun:ffi struct-by-value shim against the freshly built
-    // libllama.so. Has to come AFTER the llama build because it links
-    // against -lllama from <abiAssetDir>.
-    buildShimForAbi({
-      cacheDir: args.cacheDir,
-      abi,
-      abiAssetDir,
-      llamaIncludeDir: path.join(srcDir, "include"),
-      log: console.log,
-      spawn: run,
-    });
   }
 
   // Cross-compile the SIGSYS-handler shim + loader-wrap for x86_64. ARM64
@@ -1319,11 +2225,201 @@ export async function main(argv = process.argv.slice(2)) {
   await compileShimMain(["--skip-if-present"]);
 
   console.log(
-    `[compile-libllama] Built libllama.so + libeliza-llama-shim.so + llama-server for ` +
+    `[compile-libllama] Built libllama.so + libggml*.so + llama-server for ` +
       `${args.abis.join(", ")} (${srcDescription}).`,
   );
 }
 
+/**
+ * Explicit-triple entry point: runs the build for one or more
+ * `android-<arch>-<backend>[-fused]` targets. Mirrors the mtp build
+ * script's `--target` semantics one-for-one so an operator running the
+ * desktop fused build and the mobile fused build invokes the two scripts
+ * with the same target string.
+ *
+ * Build flow per target:
+ *   1. Resolve the llama.cpp source tree (--src-dir / in-repo submodule /
+ *      standalone clone — same logic as the bulk --abi path).
+ *   2. For `*-fused`: the merged in-fork tree at `tools/omnivoice/`
+ *      already declares the omnivoice + elizainference targets; just add
+ *      the CMake flags via `fusedExtraCmakeFlags()`.
+ *   3. Run `buildLibllamaForAbi()` (which also configures + links the
+ *      llama-server target — required for fused so omnivoice_lib links
+ *      into the same binary).
+ *   4. For `*-fused`: run `verifyFusedSymbols()` against the install dir,
+ *      asserting libelizainference.so carries `llama_*` + `ov_*` +
+ *      `eliza_inference_*` exports.
+ *
+ * Dry-run prints what each step WOULD do without touching the filesystem
+ * or running cmake / the NDK.
+ */
+export async function mainTargets(args) {
+  // Resolve the source dir up front so dry-run can report a real path.
+  let srcDir;
+  let srcDescription;
+  if (args.srcDir) {
+    if (
+      !args.dryRun &&
+      !fs.existsSync(path.join(args.srcDir, "CMakeLists.txt"))
+    ) {
+      throw new Error(
+        `[compile-libllama] --src-dir ${args.srcDir} does not contain a CMakeLists.txt; ` +
+          `expected a llama.cpp checkout.`,
+      );
+    }
+    srcDir = args.srcDir;
+    const isSubmodule =
+      path.resolve(srcDir) === path.resolve(LLAMA_CPP_SUBMODULE_DIR);
+    srcDescription = isSubmodule
+      ? `submodule plugins/plugin-local-inference/native/llama.cpp`
+      : `external src-dir ${srcDir}`;
+    if (!args.dryRun) assertSwaSpecDecodeFallback({ srcDir });
+  } else if (args.dryRun) {
+    // In a dry run with no --src-dir and no submodule, just describe the
+    // intended cache path; we never clone in dry-run.
+    srcDir = args.cacheDir;
+    srcDescription = `cache ${args.cacheDir} (would clone ${LLAMA_CPP_TAG})`;
+  } else {
+    srcDir = ensureLlamaCppCheckout({
+      cacheDir: args.cacheDir,
+      log: console.log,
+      spawn: run,
+    });
+    srcDescription = `llama.cpp ${LLAMA_CPP_TAG} / ${LLAMA_CPP_COMMIT.slice(0, 12)}`;
+  }
+
+  // omnivoice.cpp clone lives at <cacheRoot>/omnivoice.cpp; we use the parent
+  // of the llama.cpp cache dir so both clones live under one cache root, the
+  // same shape the mtp build path uses (cacheRoot=path.dirname(args.cacheDir)).
+  const omnivoiceCacheRoot = path.dirname(args.cacheDir);
+
+  if (!args.dryRun) {
+    const zigVersion = probeZig();
+    console.log(`[compile-libllama] Found zig ${zigVersion}`);
+  } else {
+    console.log(`[compile-libllama] (dry-run) skipping zig toolchain probe`);
+  }
+
+  for (const parsed of args.targets) {
+    const abiAssetDir = path.join(args.androidAssetsDir, parsed.androidAbi);
+    if (args.dryRun) {
+      describeAndroidTargetDryRun({
+        target: parsed.target,
+        srcDir,
+        cacheDir: args.cacheDir,
+        abiAssetDir,
+        jobs: args.jobs,
+      });
+      if (parsed.fused) {
+        console.log(
