Find protocol implementors and union members in referenced assemblies, and enforce placement (F3)

[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 F3, the last feature step. It closes the last pinned failure in the cross-assembly baseline.

What changes

The generator now sees across assembly boundaries in both directions the design asked for, and enforces where a serializer may live. The rules are Decisions 19 and 21 in the design record and Decisions E and G on the decision page.

• Protocol implementors in referenced assemblies count. For each referenced assembly that itself references Akka.Serialization.V2, the generator walks its public types once, collects the marked implementors of each protocol and each marked union base, and adds them to the serializer's closed set, with schemas from the F1 metadata step. Baseline case 6, the last pinned failure, flips: an implementor declared only in a referenced assembly gets a dispatch arm, helpers, and a binding.
• [AkkaUnion] with no member list means every [AkkaSerializable] implementor the generator can see, on both sides of the boundary. AkkaUnionAttribute gains a parameterless constructor, extend-only. The explicit list keeps working, and the existing union diagnostics run over the combined set.
• Placement rules. AKKASG043, error: a type implements a protocol an upstream serializer already owns, and that assembly cannot see this one, so the type would never be dispatched. AKKASG044, warning: a serializer with no messages anywhere and no registrations. Rule 3 extends AKKASG031 across assemblies with a new message, as the decision page labels it. Rule 4 has no id: the generated "unsupported type" exception now names both the value's assembly and the generating assembly. That is why six existing golden files changed by two lines each.
• The startup one-owner check. When registrations are composed into a setup, one dictionary pass throws with both serializer names if two registrations claim the same message type.
• The walk is cheap. Local marked implementors come from the cached messages, not a symbol walk. The referenced-assembly walk is cached per metadata reference, so it runs once per assembly per IDE session. A test proves an edit to the user's own source does not re-walk an unchanged reference, and that a replaced reference does.
• Docs and design record. Guide sections on marked union bases, placement, and the startup check. Decisions 19 and 21 gain addenda, including one disagreement: the decision page's fix list for AKKASG043 mentions declaring a serializer part, which does not exist yet, so the message leaves it out.

What it took

Three rounds. The first cut allocated 18 percent more per keystroke by walking the compilation for types the extraction step had already produced as cached values. The cache was then keyed on the assembly symbol, which Roslyn reuses only while the old symbol is still alive; under memory pressure the proof test failed. It is now keyed on the metadata reference. The proof test then failed in the full suite because this project runs tests at the same time despite its runner file, so a global counter picked up other tests' walks. The test now counts walks by the assembly names it compiles. The runner setting is untouched and worth its own look.

Known limit, scheduled as S7

The ManifestPrefix expansion from F2 still finds its closed set with a walk inside the serializer's per-keystroke step. It is correct across assemblies now, proven by a golden case, but it does not consume the cached facts. S7 moves it.

Cost

Base is the F2 tip. The corpus references Akka.Remote, which references V2, so the referenced-assembly walk is in play on every row. The cache is what keeps the rows flat. Allocations measured twice on the same machine.

Row	Base (F2)	This PR
Fresh driver, full corpus	22.9 ms, 19.00 MB	23.8 ms, 19.07 MB
Warm driver, one field renamed	16.9 ms, 9.57 MB	18.2 ms, 9.64 MB
Warm driver, comment edit	9.4 ms, 3.55 MB	9.7 ms, 3.61 MB
Warm driver, unrelated file edited	7.7 ms, 3.53 MB	8.0 ms, 3.59 MB

How it was checked

369 tests pass (350 plus 19: placement diagnostics, marked union bases, the startup check, three golden cases, a caching scenario, and the cache-proof test), three runs in a row. V2, the generator, and Akka.Remote build with warnings as errors. Akka.API.Tests: 18 pass. cspell passes on the guide. Local runs used NuGetAudit=false on the command line during the restore break that #8535 fixed; nothing of that is committed.