[hallelx2]

The rebuild completed ten steps and then died one from the end:

usage: pipeline.py [-h] [--dataset DATASET] [--holdout HOLDOUT]
                   [--output-dir OUTPUT_DIR] [--no-umap]
pipeline.py: error: unrecognized arguments: --from dataset_assembly --force --no-zenml

_run_module calls each step's entrypoint in-process. Those are written as main(argv=None) and fall through to parse_args(argv) — and with argv=None, argparse reads sys.argv[1:], which belongs to the pipeline, not the step. So clustering was handed --from dataset_assembly --force --no-zenml, didn't recognise them, and SystemExit-ed. figures never ran.

Why four hours of work hit this and nothing before did

The systemd unit invoked the pipeline with no arguments:

ExecStart=/usr/local/bin/uv run --with mlflow python -m cancerag.pipeline

sys.argv[1:] was empty, argparse was satisfied, and every historical run took the one path that works. The first run with --from/--force hit it immediately. A latent bug in a step that only fires when you use the CLI as designed.

Change

The guard goes in _run_module, not the call sites. Three modules parse args this way — unsupervised_clustering, chain_verification, binding_site — and patching three places leaves the next one written exposed. The step is invoked as a function; it should not see a command line at all.

sys.argv is restored in a finally, so a failing step cannot leave it clobbered for the steps after it.

_script and _module_cli use subprocess.run([sys.executable, "-m", module]), which gets a fresh argv — unaffected.

Verification

• Tests checked against the unfixed code. Reverting the guard reproduces the production error verbatim:

  FAILED test_step_with_argparse_survives_pipeline_flags - SystemExit: 2
  usage: pipeline.py ... error: unrecognized arguments: --from dataset_assembly --force --no-zenml
  

• pytest tests/unit — 481 passed, 31 skipped (was 474; +7 here)
• ruff check clean on both touched files
• Covers: argparse step survives pipeline flags; argv restored normally; argv restored when the step raises; real **kwargs still reach the entrypoint; and a guard test that flags if any of the three exposed modules stops parsing args, so the issue's exposure list can't silently rot

Next: re-run --from clustering to finish clustering and figures.

Closes HAL-4194