Cache one resolved model per serializer so one message edit regenerates one file (S3)

[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 S3.

What changes

Renaming a field in a message now regenerates one file, the serializer that owns that message. Before, it regenerated every serializer's file, and the unchanged ones matched byte for byte only by luck. Generated text is byte-identical. No golden, wire, or model snapshot changed. The diagnostic set is unchanged.

• One cached model per serializer. ResolvedSerializer holds everything emission needs: the gate result, the messages this serializer can reach, the top-level set as a ClosedSet, the closed-generic registrations after resolution, formatter overrides, a typed UnionPlan in place of a string-keyed dictionary, the derived method names, and the validation diagnostics. It compares by content.
• Emission reads only that model. No compilation, no symbols, no other serializer's data. Diagnostics that span serializers (AKKASG013, 031, 037) get their own small output and report once. The protocol-coverage output reads the gate result from the model instead of recomputing it.
• One ClosedSet type for the top-level dispatch set and the union sets. Decision 18 later expands over this object.

Why

Roslyn reuses a step's output when its inputs compare equal. A model that holds only what one serializer needs stays equal when another serializer's message changes. A first version stored the whole message table in every serializer's model, and a probe test showed every model changing on every edit.

Savings

Short run, 5 serializers x 100 messages, machine at moderate load.

Scenario	S2	S3
Fresh driver, full corpus	75.3 ms, 23.19 MB	89.8 ms, 22.61 MB
One field renamed in one message	75.8 ms, 23.17 MB	69.4 ms, 13.15 MB
Comment edit inside a message	60.2 ms, 7.82 MB	59.1 ms, 7.79 MB
Unrelated file edited	59.0 ms, 7.77 MB	57.1 ms, 7.76 MB

The rename row drops from the full-corpus figure to the no-change floor plus one serializer's emission. Wall clock carries noise at this load.

How it was checked

Scenario b in the caching tests flips: the rename runs one serializer's emit step, and the other never runs. 293 tests pass (287 plus 6): resolve tests on models with no driver, and a Verify snapshot of each serializer's resolved model over the golden corpus. Generator builds with warnings as errors. Akka.Remote builds.

Known follow-up

The protocol-coverage scan (AKKASG029) now runs once per serializer instead of once per compilation. Same output, more scanning when a project has several serializers. S5 moves it into one cached step.