+          `  fused-graft cacheRoot=${omnivoiceCacheRoot} (omnivoice.cpp clone)`,
+        );
+      }
+      continue;
+    }
+
+    // Pre-cmake: run the omnivoice graft for fused targets. Same call
+    // sequence as the mtp linux-x64-cpu-fused path; the graft is
+    // toolchain-agnostic (CMake snippet + source layout).
+    let omnivoiceInfo = null;
+    if (parsed.fused) {
+      omnivoiceInfo = applyOmnivoiceGraft({
+        srcDir,
+        omnivoiceCacheRoot,
+        log: console.log,
+      });
+    }
+
+    // Vulkan target: graft the eliza-1 qjl/polar Vulkan compute shaders +
+    // ggml-vulkan dispatch patches into the source, and assemble the
+    // GGML_VULKAN CMake flags (NDK glslc + headers + aarch64 loader). The
+    // libggml-vulkan.so the build emits is glob-staged alongside the rest of
+    // the libggml family by buildLibllamaForAbi.
+    let vulkanCmakeFlags = [];
+    if (parsed.backend === "vulkan") {
+      console.log(
+        `[compile-libllama] Patching eliza-1 Vulkan kernels into ${srcDir} for ${parsed.target}`,
+      );
+      patchVulkanKernels(srcDir, { target: parsed.target });
+      vulkanCmakeFlags = resolveAndroidVulkanCmakeFlags({
+        stagingDir: path.join(args.cacheDir, "vulkan-headers"),
+      });
+    }
+
+    // The existing per-ABI build helper handles the cmake configure +
+    // build + per-ABI install for libllama + ggml + llama-server. We
+    // reuse it as-is; the fused cmake flags + extra targets are applied
+    // below via a thin override hook so the non-fused path stays
+    // byte-for-byte identical.
+    buildLibllamaForAbi({
+      srcDir,
+      cacheDir: args.cacheDir,
+      abi: parsed.androidAbi,
+      abiAssetDir,
+      jobs: args.jobs,
+      log: console.log,
+      spawn: run,
+      // The fused path needs `-DELIZA_FUSE_OMNIVOICE=ON` on the configure
+      // line and the omnivoice-core + libelizainference + fused
+      // llama-server targets on the build line. Pass-through hooks let
+      // the caller layer those in without forking the helper. The Vulkan
+      // target adds GGML_VULKAN=ON + the NDK toolchain paths and asks the
+      // build to also produce ggml-vulkan (libggml-vulkan.so).
+      extraCmakeFlags: [
+        ...(parsed.fused ? fusedExtraCmakeFlags() : []),
+        ...vulkanCmakeFlags,
+      ],
+      extraBuildTargets: [
+        ...(parsed.fused
+          ? fusedCmakeBuildTargets().filter(
+              (t) => t !== "llama" && t !== "llama-server",
+            )
+          : []),
+        ...(parsed.backend === "vulkan" ? ["ggml-vulkan"] : []),
+      ],
+    });
+
+    // Post-build: for fused targets prove libelizainference.so exports both
+    // `llama_*` and `ov_*` (and the eliza_inference ABI surface). Hard error
+    // on a half-fused artifact — same contract as the mtp build path.
+    if (parsed.fused) {
+      const verification = verifyFusedSymbols({
+        outDir: abiAssetDir,
+        target: parsed.target,
+      });
+      console.log(
+        `[compile-libllama] omnivoice symbol-verify: ` +
+          `library=${verification.library} ` +
+          `llama=${verification.llamaSymbolCount} ` +
+          `omnivoice=${verification.omnivoiceSymbolCount} ` +
+          `abi=${verification.abiSymbolCount}`,
+      );
+      if (omnivoiceInfo) {
+        console.log(
+          `[compile-libllama] omnivoice mode=${omnivoiceInfo.mode ?? "merged"} source=${omnivoiceInfo.source ?? "tools/omnivoice"}`,
+        );
+      }
+    }
+  }
+
+  if (args.dryRun) {
+    console.log(
+      `[compile-libllama] (dry-run) plan complete: ${args.targets.length} target(s) (${srcDescription}).`,
+    );
+    return;
+  }
+
+  // SIGSYS-handler shim only needed when an x86_64 ABI was built (matches
+  // the bulk --abi path's behavior — see the comment in main()).
+  if (args.targets.some((t) => t.androidAbi === "x86_64")) {
+    await compileShimMain(["--skip-if-present"]);
+  }
+
+  console.log(
+    `[compile-libllama] Built ${args.targets.map((t) => t.target).join(", ")} (${srcDescription}).`,
+  );
+}
+
 const isMain =
   process.argv[1] &&
   path.resolve(process.argv[1]) === fileURLToPath(import.meta.url);