build: pin MatrixOne embedding to Pixi

[aunjgr]

Description

Pin the MatrixOne embedding build to Sirius's mo Pixi profile: CUDA 13.3, GCC 14, cuDF/cuVS 26.08, RMM, and packaging tools come from one locked environment. Ordinary Sirius CUDA 13.2 and CUDA 12 profiles remain unchanged. Pixi is the sole GPU dependency provider for the combined MO/Sirius build; the SDK no longer exports a second GPU toolchain JSON manifest.

The existing embedding SDK link.json continues to record the verified C consumer, compiler, link arguments, native artifacts, and source provenance. The mo task disables the CUDA sccache launcher that loses NVCC temporary PTX files; ordinary presets retain their launchers. Driver libraries and linker stubs remain host/link-only inputs, not runtime implementations.

Validation on the updated branch: pixi run --frozen -e mo mo-build-embedding-sdk, eight SDK exporter tests, and changed-file pre-commit hooks passed. The rebuilt SDK contains no GPU toolchain manifest. MatrixOne's Pixi-only branch completed a full GPU native and binary build plus a linked CGo test under this same environment. The combined bridge packaging and two-stream coexistence gate must be rerun after draft MO #29203 is updated; prior manifest-based results are not claimed for this head.

Checklist

• Read CONTRIBUTING.md and ensure PR meets the reviewability checklist
• Cover changes with new or existing tests
• Document configuration changes and summarize them above
• Update the embedding GPU toolchain documentation

References

Refs matrixorigin/matrixone#28966 and its Pixi-only decision.

[aunjgr]

Three Sirius #22 build checks remain queued on self-hosted runners. The CUDA 13/default job has waited since 2026-09-23 10:48 UTC with labels self-hosted, ubuntu-24.04, cpu-xl and no runner assigned; the two x64 build jobs are likewise queued. This is not a build/test failure. The frozen mo Pixi SDK build, eight exporter tests, and changed-file hooks passed locally at this PR head.

Could a maintainer restore the matching runner capacity or clarify whether these queued jobs are required for merging? I have not changed Sirius code or CI to work around runner availability.