Open NPU Lab and correctness-gated local RAG sidecar

[Jonas-Augustinus-Linus]

Summary

This turns the v1.0 compute toolkit into an Open NPU Lab for XDNA1, XDNA2, and evidence-driven future-device work.

• states the MIT/public-handoff mission and gives 15-minute/day/week/research entry paths
• adds a primary-source research map and evidence-labelled hybrid local-LLM roadmap
• corrects stale ecosystem claims, separating this repo's pins, AMD IRON upstream Phoenix results, and published research
• adds a real persistent NPU local-RAG scoring sidecar with complete CPU-oracle and top-k gates
• adds fail-closed corpus/endpoint boundaries and hardware-free CI regression tests
• records the v1.1.0 release and adds an open-experiment issue form

Hardware evidence

Final Strix Point RyzenAI-npu4 candidate:

• rebuilt the toolchain-keyed 256x256x256 bf16 VMFB
• all 65,536 splat outputs matched the CPU reference exactly
• three semantic queries x seven persistent calls
• all 65,536 outputs checked per call
• exact CPU top-3 parity
• worst normalized-max errors: 0.138%, 0.149%, 0.115% (5% gate)
• ordinary Open NPU Lab corpus query passed the same complete gate

The original XDNA1 evidence remains published, but the current exact-lock full sidecar rerun is explicitly marked outstanding.

Validation

• ./scripts/validate-repo.sh: 17 shell checks and 113 content/link checks PASS
• 19 hardware-free local-RAG boundary/policy tests PASS
• isolated clean NumPy 2.5.2 venv test PASS
• git diff --check and shell syntax PASS
• two independent final reviews: blocker/high/medium = 0

Honesty boundaries

This does not claim a turnkey whole-model XDNA1 server, trained embeddings, CPU speedup, or repo-measured energy savings. CPU fallback, synthetic examples, compile-only work, current-lock XDNA1 gaps, and upstream/paper evidence are all labelled separately.

Planned release: v1.1.0.