[pmeems]

Phase 3 of the Map Model and Basic Navigation plan: multi-layer rendering.

Stacked on #35 (Phase 2, viewport), which is stacked on #34 (Phase 1, map model). Merge those first.

Motivation

Before this change the control kept one dataset, one render scene, and its own extent-fitting ViewTransform. Two layers could not be shown together, and the view state lived in ActiveX instead of in Core.

What changes

• mwg::MapView (Core). It owns the Map, the one Viewport that all layers are drawn through, and the render scene. It is now the only map and view state: the control has no extent or transform of its own.
• rendering::CreateMapRenderScene (Core). It builds an owned snapshot of all visible layers, bottom layer first. Each layer is converted separately. Hidden layers, empty layers, and layers that fail (for example, non-polygon layers) are listed in skippedLayers and do not block the others.
• Backends take the scene plus a Viewport. ViewTransform is now just a per-frame copy of the viewport, using the same arithmetic as Viewport::MapToScreen. GDI checks each layer before drawing it, skips a layer it cannot draw, and reports it through GetFailedLayers and RenderingStatus. OpenGL still renders the whole frame or falls back to GDI.
• Redraws. Map changes go through MapView, which marks the scene stale. Each mutator returns whether the frame changed, and the control calls FireViewChange only when it did. A resize keeps the center and resolution and does not rebuild the scene.
• Resizing. The viewport follows WM_SIZE, and also the drawn bounds, because windowless hosts never send WM_SIZE. Sizes are clamped to Viewport::kMaxDimension. A fit requested before the control has a size, or while it is minimized, happens on the first resize that gives it a size.
• COM behavior change. LoadVectorDataset now adds the dataset's first layer on top of the map instead of replacing what is loaded. FeatureCount is the total over all layers. Until navigation is exposed (Phase 5), every load fits the view to the whole map. Non-polygon layers are still rejected at load time.

Design decisions

• MapView lives in Core rather than in the control, so the Phase 4 navigation code can build on one tested owner of the map and viewport (plan decisions 1 and 4).
• The render scene holds no view state, so view changes never re-read data.
• One backend draws the whole frame, so every layer gets the same transform and the same pixel conventions.

Testing

• Debug|x64 builds with 0 warnings and 0 errors.
• 255 of 258 native tests pass; the other 3 are the usual opt-in skips. New tests: MapRenderSceneTests, MapViewTests, and multi-layer GDI/OpenGL pixel tests.
• The opt-in rendering baseline passes.
• The registration-free C# host passes all 14 checks.

Remaining limitations are listed in the plan under Phase 3. The main ones: there are no COM methods to remove or hide layers yet, the scene is rebuilt for all layers on any map change, and there is no clipping, so GDI skips a layer whose device coordinates go out of range at deep zoom.