[cailmdaley]

Tomographic bins need z-band photometry, so the binned sample covers 3,815 of the 4,690 deg² the whole catalogue covers. Every tomographic covariance used the whole-sample area and footprint. On UNIONS v2 (size bound 9), that made the analytic shape-noise term about 1.32× too large.

What changes. cov_th.tomo can carry the binned sample's footprint, set as mask and A together. When it does, every tomographic consumer reads it:

• the workflow's tomographic OneCovariance ini, for both the area and mask_file_lensing;
• CatalogCharacterization.calculate_n_eff_gal(tomography=True), so per-bin n_eff is a density over that same footprint;
• the pseudo-Cℓ OneCovariance call when tomographic.

The workflow refuses A without mask, or mask without A. Entries without them, and every non-tomographic covariance, are unchanged.

n_eff and A must describe the same footprint. If only A is corrected, the noise term goes wrong the other way, by (4690/3815)² ≈ 1.51. The refusal and the shared footprint enforce this.

Covariance diagonal / 100-patch jackknife variance against θ, per bin pair (ξ+ solid, ξ− dotted). Grey is the OneCovariance total with the whole-sample A. Black is the re-run with the binned sample's A, mask and n_eff (Nibi job 25591046, same ξ± covariance inputs as this PR). Blue is the predicted shape-noise term. Median over pairs: θ < 10′ 1.33 → 1.08, 12–83′ 1.21 → 1.03. The remaining ~8% at small scales is weight variation inside the footprint, which only pair-count noise captures.

Consequence for the v2 null tests. The corrected covariance turns the joint pure-mode B-mode tests from pass to fail: ξ−B χ²/dof 85/70 (PTE 0.11) → 105/70 (0.005), ξ±B PTE 0.21 → 0.004. COSEBIs move from 0.97 to 0.44. The pseudo-Cℓ covariance (iNKA, catalogue) doesn't depend on area and is unaffected.

Tests.

• workflow/tests/test_onecov_inputs.py checks the per-bin area in the tomographic ini and the whole-sample area in the non-tomographic one, and that A without its footprint is refused. Workflow tests: 20/20.
• Container tests touching covariance, characterization and pseudo-Cℓ: 112 pass.

🤖 Generated with Claude Code

https://claude.ai/code/session_01JK7o9EG5R1yAPT98V9j2GJ