perf(pipeline): shared, disk-persisted VkPipelineCache (#69)

vkCreateComputePipelines and vkCreateRayTracingPipelinesKHR were passed
VK_NULL_HANDLE, so the driver couldn't reuse compiled shader binaries
across the pipelines built at startup (4 UI shaders + user/RT pipelines)
or across runs.

Add one process-wide Device::pipelineCache. LoadPipelineCache (called
from Initialize) seeds it from pipeline_cache.bin when the file's
VkPipelineCacheHeaderVersionOne header matches this device's
vendorID / deviceID / pipelineCacheUUID; a stale or foreign blob is
discarded so the driver never rejects it. SavePipelineCache is registered
with std::atexit, serialising the cache at process shutdown without any
explicit teardown hook. Both create sites now pass the cache; a null
handle stays valid, so no create-site null check is needed.

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

interfaces/Crafter.Graphics-Device.cppm

@@ -161,6 +161,27 @@ export namespace Crafter {
 
         inline static VkPhysicalDeviceMemoryProperties memoryProperties;
 
+        // Core physical-device properties, captured once at Initialize from the
+        // VkPhysicalDeviceProperties2 query. Kept because the pipeline-cache
+        // persistence path needs vendorID / deviceID / pipelineCacheUUID to
+        // decide whether an on-disk blob was written by this exact device.
+        inline static VkPhysicalDeviceProperties deviceProperties = {};
+
+        // One process-wide pipeline cache fed to every vkCreate*Pipelines call
+        // (compute UI/user shaders + RT pipelines). Pipeline compilation is a
+        // one-time cold-start cost, not a per-frame one; a single shared cache
+        // lets the driver reuse compiled shader binaries across the several
+        // pipelines built at startup and — once persisted to disk (see
+        // LoadPipelineCache / SavePipelineCache) — across runs. VK_NULL_HANDLE
+        // until LoadPipelineCache runs; passing VK_NULL_HANDLE to a create call
+        // is valid and simply means "no cache", so the create sites need no
+        // null check.
+        inline static VkPipelineCache pipelineCache = VK_NULL_HANDLE;
+        // Where the serialized cache lives. Relative to the working directory,
+        // matching the gpu_crash_dump-* convention. Set before Initialize to
+        // relocate it.
+        inline static std::filesystem::path pipelineCachePath = "pipeline_cache.bin";
+
         inline static VkPhysicalDeviceDescriptorHeapPropertiesEXT descriptorHeapProperties = {
             .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_HEAP_PROPERTIES_EXT
         };
@@ -199,6 +220,23 @@ export namespace Crafter {
         // ComputeShader read the offset off the pipeline they record.
 
         static void CheckVkResult(VkResult result);
+
+        // ─── Pipeline cache persistence (issue #69) ─────────────────────
+        // LoadPipelineCache creates `pipelineCache`, seeding it from
+        // pipelineCachePath when the file exists *and* its header matches this
+        // device (PipelineCacheDataCompatible) — a stale or foreign blob is
+        // discarded so the driver never rejects it. Called once from
+        // Initialize. SavePipelineCache writes the driver's current cache blob
+        // back out; Initialize registers it with std::atexit so the cache is
+        // serialized at process shutdown without any explicit teardown call.
+        static void LoadPipelineCache();
+        static void SavePipelineCache();
+        // True when `data` is a VkPipelineCache blob whose header was written by
+        // a device with the same vendorID / deviceID / pipelineCacheUUID as
+        // `deviceProperties`. Pure logic over the standard 32-byte cache header,
+        // so it is driven directly by the PipelineCacheValidation test with no
+        // GPU device. A blob too short to hold the header is incompatible.
+        static bool PipelineCacheDataCompatible(std::span<const std::byte> data);
         // Selects a memory type index from typeBits that satisfies `required`.
         // When `preferred` bits are also given, a type satisfying both is
         // chosen first; if none exists we fall back to required-only rather