Support message types from referenced assemblies as nested fields and union members (F1)

[Aaronontheweb]

Stack, bottom to top: S1 #8525 -> S2 #8526 -> S3 #8527 -> object elements #8528 -> S4 #8530 -> S5 #8532 -> S6 #8533 -> F1 #8534 -> F2 #8536 -> F3 #8537 -> S7 #8538. Each PR is one commit on top of the one below it. This PR is F1, the first feature step.

What changes

A serializer can now use message types declared in a referenced assembly as nested field types and as union members. Before, both cases failed, and with the wrong diagnostic id. The design record is Decisions 16 to 21 in openspec/changes/messagepack-sourcegen-validation/design.md.

• Schemas from metadata. A new step, MetadataSchemas, finds the referenced types the local messages and serializers name, reads each one's [AkkaSerializable] and [AkkaField] attributes from metadata, and builds its schema with the same code local types use. The schema is identical to what the other assembly's own generator saw. It follows nested references breadth-first. Its output is a table of plain values keyed by TypeKey, and it feeds the resolve step directly.
• Caching holds. The step runs on every edit, but its output compares equal whenever the set of referenced type keys is unchanged, so resolve and emit stay cached. Two scenarios prove it: an unrelated edit regenerates nothing, and an edit to the referencing message regenerates only its serializer.
• AKKASG039, referenced type or member is not accessible. An error at the local property or union member, naming the upstream type and member. It fires for an internal type without InternalsVisibleTo, or a property with a restricted getter. One limit is in the design record: Roslyn's metadata reader hides a member this compilation cannot see at all, so a wholly inaccessible property drops out of the schema silently. A restricted getter on a visible property is caught.
• The two wrong ids are gone. A nested type from a referenced assembly reported AKKASG023 and a union member reported AKKASG015. Both cases now succeed. A referenced type that is not marked stays an error with the existing cross-assembly hint.
• Golden output. A new corpus case covers a nested type and a union member from a referenced assembly, and a generic definition from a referenced assembly closed locally. The emitted code matches the local case byte for byte except for the namespace. No existing golden, wire, or model snapshot changed.
• Docs. The guide's cross-assembly section now describes what works and where errors show up. Decision 16 gets a short addendum, and its task is marked done.

Baseline cases

CrossAssemblyBaselineSpec pins six cross-assembly behaviors with diagrams. Cases 1 (nested field type) and 2 (union member) flip to success, and their diagrams are rewritten. The plan listed case 5 as a third flip; that was a numbering slip, since the generic-definition case already worked. Cases 3 to 6 stay for F2 and F3.

Cost

Base is the S6 tip. The corpus has no referenced-assembly types, so every row should be flat, and it is. A first cut allocated up to 18 percent more on some rows because the reference walk was an iterator method, which allocates its state machine even when it yields nothing. It now writes into a reused set.

Row	Base (S6)	This PR
Fresh driver, full corpus	22.8 ms, 18.93 MB	22.3 ms, 18.93 MB
Warm driver, one field renamed	16.1 ms, 9.51 MB	16.1 ms, 9.51 MB
Warm driver, comment edit	8.6 ms, 3.54 MB	8.8 ms, 3.55 MB
Warm driver, unrelated file edited	7.2 ms, 3.52 MB	7.4 ms, 3.52 MB

How it was checked

334 tests pass (327 plus 2 golden, 3 for AKKASG039, 2 caching scenarios). Generator and Akka.Remote build with warnings as errors. No public API change. cspell passes on the guide.