perf(rt): allocate TLAS metadata buffer in BAR/VRAM, not system RAM (#75)

The TLAS metadata buffer is CPU-written every frame (one userMetadata word
per element in BuildTLAS's copy loop) and read by the ray shaders as a
STORAGE_BUFFER via device address every frame. Allocated
HOST_VISIBLE | HOST_COHERENT (system RAM), every per-frame shader read
traversed PCIe — the same cost the sibling instance buffer paid before #65.

Upgrade the request to HOST_VISIBLE with a best-effort DEVICE_LOCAL preference
(BAR/VRAM), keeping the single persistently-mapped buffer. The CPU's writes are
write-only, sequential, never-read-back — the ideal write-combined BAR load —
and the shader reads now hit local VRAM.

Unlike the instance buffer there is no GPU co-write, so the direct upgrade is
unconditional. Staging is still deliberately avoided: this buffer is rewritten
by the CPU every frame, so a per-frame stage+copy+barrier would defeat the
purpose (#75 caveat 2).

Correctness:
- DEVICE_LOCAL is a preference only (depends on #59): GetMemoryType falls back
  to plain HOST_VISIBLE (the previous behaviour) on GPUs without resizable BAR.
- HOST_COHERENT is dropped from the required set — the combined type may not be
  coherent. A FlushDevice() is added after the copy loop to make the host writes
  available on a non-coherent type; it self-gates to a no-op on a coherent type
  (#60), so the previous behaviour is preserved wherever the type ends up
  coherent. The metadata buffer is not an AS build input and has no GPU
  co-write, so it needs neither the build-stage barrier the instance buffer
  takes nor a HOST->shader command barrier: the queue submit already orders host
  writes made available before it against every command in the submission,
  exactly as it did when this buffer was HOST_COHERENT.

Extends the TLASHighWaterMark hardware test with a metadata-buffer assertion
block mirroring #65's instance-buffer check: the chosen memory type must equal
what GetMemoryType resolves for the HOST_VISIBLE | prefer-DEVICE_LOCAL request,
and must be DEVICE_LOCAL wherever a host-visible device-local type is reachable.
The zero-validation-errors check covers the dropped HOST_COHERENT path.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

implementations/Crafter.Graphics-RenderingElement3D.cpp

@@ -121,10 +121,23 @@ void RenderingElement3D::BuildTLAS(VkCommandBuffer cmd, std::uint32_t index, RTB
         // visibility, gating its flush-skip on the *chosen* type's flags. The
         // single shared buffer is preserved: a whole-buffer copy would clobber
         // the GPU-written transforms, so staging is deliberately avoided.
+        //
+        // metadataBuffer (#75) gets the same HOST_VISIBLE | prefer-DEVICE_LOCAL
+        // upgrade: it is CPU-written every frame (the copy loop below) and read
+        // by the ray shaders as a STORAGE_BUFFER every frame, so it has the same
+        // per-frame-shader-read-over-PCIe cost. Unlike instanceBuffer there is
+        // no GPU co-write, so the direct upgrade is unconditional — but staging
+        // is still wrong here: a per-frame-rewritten buffer would pay a stage+
+        // copy+barrier every frame, defeating the point. HOST_COHERENT is
+        // likewise dropped; the FlushDevice() after the copy loop makes the host
+        // writes available before the consuming submit on a non-coherent type
+        // (no-op when the chosen type is coherent — the previous behaviour),
+        // and the queue submit carries the host-write -> shader-read ordering as
+        // it always has.
         if (tlas.instanceBuffer.buffer == VK_NULL_HANDLE
             || primitiveCount > tlas.instanceBuffer.size / sizeof(VkAccelerationStructureInstanceKHR)) {
             tlas.instanceBuffer.Resize(VK_BUFFER_USAGE_SHADER_DEVICE_ADDRESS_BIT | VK_BUFFER_USAGE_TRANSFER_DST_BIT | VK_BUFFER_USAGE_ACCELERATION_STRUCTURE_BUILD_INPUT_READ_ONLY_BIT_KHR | VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT, primitiveCount, VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT);
-            tlas.metadataBuffer.Resize(VK_BUFFER_USAGE_STORAGE_BUFFER_BIT | VK_BUFFER_USAGE_SHADER_DEVICE_ADDRESS_BIT, VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT | VK_MEMORY_PROPERTY_HOST_COHERENT_BIT, primitiveCount);
+            tlas.metadataBuffer.Resize(VK_BUFFER_USAGE_STORAGE_BUFFER_BIT | VK_BUFFER_USAGE_SHADER_DEVICE_ADDRESS_BIT, VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT, primitiveCount, VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT);
         }
     }
 
@@ -147,6 +160,18 @@ void RenderingElement3D::BuildTLAS(VkCommandBuffer cmd, std::uint32_t index, RTB
 
     tlas.instanceBuffer.FlushDevice(cmd, VK_ACCESS_MEMORY_READ_BIT, VK_PIPELINE_STAGE_ACCELERATION_STRUCTURE_BUILD_BIT_KHR);
 
+    // Make the per-frame metadata host writes available before the consuming
+    // submit. metadataBuffer is not an AS build input (the build never reads it)
+    // and has no GPU co-write — only the ray shaders read it, in a later pass —
+    // so it needs neither the build-stage barrier instanceBuffer takes above nor
+    // a HOST->shader command barrier: the queue submit already orders host
+    // writes made available before it against every command in the submission,
+    // exactly as it did when this buffer was HOST_COHERENT. This plain host
+    // flush is what now makes those writes available on a non-coherent
+    // (BAR/VRAM) type; it self-gates to a no-op on a coherent type (#60),
+    // preserving the previous behaviour where the type ends up coherent.
+    tlas.metadataBuffer.FlushDevice();
+
     VkAccelerationStructureGeometryInstancesDataKHR instancesData {
         .sType = VK_STRUCTURE_TYPE_ACCELERATION_STRUCTURE_GEOMETRY_INSTANCES_DATA_KHR,
         .arrayOfPointers = VK_FALSE,