Limonata resident-stage Windows Vulkan pilot — experimental hardware-test build.

The stage, patched llama.cpp, ggml and Vulkan backend were compiled together
with one pinned LLVM-MinGW toolchain and static C++ runtime. This package does
not mix the stock MSVC llama.dll with another compiler's internal C++ objects.
Windows execution and physical AMD/NVIDIA Vulkan GPU execution are NOT tested.
The Vulkan loader vulkan-1.dll must be supplied by the installed GPU driver.
No driver installation, credentials, models or production provider changes are
included. Do not replace your installed Limonata provider/runtime with this file.

This is a framed-stdio engine, not the public contribution app. It requires a
compatible pinned Qwen3 shard/full model and the experimental coordinator worker.
Run '--model MODEL.gguf --gpu-layers 999' through that worker; actual transformer
weights and final normalization/head must be on GPU or startup rejects fallback.
The host shader compiler supports KHR cooperative matrices, but newer NV
cooperative-matrix2, integer-dot, bfloat16 and FP8 shader optimizations are disabled
by upstream feature detection. Generic GPU kernels remain compiled. Performance
and compatibility must be checked on actual Windows hardware.

Third-party notices are in licenses/: LLVM/libc++/compiler-rt Apache2 with LLVM
exceptions; MinGW-w64 runtime notices; Vulkan/SPIRV header licenses; nlohmann JSON
MIT. llama.cpp/ggml and the embedded shaders use LLAMA-LICENSE.txt.
Consult the prototype repository README and build-provenance.json before testing.
