Status: accepted · Date: 2026-09-10

#444 opens Phase 6, a GPU package over ILGPU. CLAUDE.md's first cross-cutting fact is the rule it breaks:

Everything ships net10.0;netstandard2.0 in a single package. netstandard2.0 reaches equivalent behaviour through conditional compilation, never a reduced API.

Fifteen packages honour that today, and the rule is not decoration: the *.NetStandard.Tests mirrors exist so the assemblies shipped to .NET Framework, Mono and Unity are executed rather than merely compiled, and #529 is what happens when that pin slips.

ADR 0074's protocol, applied to the dependency rather than to an incumbent. ILGPU 1.5.3, read from the package on 2026-09-10:

lib/	net471, net5.0, net6.0, net7.0, netstandard2.1
netstandard2.0 asset	none
newest modern asset	net7.0 — consumable by net10.0, but ILGPU itself targets nothing newer
licence	University of Illinois/NCSA Open Source License, read from LICENSE.txt inside the package
declared SPDX expression	none — the nuspec carries <license>LICENSE.txt</license> and a deprecated licenseUrl
1.5.1 → 1.5.2 → 1.5.3	2023-09-13, 2025-03-06, 2025-07-12

Two of those rows are findings rather than facts in passing.

The licence had to be read from the artefact. No SPDX expression is declared, which is the trap decision 0075 named and 0099 met again: metadata silence is not a licence. NCSA is permissive — a BSD-family grant with a no-endorsement clause — so decision 0003 is satisfied. Had LICENSE.txt said GPL, this ADR would be a refusal instead.

netstandard2.0 is not a gap to polyfill, it is absent upstream. The usual order — PolySharp, then a framework-conditional package reference, then a hand-written fallback — closes language and BCL gaps. None of it produces a CUDA or OpenCL accelerator. There is nothing to write.

Lodestar.Gpu targets net10.0 alone, and is the only package in this repository that does not ship netstandard2.0.

net10.0 rather than net8.0, which #444 proposed: ILGPU's newest asset is net7.0 and is consumed identically by either, so a second target framework would add a matrix dimension and test nothing ILGPU distinguishes. One target, the repository's own.

The rule protects a promise to a consumer: take Lodestar.Text on .NET Framework and the API is the same API. Lodestar.Gpu makes no such promise because it cannot be taken there at all — and a package absent on a framework breaks no promise about its shape on that framework. The failure mode the rule prevents is a package that ships to netstandard2.0 with less in it, which is the opposite of this.

Lodestar.Gpu is therefore additive and never a dependency of a core package. No src/ project may reference it: a core package taking an edge to a net10.0-only library would make netstandard2.0 unreachable for that package, which is the breach this decision is careful not to be. The edges run the other way — Lodestar.Gpu → Lodestar.Abstractions, and later → Lodestar.Embeddings and → Lodestar.Text — so the SIMD path stays the only path a netstandard2.0 caller has, and it stays complete.

tools/check_netstandard_guards.py iterates the mirrors that exist; it does not require one per package. So Lodestar.Gpu arriving with no mirror is silence, not a finding — indistinguishable from a mirror somebody forgot. This record is what makes the absence deliberate, and CLAUDE.md's package table gains a tier column entry saying so in the same change.

• A sixteenth package, net10.0 only, satellite tier by 0076 and interop-named by 0089's reasoning, though it adapts hardware rather than another library's types.
• No *.NetStandard.Tests mirror, deliberately.
• THIRD-PARTY-NOTICES.md gains ILGPU with NCSA as read from LICENSE.txt, not from the nuspec, and a note that the package declares no expression.
• A kernel's baseline is the SIMD path on the same machine, never the scalar loop — 0102 has the protocol.