docs(release): spike downloadable server binary matrix

[localai-bot]

Claim

Claims ENG-RELEASE-BINARIES for the accepted records/spec-only release spike. Tracks #117. Row remains SPIKE; no archive or implementation is claimed.

User-reviewed primary artifact contract

• one primary adaptive CPU binary per OS+host ABI, built to a conservative baseline and runtime-dispatching only to compiled, exactly probed ISA tiers
• x86_64 baseline works without AVX2; current F16C/AVX2/AVX-512 tiers are inventoried, while VNNI/AMX or later tiers are named only where real kernels and CPU+OS-state gates exist
• aarch64 baseline is portable/NEON with independent HWCAP/sysctl-gated DotProd/i8mm and future real-kernel tiers
• no global -march=native; forced-tier mutation gates execute feature-poor and feature-rich paths/emulation, and the manifest lists compiled tiers and required OS state
• one primary fat CUDA binary per OS+host ABI; at least separate Linux x86_64 and Linux aarch64 ELF artifacts
• each fat CUDA artifact covers all supported SMs: 80, 86, 87, 89, 90a, 100a, 103a, 110, 120a, 121a
• cross-family per-source gencode must land first; fast-path TUs compile only for compatible SMs and portable TUs cover the full list
• Triton AOT packaging/runtime dispatch embeds and exact-SM-selects every available tree (80/86/89/90a/100a/121a); 87/103a/110/120a retain portable fallbacks
• optional single-SM archives are diagnostic/performance variants, not primary downloads
• build/runtime/correctness/performance evidence remains independent per SM/tier; build-only paths stay preview

Package and supply-chain contract

• staged archive help/health/model smoke plus dependency and RPATH/install-name audits
• SHA256, SPDX SBOM, provenance, licenses, version and independent evidence booleans
• least-privilege plan/build/verify/attest/publish stages for dry runs and tags
• ffmpeg and accelerator drivers remain explicit external runtime boundaries
• experimental literal-static musl CPU lane remains preview; ROCm remains blocked

Work order

W1 per-source CUDA gencode -> W2 multi-SM AOT dispatch -> W3/W4 x86_64/aarch64 ISA audits -> manifest/install/archive/workflow -> primary CPU and fat-CUDA bundles -> optional per-SM diagnostics.

Scope

Documentation and canonical records only. No CMake, workflow, source, test, build, GPU, download, service or artifact implementation.

Verification

• scripts/agent-preflight.sh --staged -> EXIT=0, all gates green
• pre-push scripts/agent-preflight.sh && git push -> EXIT=0
• head: 88c21f0e40e0d982423e3047c6234286feec04e7

FOLLOWING_AGENTS_PROTOCOL

[localai-bot]

Integration blocker found after reconciling with current main: CI job device-leakage fails at the PR merge ref because origin/main f161bcc0 itself has two new DeviceType::kCUDA references in src/vllm/multimodal/minimax_h3_video.cpp while the ratchet baseline remains zero. Reproduced directly on origin/main; #129 does not touch this code/checker and remains spec-only. A fresh helper is preparing the red-first fix on the owning H3/surface row. After that fix lands, #129 will be reconciled again, rerun, mutation-reviewed, and only then readied/merged.