lingua-engine/docs/kernel-contract.md

Kernel Contract v0

A draft, not a final decision. The microkernel, the plugin contract, and the hot-reload model for Lingua Engine — a modular engine on C#/.NET, with a constraint that shapes half the decisions below: most of the code, kernel and plugins alike, will be written by an LLM agent rather than a human.

• Stack — .NET 9, C# 13
• Platforms — Linux, Windows
• Kernel — BCL only, no dependencies
• Primary author — an agent
• Goal — Play-in-editor with no domain reload

---

1. Principle: the kernel is a shared language, not "the engine minus plugins"

The tempting version of "everything is a plugin" makes even the ECS a plugin. That's a trap: if every other plugin depends on the ECS plugin, the ECS is the kernel already — just with an extra layer of indirection and none of the stability guarantees a kernel should provide.

What becomes a plugin is behavior, not the shared data model.

The kernel is a lingua franca: the minimal set of types and mechanisms plugins need in order to understand each other at all. Anything two independent plugins are required to agree on lives in the kernel. Everything else lives outside it.

This is how every plugin architecture that survived contact with reality is built — Eclipse, VS Code, OSGi, Bevy: a small, stable, extensible kernel plus everything else on top. Trying to make the shared language itself swappable produces either indirection overhead or a kernel so empty it guarantees nothing.

2. Scope: what's in, what's out

Kernel — roughly 4,000 lines, BCL only

#	Piece	Role
01	World	ECS storage: entities, components, queries. Sparse sets on typed arrays, zero unsafe — see §7.
02	Scheduler	Frame stages, topological system ordering, parallelism from access conflicts, and debug-mode enforcement of declared access — see §7.
03	Plugin Host	Manifest parsing, dependency resolution, ALC loading, unloading, reload.
04	Service Registry	Publishing and discovering interfaces between plugins. Control path, not the hot path.
05	Event Bus	Decoupled notifications: entity created, asset reloaded, plugin unloaded.
06	Time & Log	Frame clock, fixed-step accumulator, logging interface. Kept minimal.

Plugins — everything else, no exceptions

windowing · render · physics · audio · input · assets · scene-format · animation · ui · scripting · editor-shell · inspector · gizmos · profiler · introspect · build-pipeline · the game itself.

The editor is also just a set of plugins over the same kernel. This is the architecture's real test: if the editor can't be assembled as plugins, the extensibility claim is decorative. A game build is the same kernel minus the editor plugins.

Two channels, two costs

Plugins talk to the kernel — and to each other — through two paths with deliberately different prices:

Channel	For	Cost	Frequency
World (ECS components)	Anything per-entity: transforms, meshes, colliders, health. Render reads what physics wrote without knowing physics exists.	Direct memory access, zero allocation, zero dispatch	10⁴–10⁶ / frame
Services (interfaces)	Commands and resources: load an asset, open a window, compile a shader, open an editor panel.	Virtual call, negligible	a handful / scene
Events (bus)	Facts with no fixed consumer at design time: asset reloaded, plugin unloaded, entity destroyed.	Allocation + fan-out to subscribers	tens / frame

The line that must never be crossed. Never write IPhysicsService.GetPosition(entity). A single call is cheap, but that shape of API invites calling it in a loop over entities — and now the plugin boundary sits in the hot path. Position is a component in World, not a service method. Services hand out capabilities; World hands out data.

3. The plugin contract

A plugin is two entry points and a manifest next to them. The manifest is a separate file, not assembly attributes — the host has to build the dependency graph before loading anything, or plugin load order becomes a chicken-and-egg problem with ALC loading itself.

// Engine.Kernel / IPlugin.cs

// A plugin holds no game state. None.
// State lives in World; the plugin is code that operates on it.
public interface IPlugin
{
    // Registration: services, systems, component types.
    void Configure(IPluginContext ctx);

    // Full undo of Configure. Whether this method is honest
    // determines whether the ALC unloads at all — see §5.
    void Shutdown(IPluginContext ctx);
}

public interface IPluginContext
{
    IWorld           World    { get; }  // data
    IServiceRegistry Services { get; }  // Provide<T> / Require<T>
    ISchedule        Schedule { get; }  // systems and ordering
    IEventBus        Events   { get; }
    ILogger          Log      { get; }
}

// plugins/engine.render/plugin.json
{
  "id":        "engine.render",
  "version":   "0.3.1",
  "contracts": "Engine.Render.Contracts.dll",  // Default ALC
  "assembly":  "Engine.Render.dll",             // Collectible ALC
  "dependsOn": {
    "engine.windowing": "^0.3",
    "engine.assets":    "^0.2"
  },
  "reloadable": true
}

// plugins/engine.render/RenderPlugin.cs

public sealed class RenderPlugin : IPlugin
{
    public void Configure(IPluginContext ctx)
    {
        // control plane: hand out an interface, take one in
        var window = ctx.Services.Require<IWindow>();
        ctx.Services.Provide<IRenderer>(new VulkanRenderer(window));

        // data plane: the system reads components directly.
        // Reads/Writes are declared explicitly — the scheduler
        // builds a conflict graph from them and parallelizes the
        // frame, and in debug builds enforces the declaration.
        ctx.Schedule.Add(Stage.Render, SubmitDrawCalls)
           .After("engine.transform:propagate")
           .Reads<Transform, MeshRenderer>();
    }

    public void Shutdown(IPluginContext ctx)
    {
        // undo everything: systems, services, subscriptions, GPU resources
        ctx.Services.Revoke<IRenderer>();
        ctx.Schedule.RemoveAllFrom("engine.render");
    }

    static void SubmitDrawCalls(in Frame f, Query<Transform, MeshRenderer> q)
    {
        foreach (var (xf, mesh) in q)   // ref access, no boxing
            f.Draw(mesh.Handle, xf.Matrix);
    }
}

4. Hot reload: why every plugin is two assemblies

A collectible AssemblyLoadContext only unloads once nothing references its contents. One forgotten event subscription, one live Task, one cached Type — and the unload silently fails to happen, leaking a little more memory on every reload.

The most treacherous reference isn't a subscription — it's the component structs themselves. If a plugin declares struct Transform and World stores a Transform[], the kernel holds a reference to a type from the context you're trying to unload. That plugin will never unload.

This is why every plugin splits into two assemblies:

• Contracts (*.Contracts.dll) — component structs, service interfaces. Loaded into the Default ALC, which lives for the process lifetime and never unloads. World owning references into it is fine, because it isn't supposed to unload.
• Implementation (*.dll) — systems, service implementations. Loaded into a collectible ALC, recreated on every reload. No static state, no data — only code.

References only point from implementation to contracts, never the reverse, which is what lets Unload() actually succeed. In practice, ~95% of iteration is logic changes: instant reload. Changing a component's fields is rare and requires an editor restart.

Reload sequence

1. A file watcher sees a freshly built Engine.Render.dll. The build happens externally, via plain dotnet build — the editor doesn't need its own compiler.
2. The scheduler finishes the current frame and pauses. Reload never happens mid-stage.
3. Shutdown() runs: systems, services, subscriptions, and native resources are torn down. Anything Configure registered has to be undone here, or step 5 fails.
4. A snapshot of this plugin's component data is taken — only if contracts were also rebuilt. Component arrays are copied along with a field descriptor.
5. alc.Unload() + GC.Collect(), then a WeakReference check. If the context doesn't collect, that's a loud error naming the pinning reference — not a silent leak.
6. A new ALC, the new assembly loads, Configure() runs. The plugin doesn't know it was reloaded.
7. Data is restored: old and new field layouts are matched by name. Matches are copied, new fields get default values, removed fields are dropped.
8. The scheduler rebuilds its ordering graph and resumes. Typical budget: 200–400 ms, almost all of it spent waiting on the build.

Leak testing belongs in CI from day one. Load and unload a test plugin 200 times in a row; after each cycle, verify the ALC's WeakReference is dead and working-set memory hasn't grown. This is the one thing that keeps the architecture from slowly degrading — ALC leaks accumulate invisibly and surface months later, by which point the cause is indistinguishable from noise.

5. Play mode without domain reload

Unity's Play-mode wait isn't about compilation — it's about serializing all script state, tearing the domain down, and recreating it. That step doesn't exist here: state never lived in plugin code to begin with. It lives in World, owned by the kernel, untouched by reload and untouched by entering Play.

// Engine.Editor / PlayMode.cs

// Entering Play is a memory copy, not a runtime rebuild.
void EnterPlay()
{
    _snapshot = world.Snapshot();   // array copy, low single-digit ms
    schedule.SetGroup(SystemGroup.Play);
}

void ExitPlay()
{
    world.Restore(_snapshot);       // Play-mode edits roll back
    schedule.SetGroup(SystemGroup.Edit);
}

Play becomes a system-group switch, not a world rebuild. A side effect of the same decision: system code can be edited during Play without restarting — state is preserved. That's the feedback loop the whole engine exists to enable.

6. Where this breaks

Risk	The problem	Mitigation
ALC leaks — the main killer	Unload silently fails from one forgotten reference. Symptom: memory growth after N reloads; cause takes days to find.	200-cycle test in CI. Diagnose pinning references in the host itself, not via an external profiler.
Scope	The kernel is 3–5k lines and a couple of months. The renderer, asset pipeline, and editor are years, and they decide whether the engine ships.	Don't write your own RHI. Silk.NET or Veldrid underneath; originality goes into the architecture on top.
GC in the hot path	Collector pauses against a 16.6 ms frame budget. The managed-runtime tradeoff is accepted, but it demands discipline.	unmanaged structs for components, Span<T> in systems, allocation only at load time. Server GC.
Creeping abstraction	The temptation to hide World behind a "cleaner" interface. Kills performance invisibly and irreversibly.	The rule in §2 is law. Review rejects any service method that takes an Entity.
Plausible-but-wrong code — agent-specific	The agent produces code that compiles, passes a smoke test, and breaks on someone else's GPU — sync, barriers, resource lifetime.	Minimize new subsystems; Silk.NET/Veldrid is risk management, not time-saving. A conformance harness gates every plugin merge.
Contract drift	A contract change requires updating every dependent plugin, and a stale implementation keeps compiling while silently diverging from spec.	Versions in the manifest, plus running every plugin's harness on every build, not just the changed one.

