build: pin MatrixOne embedding to Pixi

docs/embedding-gpu-toolchain.md

@@ -0,0 +1,31 @@
+# Pixi provider for MatrixOne embedding
+
+Refs [MatrixOne #28966](https://github.com/matrixorigin/matrixone/issues/28966).
+
+The `mo` Pixi environment is the sole CUDA/RAPIDS build provider for the
+combined MatrixOne/Sirius binary. It locks CUDA 13.3, GCC 14, cuVS/cuDF
+26.08.01 and RMM. Sirius's ordinary CUDA 13.2 and `cuda12` profiles are
+unchanged; a compatible NVIDIA driver and GPU remain host requirements.
+
+```sh
+pixi run --frozen -e mo mo-build-embedding-sdk
+```
+
+The profile-specific configure task disables `sccache` only for CUDA. The
+ordinary presets retain their launchers; the `mo` SDK avoids the observed NVCC
+temporary-PTX failure with CUDA 13.3 and sccache 0.15.
+
+The existing `embedding-sdk/link.json` records the exact C smoke compiler,
+link arguments, native artifacts, and source provenance. It does not export a
+second GPU toolchain manifest. Build MO's GPU native libraries under the same
+activated `mo` Pixi environment. The MatrixOne bridge verifies its native
+generation and that the SDK compiler belongs to this prefix, then hashes and
+packages the actual ELF dependency closure of both the SDK and `libmo`.
+Linker stubs and NVIDIA driver libraries are never packaged as runtime
+implementations. The deployed binary uses relative paths to staged shared
+libraries and still requires the host driver.
+
+The embedding SDK is build-tree metadata, not a relocatable artifact. Rebuild
+after changing the Pixi lock or environment. Compilation and packaging do not
+establish GPU execution: the combined MO cuVS/Sirius test with two streams is
+the runtime gate.