Release checklist: cut 0.39.0 to ship the Android JNI backend + mobile fixes (unreleased on develop)

[michalharakal]

Tracking checklist for the first engine release that carries the mobile/on-device work merged since 0.38.0. Several fixes are on develop but unreleased, so no downstream consumer (incl. SKaiNET-transformers, which pins skainet = "0.38.0") can see them until this ships.

What this release propagates to consumers

• createRandomAccessSource returns null on Android: full-file heap load OOMs on a 138 MiB GGUF (working ~40-line fix included) #922 Android createRandomAccessSource — fixes the streaming-load OOM (the transformers Android LLM path hits this directly)
• StreamingGgufParametersLoader silently skips tensors it can't load (Q4_1, but also Q4_0/Q5_0/Q5_1) — load "succeeds" with missing weights, crash comes later #919 GGUF loader fail-fast (silent-missing-weights → hard error)
• StreamingGGUFReader maps only Q4_K/Q8_0 to real TensorEncodings — Q6_K/Q5_K/Q4_0/Q5_0/Q5_1 become Opaque(…, 0) and corrupt every memory report #928 GGUF → TensorEncoding mapping (correct memory reports)
• fix(backend): embed the kernel static archive into published K/N klibs (#941) #942 cinterop archive embedding — published K/N klibs become linkable (matters for transformers iOS: kllama/kgemma)
• feat(backend): NEON body for the Q4_0 matmul kernel (#920) #939 Q4_0 NEON kernel
• TensorStorageFactory violates its own copy/borrow contracts: "zero-copy" borrowFloatArray copies, "borrowed" fromTensorData owns #927 / TensorStorage transfer API cannot perform its core transfer: copyToHost() throws for FileBacked (MMAP_WEIGHTS), copyMaterialize() misses Aliased #929 / TensorData.copyToFloatArray() default implementation is rank-broken — flat index into vararg get() trips every impl's arity check #930 / Memory diagnostics: MemoryTracker.recordCopy discards the source label every caller passes; ActiveMemoryTracker is a mutable global #931 lang-core storage/diagnostics hygiene
• feat(backend): skainet-backend-jni-cpu — Android JNI bridge for the NEON kernels (#920) #943 (+ fix(jni): eager library init; emulator-validated parity suite; gate q4k/q6k on #944 (#920) #945) skainet-backend-jni-cpu — Android JNI NEON bridge. Measured on a Pixel 8a: ~24 tok/s SmolLM2-135M Q8_0 decode projection (V82_DOTPROD tier) vs 3.8 tok/s scalar — clears the field-report usability gate ~6x. This is the artifact that turns Android inference on.

Release checklist

• Merge fix(jni): eager library init; emulator-validated parity suite; gate q4k/q6k on #944 (#920) #945 (JNI parity re-enabled with the RMS gate; on-device + emulator green).
• Close the AAR-publishing gap — the publish job runs ./gradlew publish on macOS, which now includes the skainet-backend-jni-cpu Android library, but the runner has no pinned Android SDK/NDK, so the AAR's .sos won't build reliably. PR incoming; needed before the release actually ships the JNI artifact.
• Decide the version number (0.39.0 suggested — new feature: JNI backend).
• Bump VERSION_NAME in gradle.properties; update CHANGELOG.md [Unreleased] → the version.
• Confirm the release builds & publishes all artifacts, including the new sk.ainet.core:skainet-backend-jni-cpu AAR (with arm64-v8a + x86_64 .sos and consumer ProGuard rules) and the linkable K/N klibs from fix(backend): embed the kernel static archive into published K/N klibs (#941) #942.
• Tag & publish to Maven Central.
• Follow-up in SKaiNET-transformers: bump skainet = "<version>" in gradle/libs.versions.toml; then transformers#272 (SmolLM2 reproducer) can prove the end-to-end Android generation path with NATIVE_OPTIMIZED + the JNI AAR dependency (the provider auto-registers via the PlatformCpuOpsFactory.android ServiceLoader install — no transformers code change needed, just the dependency).

Note on the mobile win reaching apps

For a transformers-based Android app to actually get the speedup it needs to (a) depend on this release, and (b) add sk.ainet.core:skainet-backend-jni-cpu as a dependency so the .sos ship in the APK. The provider is then discovered automatically. Worth a short "using native kernels on Android" doc note alongside the release.

[michalharakal]

Done: 0.39.0 is tagged and published (Central verified) — Android JNI backend AAR, streaming/random-access IO fixes, packed Q4_0/Q5_0/Q5_1 loading and the Q4_0 NEON kernel all shipped. Downstream consumed it same-day (transformers 0.39.x line, sl2610 pin bump).