ADR-0003: The .NET extractor falls back to a C# dotnet tool on a measured trigger

• Status: Superseded by ADR-0022
• Date: 2026-09-20
• Derives from: design.md § The fallback, decided now rather than under pressure; § Waves (wave 0 exit criterion)
• Constrains: architecture.md § Extractors, § Risks and their mitigations
• Implemented by: Wave 0 plan § Spike B

Context

Writing an ECMA-335 metadata and portable PDB reader in Rust is a bounded but real cost, and it is the one place the language choice in ADR-0002 could stall the project. The design fixes the decision rule now so it is not made under pressure.

Decision

The wave 0 spike measures, on the .NET oracle repositories at pinned commits, the share of types the Rust reader attributes to a source file through the portable PDB.

• If at least 99% of types are attributed within three weeks of part-time work, the Rust reader in rb-extract-dotnet is the .NET extractor.
• Otherwise the .NET extractor becomes Rulebearing.Extract, a C# dotnet tool over System.Reflection.Metadata that writes the same graph document, and the Rust binary consumes it through rb-ingest. The rule engine, reporters, config, TypeScript and Python sides are unchanged.

The trigger, the measurement script and the result are recorded in the wave 0 plan's status table, and the outcome is recorded as a new ADR that supersedes the open branch of this one.

Consequences

• The extractor boundary in ADR-0010 must be strict enough that the swap touches nothing outside rb-extract-dotnet and rb-ingest.
• The fallback costs two toolchains to build and a .NET runtime on the host, which every .NET repository already has.

Alternatives considered

• Decide after wave 1. Rejected: the .NET side is wave 2's critical path, so the decision must precede it.
• C# from the start. Rejected: it forfeits the single binary for the .NET path when the Rust reader may well succeed.