ADR-0022: The .NET extractor stays the Rust metadata reader; the C# fallback is not invoked

• Status: Accepted
• Date: 2026-09-22
• Supersedes: the "otherwise" branch of ADR-0003
• Derives from: ADR-0003 (the trigger), wave 0 plan § 1.6 (the denominator rule, fixed before measuring), design § The fallback, decided now rather than under pressure
• Constrains: rb-extract-dotnet, Wave 2 plan (its .NET sub-waves stay on the Rust reader)
• Implemented by: Wave 0 plan, sub-waves 0D and 0E
• Requirements: FR-EXT-DN-02

Context

ADR-0003 fixed the rule: if the Rust reader attributes at least 99% of the .NET oracle repositories' types to a source file within the three-week window, it stays; otherwise a C# dotnet tool replaces it. The wave 0 plan operationalised "attributed" before any measurement (§ 1.6 and Step 9): the trigger figure is (pdb + inferred) / (types - excluded), where excluded is <Module> and every type carrying CompilerGeneratedAttribute, and inferred covers a nested type taking its enclosing type's file and the <TypeName>.cs naming convention.

The reader was built in sub-wave 0D and measured on 2026-09-22 over the eleven .NET oracles in testbeds/manifest.yaml at their pinned commits, each built with -c Release -p:DebugType=portable (conformance/archunitnet/scripts/spike-b-attribution.sh; per-repository JSON under conformance/archunitnet/attribution/).

Ten of the eleven repositories were measured; TNG/ArchUnitNET's global.json requires a newer SDK than the measuring machine had, and the weekly spike-b workflow measures it with the current SDK. One project targets .NET Framework (Nager.Date); no assembly had a Windows or missing PDB. TestAssembly attributes 45 of 45 types through the PDB.

Of the 32 unattributed types, most are compiler-synthesised without the attribute that would exclude them (<>y__InlineArray21, a synthesised collection's nested enumerator, <PrivateImplementationDetails>+__StaticArrayInitTypeSize=88); the rule counts them against the reader, as it was fixed to. The rest are Razor Page models (LoginModelinLogin.cshtml.cs`) and one nested enum in a type with no sequence points.

Decision

• The trigger figure, 0.9929, is at or above 0.99 inside the window, so rb-extract-dotnet's Rust reader is the .NET extractor. Rulebearing.Extract is not built, and rb-ingest does not gain the fallback entry point.
• The margin is recorded as it is: the decision depends on 39 types attributed by naming convention, and a PDB-only reading (0.9335) would not have cleared the bar. Wave 2 therefore treats attribution quality as a live requirement, not a settled one:
  • the weekly spike-b figure continues through the window and afterwards; a pooled trigger figure below 0.99 on two consecutive weeks reopens this decision in a new ADR;
  • wave 2 reports the PDB-only share alongside the trigger figure in its .NET status table, so a regression in the part that does not depend on convention is visible.

Consequences

• One toolchain and one binary for the .NET path, as ADR-0002 intended.
• Attribution by convention is a documented state (attribution: inferred), so a rule author can see which findings rest on it.
• ADR-0003's status line changes to "Superseded by ADR-0022", the only edit an accepted ADR permits (ADR-0001).

Alternatives considered

• Count only PDB attribution and invoke the fallback. Rejected: the operational definition was fixed in the plan before measuring and approved with it; changing it after seeing the number would make the pre-committed rule meaningless in both directions.
• Extend the exclusions to every compiler-synthesised name (types whose names begin with <). Rejected for the same reason; it would raise the figure to 0.9993 (29 of the 32 unattributed types have such names), and it is recorded here only so a later ADR can revisit the rule openly.