Split `gwdkir.Program` into semantic IR and target-specific plans

[cssbruno]

Summary

internal/gwdkir.Program is described as the stable single compiler handoff, but it currently combines several different abstraction levels:

• parsed/source-oriented page, component, layout, block, and raw-body records;
• normalized routes and endpoints;
• standalone raw Go endpoint declarations;
• resolved backend bindings;
• templates and parsed view nodes;
• contracts, realtime subscriptions, invalidations, and audit specs;
• client behavior and asset records;
• accumulated diagnostics.

Generators then build additional parallel projections, notably BackendAdapterIR, by copying endpoint metadata into registrations, action/API/fragment adapters, decoders, calls, responses, and fallbacks.

Architectural warning signs

• GoEndpoint exists because normalized Endpoint is explicitly lossy and validation still needs the raw kind, method, and source spans.
• Several records preserve both raw source bodies and parsed representations.
• Endpoint bindings are rejoined in app generation using composite string keys rather than stable semantic identities.
• New generated-output work is encouraged to add fields to the global Program, increasing coupling among unrelated compiler consumers.
• A change for one target can expand the shared schema used by every other target.

Proposed model

Move toward explicit compiler stages:

Lossless AST/CST
    -> SemanticProgram / HIR
    -> StaticSitePlan | ServerPlan | ClientPlan | AuditProjection
    -> target generation

The semantic layer should use stable typed identities, for example:

type PageID string
type EndpointID string
type ComponentID string
type SourceID string

Source fidelity should remain in the lossless tree or a source-map service. Target plans should reference semantic IDs instead of repeatedly copying complete endpoint registrations.

This can be migrated incrementally; it does not require replacing every consumer at once.

Suggested phases

1. Introduce typed stable IDs and assign them during program assembly.
2. Replace string-concatenated endpoint joins with EndpointID lookups.
3. Define a lossless source-reference/source-map boundary so post-lowering validation does not require duplicate raw endpoint models.
4. Extract ServerPlan from appgen's current backend adapter projection.
5. Extract corresponding static/client/audit projections as those consumers change.
6. Narrow gwdkir.Program to semantic facts shared by multiple targets.

Acceptance criteria

• Pages, components, routes, and endpoints have stable typed identities across compiler stages.
• App generation no longer joins bindings through concatenated string keys.
• Exact source spelling and spans needed for diagnostics are obtained from a lossless source model or source map, not a second raw semantic entity.
• Target-specific generation decisions live in explicit target plans rather than the global shared IR.
• Adding a field needed by only one generator does not require adding it to the shared semantic program.
• Golden tests continue to pin AST, semantic IR, target plans, and emitted artifacts at clear stage boundaries.
• The architecture documentation defines invariants for each stage and which package owns validation before and after lowering.