Share the quantity vocabulary between both generators

[matt-edmondson]

Lifts QuantityVocabulary out of Semantics.Cpp so the C# generator runs the same physics model.

What this found

Only the C++ projection knew whether a declared relationship was dimensionally true — it needed the exponents to write Quantity<D>, so it could not avoid multiplying them out. The C# generator checked that a relationship's names resolved (SEM001) and that its forms existed (SEM003), and then emitted the operator.

So Sensitivity * Pressure -> ElectricPotential has been shipping as

public static VoltageMagnitude<T> operator *(Sensitivity<T> left, Pressure<T> right) =>
    VoltageMagnitude<T>.Create(left.Quantity * right.Quantity);

Sensitivity is M⁻¹L⁻¹T²I and Pressure is ML⁻¹T⁻², so their product is L⁻²I — but ElectricPotential is ML²T⁻³I⁻¹. It compiles, it runs, and it computes the wrong physics. C++ has refused it by name since the projection existed; CLAUDE.md documents it as a pre-existing metadata bug.

With the check shared, the C# side now reports all five of the relationships CLAUDE.md lists, verbatim from the build:

SEM008: Relationship Sensitivity * Pressure -> ElectricPotential is not dimensionally true:
        L⁻¹ M⁻¹ T² I * L⁻¹ M T⁻² is L⁻² I, and ElectricPotential is L² M T⁻³ I⁻¹.
SEM008: Relationship Torque * AngularDisplacement -> Energy ...
SEM008: Relationship MomentOfInertia * AngularVelocity -> AngularMomentum ...
SEM008: Relationship MomentOfInertia * AngularAcceleration -> Torque ...
SEM008: Relationship dot(Force, Length) -> Energy reduces to a signed value ...

Shape

QuantityVocabulary, DimensionVector and the resolution around them move to Semantics.Vocabulary. Both generators keep their own reader of dimensions.json — they differ in what they carry beyond the physics, which is the point QuantityMetadata's remarks already made — and each projects onto a small DimensionDeclaration shape the vocabulary owns. That is one ~50-line adapter per side.

DimensionVector.ToCpp stayed behind as an extension in Semantics.Cpp. Writing a dimension as a C++ template argument list is that target's business.

Shared source, not a project, argued at length in Semantics.Vocabulary/README.md. Briefly: Semantics.SourceGenerators is a Roslyn component that bundles its dependencies for analysis-time loading, and its .csproj already carries a long comment about those bundled facades colliding with in-box types; adding an assembly to that bundle is the kind of thing that breaks subtly. And Semantics.Cpp is packed, so a project reference becomes a published NuGet package — a public surface commitment for what is so far an implementation detail. Nothing passes these types across a boundary and no project references both consumers, so two compiled copies cost nothing and there is no CS0433 risk.

The price is real and is documented at each site: the files compile under two analysis configurations, so everything in them stays inside netstandard2.0's surface — no System.HashCode, no range expressions, no string.Contains(char). If a third consumer appears, this should become a project and take the packaging decision on its merits.

Three judgement calls, all reversible

SEM008 reports; it does not drop the operator. Removing Sensitivity * Pressure from a shipped package is a breaking change nobody asked for, and fixing the metadata is a physics call CLAUDE.md already declines to guess at. Making it drop, matching C++, is a one-line change — say the word.

SEM008 is suppressed in Semantics.Quantities only. DiagnosticCatalog offers Warning and Error and nothing quieter, and ktsu.Sdk sets TreatWarningsAsErrors, so a diagnostic here cannot be both visible and non-fatal. Three of the five are provably unfixable by choosing exponents (the r×F versus τ·θ contradiction), so leaving it on would mean the repo does not build until a physics decision is made. The suppression is scoped to that one project with a comment naming the five, so a consumer editing metadata still sees the warning.

UnkeepableRelationshipTests pins the set to exactly those five, so a sixth fails there rather than disappearing into the suppression — and so does a fourth, which would mean somebody fixed one and the list and the NoWarn should shrink together.

One invariant changed

TheRealMetadataReportsNothing and EveryGenerator_EmitsSourcesWithTheCanonicalHeader asserted that the real metadata produced no generator diagnostics at all. That was true only because nothing was checking. Both now assert nothing unexpected, excluding SEM008 by identifier, with the set pinned by the test above. The first is renamed to match.

Verification

• dotnet build and dotnet build -c Release — clean, 0 warnings, every target framework
• Semantics.Test: 1136 tests, 0 failed (1128 passed, 8 Windows-only skipped)
• Semantics.Cpp.Test: 27 passed, including the tests that compile the whole generated vocabulary with real g++/clang++
• Generated output is byte-identical — git status on Semantics.Quantities/Generated/ is empty, so the lift moved the code without moving the behaviour

Not in scope

The C# generator still builds its own type and operator lists for emission; only the check is shared so far. Making QuantityVocabulary.Types drive C# emission is the larger half of "one physics model" and would risk changing committed generator output, so it is worth doing on its own.

🤖 Generated with Claude Code

https://claude.ai/code/session_01QGCMUrT3jBgmANHNHcPmBf

---

Generated by Claude Code

[sonarqubecloud]

Quality Gate passed

Issues
6 New issues
0 Accepted issues

Measures
0 Security Hotspots
92.5% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud