A handler returning a value type crashes a Native AOT image at startup: PrepopulateRoutingCache closes EmptyMessageRouter<Guid>, and no root can reach it #4848

Found while researching #4810, by prediction first and one native publish second. Not a user report.

What happens

A message handler that returns a value type kills a PublishAot application at startup:

System.NotSupportedException: 'Wolverine.Runtime.Routing.EmptyMessageRouter`1[System.Guid]' is missing native code or metadata.
   at System.Reflection.Runtime.TypeInfos.RuntimeTypeInfo.MakeGenericType(Type[])
   at JasperFx.Core.Reflection.TypeExtensions.CloseAndBuildAs[T](Type, Object, Type[])
   at Wolverine.Runtime.WolverineRuntime.RoutingFor(Type)
   at Wolverine.Runtime.WolverineRuntime.PrepopulateRoutingCache(IEnumerable`1)
   at Wolverine.Runtime.WolverineRuntime.<StartAsync>d__165.MoveNext()

Reproduced in Wolverine.AotSmoke.Fisher (net9.0, Generating native code asserted, then reverted) by adding exactly this and nothing else:

public record AotProbeQuery(Guid Id);

public static class AotProbeStructHandler
{
    public static Guid Handle(AotProbeQuery query) => query.Id;
}

That is the ordinary request/reply shape — await bus.InvokeAsync<Guid>(new CreateOrder(...)) — in any durable-store application. PrepopulateRoutingCache runs in every mode except MediatorOnly and Serverless, so the host never starts.

Why

Three facts, each already written down somewhere, that nobody had put together:

1. HandlerChain.PublishedTypes() yields every variable a handler creates, so a Guid return is a "published type" and HandlerGraph.AllMessageTypes() hands it to the routing warm-up. (GH-4840, GH-4841, GH-4843: sharpen the side-effect AOT roots, guard the HTTP rooting order, run a Balanced native lane #4847 removed INotToBeRouted return types from that list for the side-effect sibling of this bug. A Guid is not one of those: Wolverine will cascade it if the handler is not being invoked for a response, so it is legitimately routed.)
2. HandlerRegistryCodeFile emits MessageRouter<T> / EmptyMessageRouter<T> roots only for the chain message types, never for published types. Reference-type published types survive anyway, through canonical sharing plus a typeof token (see the measurement on Investigate a static gate for the AOT reflective-generics class, instead of relying on publish-and-run smokes (follow-up to #4765, not for the next release) #4810). A value type cannot: there is no shared template for EmptyMessageRouter<Guid>.
3. [DynamicDependency] over a value-type instantiation roots metadata, not code (Sagas do not work under Native AOT: [DynamicDependency] does not root a value-type instantiation, and SagaSchemaFor is a generic virtual method #4805, rule 1a). So adding the published types to the rooting block would not fix this one; it would emit a root that looks right and does nothing.

How much of a real graph this is

Measured on CoreTests' own handler graph on CoreCLR (a throwaway probe, not committed):

Every one of those three is a startup crash in a native image today.

Fix shapes

In the order I would try them:

1. De-genericize the routers. MessageRouterBase<T> uses T for typeof(T) and for typed overloads that IMessageRouter already exposes as object. A Type-carrying non-generic router removes the reflective close entirely, which is the audit's stated preference ("prefer de-genericizing over rooting wherever the generic buys only a typed delegate") and the only shape that covers value types without generated code. Needs a look at MessageRouter<T>'s topic and partition delegates before committing to it.
2. A generated factory for value-type published types, the Sagas do not work under Native AOT: [DynamicDependency] does not root a value-type instantiation, and SagaSchemaFor is a generic virtual method #4805 rule-3 pattern: codegen emits new EmptyMessageRouter<Guid>(runtime) by name, and RoutingFor consults it before closing. Works, but it is a second direct-construction table next to _frameworkRouterFactories and it only covers what the registry saw at codegen write time.
3. Not a [DynamicDependency] over EmptyMessageRouter<Guid>. Rule 1a; it would be a root that resolves and does not generate.

Whichever lands, the Fisher lane should keep a Guid-returning handler so this stays measured.

Related: #4810 (the research this fell out of), #4805 (rule 1a), #4847 (the INotToBeRouted sibling).

🤖 Generated with Claude Code

https://claude.ai/code/session_017DbTkhs3JxSMCWasTzPdRq