Test suite on the GPU development server, 2026-08-09
====================================================

Host:   osiris (Windows desktop DESKTOP-4LO089U, WSL2 "Pengwin"), NVIDIA GeForce RTX 3090
Build:  conda env evm-cuda (nvcc 12.9, cmake 4.4.2, ninja, g++), built with
        -DCMAKE_CUDA_ARCHITECTURES=86  (the explicit flag is REQUIRED; see
        docs/dev/packaging-notes.md for the CMakeLists defect that makes it so)
Command: PYTHONPATH=<repo>/cuda python -m pytest tests/ -q -p no:randomly

RUN 1 - 102 passed in 161.77s
    NOT a clean pre-change baseline. The working copy was synced to this machine while
    Phase 0 agents were still writing files, so it already contained
    tests/test_reference_lock.py (7 cases) but not yet tests/test_golden.py.
    Composition: 95 original cases + 7 = 102. Recorded here for honesty, not as a baseline.

RUN 2 - 111 passed, 0 skipped in 169.43s
    The real post-Phase-0 state: 95 original + 7 reference-lock + 9 golden = 111.
    Zero skips, because the compiled CUDA module is importable on this host.

For comparison, the same commit on macOS (no CUDA toolchain):
    48 passed, 63 skipped - the 63 skips are the entire tests/cuda/ suite.
    See benches/baseline_tests_2026-08-09_macos.txt.

The original 95-case suite (before Phase 0 added any test) was never captured on this
machine with zero skips; the macOS file is the authoritative record of the pre-change
count, at 32 passed + 63 skipped.
