L0 Test (PR CI reproduction)

Runs the same checks that gate a pull request in .github/workflows/cicd-main.yml and code-linting.yml:

1. Lint (host) — pre-commit run --all-files (ruff check, ruff format, mypy, copyright/EOF/whitespace, no-underscore-md).
2. L0 CPU tests (container) — distributed muon utils at nproc=4,8, plus test_scalar_optimizers.py and test_procrustes_step.py.
3. L0 GPU tests (container) — every tests/test_*.py (excluding *_cpu.py) twice — once with a random seed, once with --seed=42 — then the convergence tests/convergence/*_test.py set.

Stages 2 and 3 share a single NGC container session: lint runs on the host, then a single docker run against the image declared in docker/Dockerfile.ci (the FROM line — do not hard-code a tag, so the skill picks up version bumps automatically) with --gpus all performs both CPU and GPU passes. All test execution happens inside the container; never run tests on the host. Reasons:

• tests/test_distributed_muon_utils_cpu.py imports numpy, which is preinstalled in the NGC image but not pulled in by uv sync against a host venv.
• The NGC torch build (2.11.0a0+...nv26.02) and CUDA stack are what CI uses — host torch may differ in version, build flags, or CUDA major.
• torchrun resolution differs between host and container; we use python -m torch.distributed.run everywhere so it's deterministic.

Not covered: the copyright-check external workflow (uses NVIDIA-NeMo/FW-CI-templates) and L1 long-running convergence runs.

Args

Default (no args): run all three stages in order, stopping on first failure.