[Jammy2211]

Summary

hpc-campaign phase 4 (#1746). Stacked on #1747 (phase 3); merge that first. Both add commands to campaign/__main__.py, and that conflict is already resolved here. Site constants come from the PyAutoNerves companion PR (PyAutoLabs/PyAutoNerves#209).

• carbon.py: a Green Algorithms estimate, E = t × (n_c P_c u_c + n_m × 0.3725 W/GB) × PUE, C = E × CI.
  • Four cases: low; central (usage basis, requested memory); allocation (u = 1); high (high inputs on the allocation basis).
  • Ranges come from the site's low/high values. Carbon-intensity bounds are the spread across bases (ruling F6).
  • Energy is split into successful, failed/repeated and in-progress work, with values per completed task.
  • There are no Green Algorithms default fallbacks: if the site has no carbon-intensity table the output is energy only, and if W per CPU or PUE is missing there is no estimate. --pue, --tdp and --ci override the site values.
  • The NESO fetch is optional and cached (--neso), the --begin-hint is optional, and nothing uses the network by default.
  • forecast(StageForecast…) estimates a planned campaign.
• report.py: python -m autofit.campaign report works from a live or retro digest, or a campaign.yaml. It writes report.md + report.json covering compute, energy and carbon, the failure census, per-stage tables and results counts. It ends with a paste-ready paper paragraph that says "estimate", gives the range, names the carbon-intensity basis and the other bases, and lists what is excluded. "Saving" appears nowhere, and a test asserts it.

Heart RED override (development only)

Heart was RED at the start of this task (2026-10-30): release validation FAILED (stage integrate); PyAutoHeart/workflows/Release Integrate failure; PyAutoNerves test_fitsable.py::test__output_to_fits red; unit-test timing 1 regression (>3× baseline), 14 slow; autogalaxy red; release validation NOT release_ready (v2026.10.30.1.dev85101); autolens_profiling red; worktree drift 1 dirty. The human authorized the Human override for Heart RED (development only) live for hpc-campaign phases 3 and 4 ("Override, proceed"). This PR touches no RED item and does not fix Heart. Merging still needs green CI and a separate /prm; it permits no release.

Decisions taken (please check)

1. Carbon-intensity range: the range is the spread across bases (F6). For RAL, regional (188) is the highest basis, so the high bound uses no intensity above central. Half-hourly p5/p95 values are stored but not used, because a campaign averages over them.
2. Earlier attempts are charged with the final attempt's CPUs and memory. Tasks with no CPU time count at u = 1, GPUs at full TDP, and a partition the site doesn't list takes the widest range of the partitions it does list.
3. Successful work means the final attempt of a completed task that passed its success check. Pending, running and timeout-pending tasks go in an "in progress" bucket that counts toward neither share.

Test plan

• test_autofit/campaign: 151 passed on this stacked head, including report §1.4's DR1 worked estimate reproduced exactly (1,239 / 1,717 / 4,494 kWh; 124 / 223 / 854 kgCO2e). Also covered: both bases, the split, NESO mocked, and the CLI.
• Full PyAutoFit suite: 4,143 passed. Smoke test on the autofit workspace: 10 passed, 0 failed.
• Independent review: FINDINGS, fixed in 8e7a1f5:
  • a null carbon-intensity basis no longer crashes;
  • a report with no finished energy no longer crashes;
  • running tasks are no longer counted as failed;
  • the --ci basis label is correct;
  • report with no source gives a usage error.
• Witness W4: a DR1 estimate from the W2 retro digest, in Workspace/Science/scratch/hpc_campaign_p4_w4/ (not committed).
  • ≈3.5 MWh (range 1.7–6.0); ≈651 kgCO2e (range 221–1,119) on the regional basis.
  • 30% of the energy of finished work went on failed, timed-out or repeated runs.
  • That's about 2× the §1.4 forecast because of workload (22k vis_lp and 23k SED tasks against 5k each), not the model.

Refs #1746

🤖 Generated with Claude Code