Ship the aarch64-verified NEON kernels to mobile: Apple targets + Android JNI for skainet-backend-native-cpu (measured 21 → 1.0 → 0.11 tok/s cliff) #920

The numbers

Same model (SmolLM2-135M-Instruct Q8_0), same QuantPolicy.NATIVE_OPTIMIZED, same
commonMain inference code, SKaiNET 0.38.0:

Output is correct everywhere — this is purely a kernel-availability gap, and it is
what keeps our shipped app's AI feature switched off on phones (we gate on
3 tok/s at startup; both mobile platforms fail the gate). Roughly 200x between JVM
and Kotlin/Native on comparable CPUs.

Why it happens

skainet-backend-native-cpu publishes only -jvm, -linuxx64, -linuxarm64.
There is no Apple or Android artifact, Android has no java.lang.foreign / Vector
API, and every non-JVM PlatformCpuOpsFactory actual registers only
ScalarKernelProvider (priority 0). So both mobile platforms run scalar Kotlin
matmul while the JVM gets SIMD.

Why this is closer than it looks

Most of the hard part already exists in-tree:

• The C kernels (native/src/*.c) are portable C11 with NEON bodies gated on
  __ARM_NEON / __ARM_FEATURE_DOTPROD — fp32, q8_0, q4k, q5k, q6k — and carry an
  "AARCH64-VERIFIED" banner: parity-checked under qemu-aarch64 and on a physical
  Cortex-A55, -march=armv8.2-a+fp16+dotprod, deliberately no +i8mm.
• CMakeLists.txt keys the -march flags off CMAKE_SYSTEM_PROCESSOR, which NDK
  and Apple toolchains both set correctly — the flags come for free.
• The Kotlin/Native side (NativeKnKernelProvider + wrappers + the nativeTest
  parity tests) lives in target-agnostic nativeMain/nativeTest and should
  compile for Apple/Android-native targets unchanged.
• The publish workflow already runs on a macOS runner.

What actually needs doing

We'd suggest two tracks, iOS first (it reuses the verified K/N cinterop path
as-is), Android second (different mechanism needed — see below).

Track A — Apple (iosArm64, iosSimulatorArm64, ideally macosArm64):

1. Fix the cinterop packaging gap first — this one matters beyond mobile.
   skainet_kernels.def declares no staticLibraries/libraryPaths; the archive
   is attached via binaries.all { linkerOpts(...) }, which does not travel to
   consumers. As published, the K/N klibs carry bindings but no machine code, so a
   downstream app can't link — this already affects the published -linuxarm64
   artifact today. Either embed staticLibraries in the .def or ship the archive
   in the klib. (Happy to split this into its own issue if you prefer.)
2. Apple CMake invocations: -DCMAKE_SYSTEM_NAME=iOS against the iphoneos /
   iphonesimulator SDKs (+ macosArm64), one cmake-build-<target> dir and a
   configure/build/package task triple each, mirroring the existing -PcrossArm64
   pattern.
3. Registration: K/N has no ServiceLoader; installNativeKernels() is a manual
   call, and PlatformCpuOpsFactory.apple.kt currently registers only the scalar
   provider. Needs a small design decision (documented startup call vs a hook), and
   ideally composes with the existing AccelerateCpuOps (vDSP dense FP32) so Apple
   gets NEON quant kernels and Accelerate FP32 rather than all-or-nothing.
4. CI: Apple legs in publish.yml's build-native matrix and an
   iosSimulatorArm64Test lane (currently the only automatable way to exercise the
   iOS NEON path).

Track B — Android:

Real Android apps (ART/Compose) can't consume androidNativeArm64 klibs — that
target only serves pure Kotlin/Native binaries. The practical route is a small JNI
bridge: an AAR module with externalNativeBuild over the same C sources, jniLibs
for arm64-v8a, thin JNI shims per kernel, and a priority-100 KernelProvider in
androidMain registered from PlatformCpuOpsFactory.android.kt. It works against
the existing heap FloatArray/ByteArray kernel SPI, no storage redesign needed.
(If the K/N androidNativeArm64 route is wanted anyway, note
skainet-backend-api doesn't declare androidNative targets yet, while
skainet-lang-core already does.)

Small kernel gap for either track: q4_0_matmul.c is scalar-only (no NEON
body), while q8_0/q4k/q5k/q6k have one. Q4_0 is one of the priority mobile formats,
so it deserves a NEON body as part of this work.

Scale of the ask

At 1 tok/s on Android we need roughly 5x to be usable, not the full 21x the JVM
gets — even a partial result unblocks real apps. Related: #910 (vectorized eager
CPU kernels for K/N), #722 (accelerated CPU ops for linux native targets).

We have a physical-device test bench (Android phone + iPhone) and will gladly test
any branch or snapshot and report numbers back. Happy to contribute PRs for the
tracks above, starting with the cinterop packaging fix.