Merge develop into the tomography branch

[cailmdaley]

Step 3 of the epic: merge develop (D) into feature/sp_validation-extend-to-tomography (T) on an integration branch, with a PR into T.

Resolution: keep T's API (bin-pair loops, renamed functions). Bring in D's:

• I/O: SACC writers, open_entry/_CatalogueLoader, mandatory patch_number;
• pure-E/B operator API;
• uv.lock.

Ports, as separate commits:

• Include the HEALPix pixel window in map-based pseudo-Cℓ bandpower windows #366: apply the pixel window pw² only when cell_method == "map", in both the SACC window and the iNKA fiducial.
• Seed the jackknife k-means and drop the patch-centre file #367: seeded jackknife centres, computed once per version and shared across bins.

Gate: the container test suite passes. Non-tomographic outputs (pseudo-Cℓ, ξ±, ρ/τ, COSEBIs, pure-E/B) on SP_v1.4.6.3 match D, apart from those ports.

Check shear_psf_leakage: confirm that the merged uv.lock pins a version that has CovTauTh mask support and the two-argument compute_xi_psf_sys. D's 649edf4 has both; T's 0b3c175 has neither. Then rerun the tomographic ρ/τ on the merged branch, rather than trusting earlier T outputs.

Review: Sacha. Lisa reviews the real-space/ξ± part.

Traps git won't flag (found by reading the code):

• pol_factor changed from a bool to ±1. D's config.yaml value true would silently drop the e2 flip, so set it to −1.
• summarize_bmodes swallows the KeyError from the new result shapes.
• Taking D's side in cosmo_val/pseudo_cl.py creates duplicate method definitions.
• Several workflow scripts and cv_basename still assume the old names.

— Claude on behalf of Cail

[cailmdaley]

These regression tests from the sp_validation audit should pass on the merged branch. They are on branch assay/decision-record (audit epic #377). The tests in the first group already pass on develop, so they guard against the merge reintroducing the tomography branch's versions.

Develop is right; the tomography branch regresses

• Joint xi+^B/xi-^B chi2 uses Hartlap p=n instead of 2n, so the joint PTE is too low (b_modes.py:410). Keep develop's hartlap_factor(N, 2*nbins_eff). src/sp_validation/tests/regression/test_tomo_joint_pure_b_hartlap_dimension.py
• treecorr_config drops min_top=6 and num_threads, so xi± depends on the node's CPU count: 0.14 sigma between 16 and 48 CPUs on a jackknife fixture, 0.66 sigma at 256. src/sp_validation/tests/regression/test_tomo_drops_min_top_pin.py
• The MC pure-E/B covariance rebins fine xi with a uniform mean instead of TreeCorr pair weights. sigma(xi+^B) shifts by 3-5%, and the [12,83]' B-mode PTE moves from 0.0385 to 0.0428. src/sp_validation/tests/regression/test_tomo_mc_covariance_uniform_rebinning.py
• Covariance and tau rules call scripts in an absolute external checkout, and GLASS outputs land outside the launch directory. src/sp_validation/tests/regression/test_tomo_covariance_rules_call_external_checkout.py
• run_rho_tau.py chdirs into a hard-coded external checkout and reads its cat_config.yaml and output/. src/sp_validation/tests/regression/test_tomo_rho_tau_driver_external_checkout.py

The tomography branch is right; develop is wrong

• rho/tau ignore the configured HSM_FLAG_PSF/HSM_FLAG_STAR cut. Keep the tomography branch's _get_star_mask, forwarded to rho, tau and CovTauTh. src/sp_validation/tests/regression/test_rho_tau_ignores_hsm_flags.py
• The rho/tau least-squares sample cache is not keyed on cov_estimate_method. Keep the tomography branch's cov_type key. src/sp_validation/tests/regression/test_lsq_sample_cache_ignores_cov_type.py

Defects in the tomography branch's own code

• plot_pure_eb passes the bin-keyed result to calculate_eb_statistics and raises KeyError('gg'). Keep tomographic iteration. src/sp_validation/tests/regression/test_tomo_pure_eb_mapping_keyerror.py
• plot_cosebis hardcodes compute_tomography=False, so the flag has no effect. src/sp_validation/tests/regression/test_tomo_plot_cosebis_ignores_tomography.py
• _merge_iNKA_covariance fills lower cross-polarisation blocks with COVAR_{pa}_{pb}.T instead of COVAR_{pb}_{pa}.T. All 12 cross-pol blocks are wrong. src/sp_validation/tests/regression/test_tomo_inka_merge_crosspol_transpose.py
• Gaussian-sim workers write ..._seed_{seed}.npz, but the covariance reader looks for the unsuffixed name. It fails in a fresh directory and reads stale spectra otherwise. src/sp_validation/tests/regression/test_tomo_gaussian_sims_seed_filename_mismatch.py
• Per-bin tau: Catalogs.build_catalog in shear_psf_leakage removes the full-catalogue mean before applying the bin mask, leaving a tau_0/2/5 offset in each bin. The fix belongs in shear_psf_leakage; the test also fails on develop. src/sp_validation/tests/regression/test_tomo_tau_centering_before_bin_mask.py
• The pseudo-C_ell covariance producer is broken on both branches. On develop, get_theo_c_ell now returns a dict (TypeError); see also the ell=0 convention in Fix one-multipole shift in the pseudo-Cℓ Gaussian covariance theory #372. The tomography branch calls the removed calculate_pseudo_cl_eb_cov, and cv_pseudo_cl.py omits pol_list. src/sp_validation/tests/regression/test_pseudo_cl_covariance_producer_crashes.py

Claude (Opus 5.5) on behalf of Cail