release: 0.52.0 — self-healing dispatch (ternary included), androidNative chain, decode-shaped FP32

[michalharakal]

The release branch for 0.52.0 — now the single release: since no 0.52.0 was ever published to Maven Central (the locally-consumed artifacts were an unsigned publishToMavenLocal), the previously local release branch has been merged with develop so the release includes everything on the 0.52 line: the self-healing dispatch arc (#1220/#1221), androidNative arm32/arm64 across the downstream chain (#1239), and the ternary packs joining the self-healing SPI (#1240/#1241) — folded into the one [0.52.0] CHANGELOG section, dated 2026-09-01. Supersedes #1242.

Release-prep commits follow the established pattern (changelog entry, README highlights + BOM coordinate, antora/tutorial version bumps, release: 0.52.0, regenerated matrices), inherited from the original branch and extended with the ternary fold-in.

Downstream validation

SKaiNET-transformers#370 consumes these artifacts (repointed to 0.52.0) with zero explicit kernel installs: BitNet-2B4T decodes at full speed (6.16 tok/s, two-stage) on discovery alone, and the skainet-decode sample's kernel counters light up through KernelDispatch.defaultSink (56.21 GB/s effective bandwidth measured).

After merging — finishing the release

The Central publish is triggered by pushing the annotated tag (the 0.51.0 pattern: annotated tag, SKaiNET <version> — <headline> subject, prose digest body). Suggested, from the merged release commit:

git tag -a 0.52.0 <merge-commit> -F tag-0.52.0-msg.txt   # message drafted below
git push origin 0.52.0                                    # triggers the Central release
git push origin <merge-commit>:main                       # fast-forward main to the release point

Drafted tag message (mirrors the 0.51.0 annotation style):

SKaiNET 0.52.0 — the engine stops silently running on the scalar floor

KernelDispatch now installs itself on first use through the ServiceLoader-backed
ViewKernelPack SPI — providers, the row-major Q-series packs and the ternary/BitNet
packs (BITNET_B1_58 LUT gemv, fused BITNET_PLANES lm_head) are all discovered, so an
application that never calls an install routine no longer loses every kernel to the
~1000x decoding-reference fallback, and a fallback that does happen says so once,
loudly. Dense FP32 weights are served from any storage kind, decode-shaped matmuls
get a dedicated FP32 GEMV path, and androidNativeArm32/Arm64 build and publish across
the whole downstream chain. Validated downstream: SKaiNET-transformers decodes
BitNet-b1.58-2B-4T at full speed with zero bootstrap code, token-for-token equal to
bitnet.cpp and the HF BF16 reference.

Then merge SKaiNET-transformers#370 and #371 once the artifacts resolve on Central.

Verification: self-heal dispatch test green (ternary kernels discovered cold); matrices regenerated (generateKernelMatrix/generateOperatorDocs, no content drift); REUSE compliance left to CI. Post-release issue-closure analysis in the comment below.

🤖 Generated with Claude Code

https://claude.ai/code/session_01YKeDSK4JF295y53Uvez954

[github-actions]

📖 Documentation Preview

The documentation has been built successfully for this PR.

Generated Files:

• Operator documentation: docs/modules/operators/_generated_/
• JSON schema output: operators.json

Artifacts:

• Download the documentation-preview-1243 artifact to view the complete documentation locally.

This comment will be updated automatically when the PR is updated.

[michalharakal]

What this release lets us close (open-issue analysis, 2026-09-01)

Closes with / right after this release (engine):

Issue	State after release	What remains before closing
#1166 (ternary final gate)	Every technical checklist item done: reuse lint in CI, byte-identical vendor verification, artifact notices (jar+AAR), BitNet.cpp credits, CHANGELOG/third-party section, real-2B4T verification (now token-for-token vs bitnet.cpp AND the HF BF16 reference, transformers#365)	Two human actions: the upstream courtesies to anjaustin/neogpu (text drafted in the issue, needs your go) and the klib-notices decision (jar+AAR carry notices; klibs have no META-INF — document the stance)
#1136 (ternary tracking)	All sub-issues #1137–#1141, #1150 shipped; architecture proof points held; downstream BitNet arc complete	Closes when #1166 does — close-out summary ready: 51.8 GOPS ffm-lut, 15.98× memory, 6.16 tok/s discovery-only 2B4T, three-way greedy parity
#1129 (skainet-decode sample)	Delivered in transformers#369 + #370: GGUF → decode → GenerationMetrics (TTFT 1131 ms, 6.48 tok/s, 56.21 GB/s effective bandwidth), weight forms resolved not configured	Split the Android-activity leg into its own issue, then close
#1204 (on-device sidecar cache)	The mainline it was a fallback for is fully served: AOT converter (0.51.0) + SEQUENTIAL zero-copy mmap + this release's discovery	Decision-close as not-planned (its own text: "deliberately not the first thing to build"), or keep as backlog — your call
#805 (Add REUSE)	Already satisfied by the ternary compliance arc, independent of this release: REUSE.toml, LICENSES/, reuse lint in CI, third-party notices	Closable now

Transformers-side, once this release publishes and #370/#371 merge: #360 (maturity gate — all items done), #335 (tracking: T1/T2 + fused head + two-stage + parity all delivered), #336/#337 (implemented; re-scope comments posted), #344 (add-model guide rewritten in #368), and the bookkeeping-stale migration issues #338/#339/#340/#342/#343. #346 can either close (plan items 1–4 all executed; the guide is now the reference) or stay open as the living convention reference.

Explicitly NOT closed by this release (the big adjacent ones): #920/#910/#722 (Kotlin/Native mobile kernel gap — the ternary KN bridges exist but KN still has no ServiceLoader/self-heal and dense KN kernels remain scalar), #979 (Apple hardware validation), #976/#849 (Antora versioning), #932 (SKEEP-003 storage unification anchor), the metrics lane (#1222–#1235), ground-truth/fuzzing arcs (#981–#988), and the IREE arc (#1148/#447/#261).