pixi run test fails on main: CUDA 13.2 pin lags 13.3 bindings + cython-test .so misplacement

[rparolin]

Summary

pixi run test (env cu13) does not complete from a clean checkout on main. Two independent root causes:

1. CUDA version pin lag. The bindings were regenerated against CUDA 13.3.0 (Support CUDA 13.3 #2139), adding NVRTC symbols (NVRTC_ERROR_BUSY, nvrtcBundledHeadersInfo, nvrtcGetBundledHeadersInfo), but the pixi cuda-version pins stayed at 13.2.* in cuda_bindings/pixi.toml and cuda_core/pixi.toml. The local build then compiles 13.3-referencing Cython code against a 13.2 nvrtc.h:

   'nvrtcBundledHeadersInfo' was not declared in this scope
   

   CI was unaffected because it builds wheels from ci/versions.yml (13.3.0), not via pixi run test.

2. Cython-test .so placement. tests/cython/build_tests.py runs build_ext --inplace, which writes the .so relative to cwd; pixi runs the task from the project root, so the .so lands in the package root instead of tests/cython/, where pytest imports it by bare module name. Latent since fix(test): unblock root pixi run test workflow #1978 — a stale, correctly-placed .so masked it. A clean checkout fails with ModuleNotFoundError.

Fix

PR #2180 — bumps the cuda-version pins to 13.3.* and chdirs before build_ext --inplace in both build_tests.py.

Notes

There were also two local-only stale-artifact symptoms while bisecting (stale pyclibrary parser caches cache_driver/cache_runtime, and a stale compiled .so) that don't exist on a clean checkout — captured in the follow-up issue on drift prevention.

🤖 Filed with Claude Code