[Jammy2211]

Summary

PR 2a of hpc-campaign phase 2 (#1740, epic ledger draft/research/autofit/hpc_campaign_epic.md). It adds a new autofit/campaign/ package (stdlib + PyYAML; nothing calls Slurm at import), its tests, and nothing else.

• campaign.yaml loader and validator (spec.py):
  • stages and manifests (0-based line = array index), chunked to the site's max_array_size;
  • resources set through a single --mem-per-cpu knob;
  • per-index after: dependencies (aftercorr);
  • success: checks;
  • a propose-only retry policy (ruling F5);
  • the site: layer from PyAutoNerves site.yaml.
• Scheduler protocol with three adapters:
  • FakeScheduler, for tests;
  • LocalScheduler, a subprocess pool;
  • SlurmScheduler: batched sacct/squeue/sinfo through an injectable runner, and sbatch with aftercorr, --kill-on-invalid-dep=yes, --mem-per-cpu, and --signal=B:USR1@600 forwarded to the fit.
• Partition lint (hard constraint, lint.py): CPU-only work never goes on a GPU partition. It runs on validate, submit and every pending-job update. The timing-leg exemption needs all three of: ≤8 CPUs per task, a throttle of %2 or less, and no pending GPU jobs. If any of those is unknown, it refuses.
• submit writes an atomic job ledger.
• CLI: python -m autofit.campaign {validate,submit}.

Decisions taken (please check)

1. Mixed lists are always refused. ral,gpu and gpu,ral are refused for CPU work even when the exemption conditions hold; the exemption only applies to a GPU partition on its own. This is stricter than the ledger's wording, and safe.
2. Directives are read the way sbatch reads them: parsing stops at the first command, so a late #SBATCH --gpus cannot make a CPU job count as a GPU job.
3. Array numbering: each chunk's array runs from 0, and the script exports CAMPAIGN_INDEX = CAMPAIGN_OFFSET + SLURM_ARRAY_TASK_ID. That keeps aftercorr aligned chunk by chunk.
4. squeue/sinfo use classic -o format strings rather than -O.

Test plan

• test_autofit/campaign (90 on this branch): spec validation, units, the lint matrix (every branch, ral,gpu and gpu,ral included), sacct parsing, Slurm golden sbatch text with an injected runner, submit, local scheduler, CLI, and a subprocess test that nothing runs at import.
• Full PyAutoFit suite (see comment).

Refs #1740

🤖 Generated with Claude Code

[Jammy2211]

Full PyAutoFit suite (PyAuto313, local): 2a 4,082 passed / 2b 4,098 passed (2 skipped, 1 xfailed). Smoke autofit (scripts + notebooks) green on the 2b head. Heart STALE (release validation only) acknowledged by the human for this ship.