[juliendrapeau]

Summary

• Fix two silent correctness bugs: global RNG state mutation in sweep config generation, and NaN
  values from empty CSV cells silently corrupting Snakemake rule params
• Refactor build_sif.py: fast local-apptainer path, SSH safety check, BuildKit cache, remove
  tqdm and third-party importlib_metadata
• Extend job_utils.py with full SLURM metadata capture and automatic container provenance

Resolves #7.

Changes

Fix global RNG state in sweeper.py

Config.default_metadata was calling random.seed(seed) and np.random.seed(seed) as a side
effect of generating a config. This mutated global state on every Config instantiation, causing
non-obvious test interference and making independent RNG streams impossible. Replaced with
np.random.default_rng().integers(2**32) for seed generation (no global state). Callers needing
reproducibility should use np.random.default_rng(metadata["seed"]) locally.

NaN sanitization in Snakefile

pd.read_csv represents empty CSV cells as NaN (float). Without sanitization, any optional
parameter left blank in the sweep CSV arrives at the Snakemake rule as a truthy float and silently
passes incorrect values downstream. Added _clean() helper that maps NaN → None when building
params_by_path.

build_sif.py refactored

• Fast path: when apptainer/singularity is installed locally and building for the host
  platform, uses docker buildx build --load + apptainer build docker-daemon:// instead of
  writing an intermediate OCI archive — substantially faster for native builds.
• SSH safety: _ssh_flag() only passes --ssh default when ssh-add -l confirms keys are
  loaded; previously it was always passed, causing failures on machines without an SSH agent.
• BuildKit cache: OCI fallback path now uses --cache-from/--cache-to to avoid rebuilding
  unchanged layers.
• Stdlib metadata: replaced third-party importlib_metadata with stdlib importlib.metadata;
  removed importlib-metadata and tqdm from dev dependencies.

SLURM metadata and container provenance in job_utils.py

capture_job_metadata() now records SLURM env vars (SLURM_JOB_ID, SLURM_JOB_PARTITION,
SLURM_CPUS_PER_TASK, SLURM_MEM_PER_CPU, SLURM_NODELIST, SLURM_TIMELIMIT,
SLURM_ARRAY_TASK_ID) — previously none of this was captured, making it impossible to correlate
results with cluster allocation details after the fact. All fields are None on local runs.

capture_run_metadata() now reads the container path from APPTAINER_CONTAINER/SINGULARITY_CONTAINER
(set automatically when running inside Apptainer) rather than the manually-set CONTAINER_DIGEST
env var, and computes a SHA-256 digest of the SIF file for exact reproducibility.

Version

Bumped to 0.4.0.

[github-actions]

☂️ Python Coverage

current status: ✅

Overall Coverage

Lines	Covered	Coverage	Threshold	Status
59	55	93%	90%	🟢

New Files

No new covered files...

Modified Files

No covered modified files...

updated for commit: 5e4f574 by action🐍