Racing placement: Simulation facet + dogfood composition

name: Racing placement split overview: Extract the headless racing simulation into Novolis.Simulation.Racing (platform facet, ML-agnostic). Keep neural evolution glue in novolis-dogfooding as app composition code. Do not add MachineLearning.Racing or any domain-specific ML packages. todos:

• id: sim-racing-project

content: Create Novolis.Simulation.Racing project; move ML-agnostic sim code from dogfood; rename namespaces status: completed

• id: sim-racing-tests

content: Add Novolis.Simulation.Racing.Tests (11 test files); wire slnx, packages.json, registry status: completed

• id: dogfood-slim

content: Remove NeuralRacing.Sim; keep Training + Neural controllers in app; PackageReference sim + neural status: completed

• id: dogfood-tests

content: Reduce NeuralRacing.Tests to 2 trainer test files only status: completed

• id: governance

content: Update library-boundaries, frank docs, migrate-frank-slice; optional wave-12 brief status: completed isProject: false

Problem

MachineLearning.Racing would be reverse leakage: a domain scenario (racing + evolution) packaged under a repo that should stay orthogonal building blocks like novolis-math or novolis-raylib.

Per library-boundaries.md:

• Math / Physics / Simulation = closed stack; Simulation depends only on Math + Physics.
• MachineLearning = orthogonal (not in stack); apps compose it with other repos.
• Apps own cross-repo glue: "Apps may reference any combination; they own cross-repo glue."

Racing today in novolis-dogfooding/apps/NeuralRacing is a hidden library (~55 sim files + 266 tests) with an invalid dependency: NeuralRacing.Sim ProjectReferences Novolis.MachineLearning.Neural — the sim layer should not depend on ML.

flowchart TB
  subgraph ml_repo [novolis-machinelearning]
    Core[Core]
    NeuralAbs[Neural.Abstractions]
    Neural[Neural]
    AutoMl[AutoMl]
  end

  subgraph sim_repo [novolis-simulation]
    SimRacing[Simulation.Racing]
  end

  subgraph dogfood [novolis-dogfooding]
    App[NeuralRacing app]
    Trainer[Training folder]
    NeuralCtrl[NeuralRaceCarController]
    TrainerTests[Trainer tests only]
  end

  Neural --> NeuralAbs
  SimRacing --> MathBCL[BCL numerics only at first]

  App --> SimRacing
  App --> Neural
  Trainer --> SimRacing
  Trainer --> Neural
  NeuralCtrl --> SimRacing
  NeuralCtrl --> Neural

What belongs where

[novolis-machinelearning](d:\novolis\novolis-machinelearning) stays four packable facets only:

• Novolis.MachineLearning.Core
• Novolis.MachineLearning.Neural.Abstractions
• Novolis.MachineLearning.Neural
• Novolis.MachineLearning.AutoMl

No new ML packages. README already states racing is not in this repo — keep that.

Why Simulation.Racing fits the stack

• Matches existing facets (World, Kinematics): domain runtime under Simulation, not under ML.
• RaceSimulation is orchestration over discrete ticks (like Simulation’s clock/world model), with custom kinematics — does not require Novolis.Physics.* on day one.
• Must not reference Novolis.MachineLearning.* (enables sim tests without ML on the graph).

Not Simulation.Racing + MachineLearning.Racing: the second package would re-introduce domain naming in the wrong repo and tempt Simulation ↔ ML coupling in platform code.

Phase 1 — Add Novolis.Simulation.Racing (novolis-simulation)

1. Create project at src/Novolis.Simulation.Racing/ using the same pattern as Novolis.Simulation.Kinematics.csproj:

• Import Novolis.Simulation.Packaging.props
• PackageId = Novolis.Simulation.Racing
• No project references to MachineLearning
• No project reference to Physics required initially (racing is self-contained); optional Novolis.Math.Geometry later if shared ray/grid helpers appear

1. Move + rename from NeuralRacing.Sim:

• Namespace: NeuralRacing.* → Novolis.Simulation.Racing.* (folder layout unchanged: Tracks, Race, Cars, etc.)
• Exclude: Training/, NeuralRaceCarController.cs, INeuralRaceCarController.cs

1. Add tests tests/Novolis.Simulation.Racing.Tests/:

• Port 11 files from NeuralRacing.Tests/Sim (all except the two EvolutionaryRacingTrainer* files)
• Namespace: Novolis.Simulation.Racing.Tests
• TUnit exe pattern like Novolis.Simulation.World.Tests
• Small local TestInfrastructure (copy BaseTest / StructuredTestOutput from dogfood or ML test support — do not reference ML test projects)

1. Wire repo:

• Novolis.Simulation.slnx
• .novolis/packages.json
• novolis-registry/packages/novolis-simulation-racing.json (new stub)
• dotnet build + run Racing tests in simulation CI

Phase 2 — Slim dogfooding to composition-only

1. Delete apps/NeuralRacing/NeuralRacing.Sim/ after move.

1. Restructure apps/NeuralRacing:

   apps/NeuralRacing/
     NeuralRacing.csproj          # Exe, IsPackable=false
     Program.cs
     Training/                    # 4 files from old Sim/Training
     Controllers/                 # NeuralRaceCarController + INeuralRaceCarController

• PackageReference to Novolis.Simulation.Racing and Novolis.MachineLearning.Neural (2026.1.* per Directory.Packages.props)
• Remove monorepo ProjectReference into novolis-machinelearning (aligns with dogfooding design); use local GPR or artifacts/nuget-local during dev per nuget-setup.md

1. Slim tests NeuralRacing.Tests:

• Keep only EvolutionaryRacingTrainerTests.cs + EvolutionaryRacingTrainerValidationTests.cs
• Reference app + packages (not Simulation test internals)

1. Update Novolis.Dogfooding.slnx and README apps table: NeuralRacing exercises Simulation.Racing + MachineLearning.Neural.

Phase 3 — Governance and migration script

Update docs to codify the rule “no domain packages in ML repo”:

Optional: add extraction-briefs/wave-12-simulation-racing.md for the Frank.ML racing sim slice.

Deferred (explicitly out of scope)

Success criteria

• novolis-simulation builds; Novolis.Simulation.Racing.Tests passes (~264 tests).
• novolis-machinelearning unchanged package set (4 facets); no racing code.
• novolis-dogfooding NeuralRacing is a thin exe + ~6 glue files + 2 trainer test files; no embedded sim library.
• Dependency graph: Simulation.Racing does not reference MachineLearning.*.