[spelech]

Summary

• Disable Server GC (ServerGarbageCollection=false) and enable DATAS (GarbageCollectionAdaptationMode=1) in LocalLLMServerManager.csproj so freed heap segments are decommitted back to the OS on desktop machines.
• Compact the Large Object Heap (GCLargeObjectHeapCompactionMode.CompactOnce) and trigger an aggressive blocking GC in MemoryLifecycleWatchdog when process working set exceeds 512 MB, managed heap exceeds 256 MB, or system RAM usage reaches 75%.
• Trigger LOH compaction and aggressive GC in POST /api/memory/purge before reading post-purge memory telemetry.
• Add a 3-second TTL cache in GpuTelemetryProvider (GetTelemetryAsync and GetGpuInfo) to avoid spawning
  vidia-smi thousands of times per day across concurrent polling, watchdog ticks, and health checks, while ensuring process handles are safely handled.
• Add unit tests in HardwareEndpointsTests covering /api/memory/purge and GpuTelemetryProvider.GetTelemetryAsync TTL caching.