windows: fix test linking and missing include

src/visualizer/rendering/lod_page_cache.hpp

@@ -17,7 +17,7 @@
 
 namespace lfs::vis {
 
-    class LodPageCache {
+    class LFS_VIS_API LodPageCache {
     public:
         static constexpr std::uint32_t kInvalidPage = lfs::core::SplatLodTree::kInvalidPage;
         static constexpr std::size_t kChunkSplats = lfs::core::SplatLodTree::kChunkSplats;

src/visualizer/rendering/lod_page_dequant_cuda.hpp

@@ -4,6 +4,7 @@
 
 #pragma once
 
+#include "core/export.hpp"
 #include "io/formats/rad_packed_page.hpp"
 #include "lod_pool_quant.hpp"
 
@@ -36,12 +37,12 @@ namespace lfs::vis {
     // pass through bit-exact; everything else lands as f16 — no new loss over
     // the file. Launched on the upload engine's stream between the slot copy
     // and the timeline signal.
-    cudaError_t launchLodPageDequant(const std::uint8_t* device_slot,
-                                     const lfs::io::RadPagePackedDesc& desc,
-                                     const LodPoolDeviceView& pool,
-                                     std::uint32_t page,
-                                     std::uint32_t page_splats,
-                                     cudaStream_t stream);
+    LFS_VIS_API cudaError_t launchLodPageDequant(const std::uint8_t* device_slot,
+                                                 const lfs::io::RadPagePackedDesc& desc,
+                                                 const LodPoolDeviceView& pool,
+                                                 std::uint32_t page,
+                                                 std::uint32_t page_splats,
+                                                 cudaStream_t stream);
 
     // fp32 source tensors for the in-core / pinned-root D2D fill path.
     // Pointers are pre-offset to the page's first splat; layouts match the
@@ -63,10 +64,10 @@ namespace lfs::vis {
     // Quantizes resident fp32 tensors into one canonical pool page (two
     // passes: per-band |max| reduction into the page frame, then the
     // quantizing scatter). Same canonical writers as the streamed path.
-    cudaError_t launchLodPageQuantizeFromTensors(const LodPageTensorSources& src,
-                                                 const LodPoolDeviceView& pool,
-                                                 std::uint32_t page,
-                                                 std::uint32_t page_splats,
-                                                 cudaStream_t stream);
+    LFS_VIS_API cudaError_t launchLodPageQuantizeFromTensors(const LodPageTensorSources& src,
+                                                             const LodPoolDeviceView& pool,
+                                                             std::uint32_t page,
+                                                             std::uint32_t page_splats,
+                                                             cudaStream_t stream);
 
 } // namespace lfs::vis

src/visualizer/rendering/lod_pool_quant.hpp

@@ -12,6 +12,7 @@
 
 #pragma once
 
+#include <cstddef>
 #include <cstdint>
 
 namespace lfs::vis::lodq {

src/visualizer/rendering/lod_upload_engine.hpp

@@ -4,6 +4,7 @@
 
 #pragma once
 
+#include "core/export.hpp"
 #include "io/formats/rad_packed_page.hpp"
 #include "lod_page_cache.hpp"
 #include "rendering/cuda_vulkan_interop.hpp"
@@ -27,7 +28,7 @@ namespace lfs::vis {
     // stream. The render thread polls collectPublished() and publishes
     // residency only for pages whose kernel completed, so the selector never
     // observes a partially written page. Staging slots are the backpressure.
-    class LodUploadEngine {
+    class LFS_VIS_API LodUploadEngine {
     public:
         struct DeviceLayout {
             // Payload pool (page-input regions, InputRegion order): xyz_ws,