perf: make the incremental pipeline actually incremental

[kamronbatman]

Summary

The pipeline previously combined the full Compilation into every output node and cached on symbol-bearing records that never compare equal — so every keystroke re-ran everything: full codegen for every class and a re-parse of every migration JSON (#47's baseline measured the warm re-run as more expensive than a cold run). This PR restructures the generator around a fully resolved, value-equatable SerializationModel so the expensive stages cache for real.

Design

• Transform (ForAttributeWithMetadataName → BuildSerializationModel): all symbol/compilation-dependent work — attribute processing, the rules engine, every validation diagnostic, MarkDirty/timer/after-deserialization resolution, nesting shells — resolves into an equatable model of strings, enums, and EquatableArrays.
• Generation is symbol-free: it consumes only the model, so the source-output stage caches on model equality. Editing anything that doesn't change a class's serialization surface — method bodies, comments, unrelated members — produces an equal model and regenerates nothing.
• Migration files parse once per content change in their own tracked node; a cheap compilation-combined augment resolves the one value-type fact content structs need; an equality-aware map attaches each class's migrations (SG3011/SG3012/SG3013 ride along as equatable DiagnosticInfo).
• Location is deliberately excluded from model equality so same-file edits don't defeat caching; a cached warning's position can lag until the surface changes.

Proof

• Incrementality tests (step tracking, red on main → green): editing one class leaves the rest cached; editing an unrelated file, or non-serialization code inside a serializable class, regenerates nothing; code edits never re-parse migration JSONs; a migration edit re-runs only its class. A 150-class probe pins it at scale: after an inert edit, SourceOutput: executed=0, cached=150.
• Output-identical: all 17 snapshots byte-identical; DiffTool manifest over the 3,786-source real corpus byte-identical to main.
• Benchmarks (150 classes with migrations; baseline re-measured on main after fixing the benchmark's missing analyzer-private Humanizer dependency — the earlier baseline was unknowingly measuring crash paths):

Scenario	main	this PR
Single-edit warm re-run	61.2 ms / 16.4 MB / 150 outputs re-run	47.9 ms / 10.4 MB / ≤1 output re-run
No-change re-run	~69 µs	~63 µs
Cold full run	20.7 ms / 11.3 MB	29.9 ms / 10.1 MB

The per-edit wall-time and allocation wins understate the IDE impact: the workspace no longer re-adds and re-parses every generated document on every keystroke. The cold-run regression is the transform fronting resolution work — a deliberate trade against the continuous per-edit cost.

Breaking (rule API, for the major bump)

GenerateDeserializationMethod / GenerateMigrationProperty no longer take a Compilation; IPostDeserializeMethod receives the model; timer-method resolution and its SG3008 check moved to model building.

🤖 Generated with Claude Code