v1.0.0: publish the verified XDNA2 Linux toolkit

[Jonas-Augustinus-Linus]

Outcome

Completes the public XDNA2 path alongside the existing XDNA1 work: strict device detection, pinned source builds, full CPU-reference verification, persistent native/Python invocation, application examples, and real-hardware evidence.

What changed

• detect only verified npu1_4col and Strix Point npu4 VBNV/geometry pairs
• pin source commits, Python/tool versions, critical wheel filenames, and SHA-256 values
• add one-command verify-stack.sh quick/full acceptance contracts
• make matmul, wake-word, ONNX MLP, and camera compute paths fail closed and preserve last-known-good artifacts
• publish a real Strix Point GIF plus five-language XDNA2/support documentation
• add the open LLM roadmap, security policy, issue/PR templates, Dependabot, citation metadata, changelog, and hardware-free CI

Hardware verification

Device: Ryzen AI 9 HX PRO 370, RyzenAI-npu4, Strix Point 4x8, Ubuntu 26.04, Linux 7.0.

• 1 GiB memlock: ./scripts/verify-stack.sh --full PASS
• i32 and bf16: every output matches the CPU reference exactly
• native and Python persistent runners: 16,384/16,384 outputs correct
• wake-word persistent pipeline: wake 94 vs deterministic noise 31
• ONNX MLP: 3.613% normalized max error vs bf16 CPU reference, below the documented 5% bound
• native bfp16ebs8 512 cubed and 1024 cubed: CPU-reference PASS
• npu4 camera processing core: 921,600/921,600 values match CPU

Static and failure-path verification

• ./scripts/validate-repo.sh
• clean committed-clone validation
• 16 shell entry points and 104 candidate files
• 614 local Markdown targets and all 8 GIFs decoded
• GitHub YAML and CFF 1.2 schema validation
• locked upstream commits/wheel assets and hashes independently checked
• legacy Peano21 migration, atomic VMFB failure preservation, and temporary-file cleanup regressions
• no bundled ELF/VMFB/bytecode, secrets, or private paths

Honest boundaries

• Current exact v1 lock was hardware-revalidated on Strix Point. Phoenix has earlier real-hardware evidence and offline compile regression coverage; current-lock XDNA1 reconfirmation is requested. Hawk Point has no separate hardware result.
• npu5/npu6 are rejected rather than silently mapped to npu4.
• bfp16ebs8 passes through K=1216 and first fails at K=1280 in the documented sweep.
• W4A16 remains compile-only; a full arbitrary LLM runtime is the published roadmap, not a present claim.
• The XDNA2 camera compute core is verified, but full /dev/video10 GStreamer loopback/FPS was not available on this host.

Public mission

The project stays MIT-licensed and free so students, independent developers, researchers, and small teams can reproduce the foundation, build many different Linux NPU LLM systems, and contribute evidence and kernels back to the ecosystem.