[cailmdaley]

Every cosmo_val figure can now be redrawn from cached products without rerunning a transform, PSF fit or catalogue measurement.
cosmo_val_all still requests the complete analysis; the new cosmo_val_plots aggregate requests all figures and their required caches.
Plotting modules are explicit inputs, so editing their style invalidates figures rather than computations.

Closes #429.

Cached products and plot rules

Computation / existing product	Plot-only consumer
Pure E/B → sealed SACC + NPZ with comparison vectors	cv_plot_pure_eb: four figures
COSEBIs → sealed SACC + NPZ with every cut, integration edges and exact selected indices	cv_plot_cosebis: three figures
ρ/τ fits → posterior and propagated PSF-error NPZ	cv_plot_rho_tau_fits; cv_ratio_xi_sys_xi also reads sealed reporting ξ
Footprints → sparse pixel/count NPZ; weights → histogram NPZ; objectwise regression → coefficient JSON	cv_plot_footprints (NGC, SGC, fullsky), cv_plot_weights, cv_plot_objectwise_leakage
Existing ρ/τ FITS + explicit τ covariance; reporting ξ SACC; pseudo-Cℓ SACC/covariance	cv_plot_rho_stats, cv_plot_tau_stats, cv_plot_2pcf, cv_plot_pseudo_cl
OneCovariance .dat, with its own plotting disabled	covariance_plot: correlation PDF

Figures are declared outputs, replacing plot sentinels, and renderers overwrite existing figures.
The objectwise computation now calls the spin-2 regression directly instead of the exploratory driver that also emitted undeclared plots.
Presentation fixes retain signed PSF/pseudo-Cℓ signals, include uncertainties in pure-E/B limits, align COSEBIs heatmaps with the actual integration bins, and improve labels/layout.

Blinding contract

Pure-E/B and COSEBIs caches derive only from the already-sealed integration SACC; no unsealed ξ vector is added for plotting.
Their NPZs carry the inherited blind stamp, including none for public products, using the same sacc_io.derived_stamp check as derived SACC writers.
NPZ loaders fail closed on missing stamps and use sacc_io.expect to reject mismatches before reading signal arrays.
Plot rules and the B-mode summary pass the expected catalogue blind.
Old unstamped caches require a compute pass from the sealed products; manually stamping them is not a supported migration.

Plot-only command

Use the full run's config and output roots, on an allocation with the host's container/bind settings:

plot_rules=(cosmo_val_plots cv_plot_rho_stats cv_plot_tau_stats \
    cv_plot_rho_tau_fits cv_plot_footprints cv_plot_objectwise_leakage \
    cv_plot_weights cv_plot_2pcf cv_ratio_xi_sys_xi cv_plot_pure_eb \
    cv_plot_cosebis cv_plot_pseudo_cl covariance_plot)
snakemake --profile workflow/profiles/default -s workflow/Snakefile \
    --configfile /path/to/run.yaml --cores 4 --forceall \
    --allowed-rules "${plot_rules[@]}" -- cosmo_val_plots

The restricted rule list prevents computations, and the aggregate requires every cache even when all figures already exist.
A missing cache therefore stops the command.
Ordinary incremental restyling can use cosmo_val_plots without forcing; don't use unrestricted --forceall.
The pseudo-Cℓ include flag is respected.

Validation

• 198 passed, 2 skipped (catalogue-dependent), 257 warnings, in 204.29 s: Slurm job 25027245, def-sfabbro, 4 CPUs, 16 GiB, 1 hour requested; completed with exit code 0.
• Ran all 12 test files listed in the plot-split review report, plus test_plotfix_cell.py, test_plotfix_layout.py, test_plotfix_psf.py and Decode FITS logical columns to booleans before writing HDF5 catalogues #423's test_catalog_logical_roundtrip.py.
• Used /project/6001537/cdaley/containers/sp_validation-ddd58fe0.sif with this checkout's src/ first on PYTHONPATH; host Snakemake 9.23.1 supplied the workflow dry-runs.
  Log: /scratch/cdaley/cosmo-val-v2/plotsplit-review-25027245.log.
• uvx ruff check and uvx ruff format --check: all 29 changed/new, non-deleted Python files pass.
  git diff --check passes.

The workflow regressions exercise pseudo-Cℓ enabled/disabled, exclude compute jobs from the forced plot-only DAG, and remove every required cache with all figures already present to verify refusal.
Synthetic renderer tests forbid catalogue reads; provenance tests cover public and named-blind consumers.
The branch is rebased onto develop at 158c6d30, retaining #423's logical-column decoding and #424's default mask unchanged.

Not verified in this landing pass: no real-catalogue cosmo_val run, production write, production-scale timing measurement, or full-survey numerical comparison.
The public figures below illustrate the outputs; they aren't a new production validation run.

Example figures from the public blinded v2 dashboard

Source: production dashboard.
All four image URLs were checked with curl -I and returned HTTP 200.

Pure E/B and ambiguous ξ components

COSEBIs mode amplitudes

Galaxy–PSF/residual τ correlations

NGC footprint counts

🤖 Generated with Claude Code