7. Written by an agent, not a human

This isn't an afterthought — it's an input condition. It's why §2's storage is simpler than a "fast" ECS would normally be, and it adds a surface no classic editor needs at all.

What works in our favor:

• A plugin's boundary matches a context window's boundary. Writing engine.physics only requires the kernel API, physics' own contracts, and its own code — nothing else. Modularity chosen for team reasons turns out to also be how you fit a task in an agent's head.
• Blast radius is bounded by the plugin. Plausible-but-wrong code is inevitable; the question is what it can break. The kernel is written once, tested, and frozen — the agent never touches it again after that. A bug in a plugin stays a bug in that plugin.

What has to change:

• No unsafe in the v1 hot path. Stride arithmetic, alignment, a pointer that outlives a GC-triggering call — exactly the code an LLM writes convincingly and wrong, failing as nondeterministic memory corruption. Hence sparse sets on T[] instead of archetype chunks; chunked layout stays an optimization behind the same query API for whenever a profile actually calls for it.
• Explicit over clever. Naming conventions, code generators, reflection magic save a human keystrokes but hide behavior from something that reasons over text. Verbose, explicit system and service registration is a deliberate cost. Reflection stays where it's safe: the editor inspector.

Verification instead of trust. Reads<> / Writes<> declarations must be enforced in debug builds: a system touching an undeclared component fails immediately, with a message naming the violation. For a human this is hygiene; for an agent it's structural — otherwise a wrong access declaration becomes a race that reproduces once in a hundred runs and is otherwise undiagnosable. Same principle for the plugin conformance harness: load, reload 200 times, verify Shutdown fully undoes Configure. The agent needs to be able to tell, on its own, that it's actually done.

Introspection surface

The agent doesn't look at a screen. Whatever the editor shows a human's eyes has to be available as data, or the feedback loop closes on a human and the whole point of fast reload is lost.

# run a scene headless and check an assertion about world state
engine run --headless --frames 60 \
           --scene tests/physics_stack.scene \
           --dump out/world.json \
           --assert "count(Rigidbody where sleeping) == 12"

# why a plugin won't unload — instead of guessing from a profiler
engine diag why-pinned engine.render

The loop this enables: agent edits a system → dotnet build for one plugin → ALC reload (world state untouched) → 60 headless frames → machine-readable dump + assertions → back to the agent, no human in the loop. Fast reload saves a human time on its own; paired with headless runs and a state dump, it becomes a loop the agent can close by itself — and the minutes-to-seconds iteration speedup multiplies by however many iterations the agent can now run.

The kernel is written once and frozen. It's the one place where a mistake is expensive and spreads everywhere. A small kernel isn't only an architectural preference — it bounds how much code has to be correct.

8. Build order

Each milestone ends in a working demo, not a "finished subsystem." The order is chosen so the riskiest bet — ALC unloading — gets tested first, while the cost of changing course is still zero.

	Milestone	Done when
M0	Kernel only. World on sparse sets, staged scheduler with access enforcement, plugin host with ALC, service registry, JSON world dump. No window, no graphics.	An agent runs the full loop from §7 unassisted: edits a headless plugin, rebuilds, reads the changed dump — and the 200-cycle leak test is green.
M1	Window, input, a triangle. Three separate plugins over Silk.NET. First real-load test of the data channel.	The triangle's color changes by editing system code, with no app restart.
M2	Assets and scenes. Hot-reloading asset plugin, scene format, World serialization.	Swapping a texture on disk changes the picture with nothing stopped; a scene loads and saves.
M3	Editor as plugins. Shell, reflection-based component inspector, hierarchy, gizmos, Play/Stop on snapshots.	Entering Play takes under 100 ms — the original complaint about Unity is closed.
M4	One small game, end to end. Physics, audio, a Linux + Windows build pipeline. A 20-minute game, shipped as an executable.	The build runs on both platforms with no editor plugins in the shipped binary.

---

Open questions to resolve before M0: whether Time and Log belong in the kernel or as plugins; whether the Event Bus is needed at launch or whether event-components in World cover its role; whether the set of frame stages is fixed or plugin-extensible; and at what profiling point (if ever) to move from sparse sets to archetype chunks. Assembly names in the examples are placeholders.