[LisaGoh]

Closes #305. Part of #376.

Writes tomographic ξ± as one SACC part, and puts the OneCovariance ξ± covariance (#256) into SACC in the right order.

develop already wrote the non-tomographic ξ± as a sealed part (xi_to_sacc, #251/#253). This extends that same writer to tomographic bin pairs rather than adding a second one, as agreed in #375/#376. The pure_eb, test and plotting fixes from the first version of this PR are dropped (superseded on develop), and so is the earlier save_2pcf_sacc, which targeted the old sacc_io API.

OneCovariance order ≠ SACC order
OneCovariance writes the real-space matrix type-major: all ξ+ (bin pairs i ≤ j, i outer, then θ), then all ξ−. A ξ± part is pair-major: each pair's ξ+ then ξ−. The two agree only for a single bin, so a tomographic .dat attached as-is (what assemble_sacc did) would silently misalign the covariance. The matrix is now reordered to the part's points before it's attached.

What changes

• cosmo_val/sacc_writers.py
  • xi_pairs_to_sacc: one ξ± part over any set of bin pairs, pair-major, with an optional dense covariance or diagonal. xi_to_sacc is now its single-pair form.
  • onecov_xi_covariance(part, cov): OneCovariance's matrix permuted to the part's insertion order; refuses a matrix whose size doesn't match.
• sacc_io.read_onecov_matrix: reads the .dat (np.savetxt with a # header) and checks it's square and symmetric.
• cosmo_val/real_space.py, calculate_2pcf_version(compute_tomography=True):
  • every bin pair goes into one part, with a source_i tracer per bin;
  • pairs go in as (0,0), (0,1), …, (n-1,n-1), which is OneCovariance's order;
  • jackknife covariance via treecorr.estimate_multi_cov, so the correlations between pairs are kept; shot noise gives the diagonal;
  • the part is sealed under the catalogue's blind and kept in cv.xi_tomo_parts[ver, grid];
  • the returned values are xi_correlation views of the sealed part, not raw TreeCorr objects, with the same attributes.
• cosmo_val/core.py
  • sacc_nz(version, tomography=True): bin i takes column i + 1 of redshift_path.
  • Blinded catalogues now conceal tomographic ξ± (and ⟨M_ap²⟩ derived from it) instead of refusing it; blinding.shear already works per tracer pair. Tomographic pseudo-Cℓ are still refused.
• Workflow
  • run_2pcf.py --tomo; the xi rule output is …_npatch={npatch}{tomo_suffix}.sacc (also removes a duplicated npatch param).
  • New rule xi_onecov → …{tomo_suffix}_onecov.sacc: the ξ± part with the OneCovariance matrix of its binning (NG on the reporting grid, G on the integration grid) as its covariance, reordered, keeping the part's blind. Script: workflow/scripts/xi_onecov_sacc.py (also runs from the command line).
  • assemble_sacc.py reorders --xi-cov the same way.

Requirements

• Tomographic ξ± needs an n(z) column per bin. A single-column file for a multi-bin catalogue is refused.
• SACC bin i is the i-th smallest catalogue bin ID, and must be column i + 1 of the n(z) file. OneCovariance reads the same file, so its bins line up as long as columns are in ascending bin-ID order.

Tested

• New tests:
  • the multi-pair layout;
  • the type-major → pair-major reorder, against an expected order built independently;
  • the identity case for one bin, and refusal of a mismatched matrix;
  • read_onecov_matrix;
  • a tomographic part on a synthetic catalogue: tracers and n(z) per bin, pair order, jackknife correlations between pairs;
  • refusal of a single-column n(z);
  • sacc_nz(tomography=True);
  • concealment on a blinded catalogue (needs smokescreen).
• Updated fixtures: two existing tomographic tests now carry a per-bin n(z).
• Suites: test_cosmo_val, test_sacc_writers, test_sacc_io, test_assemble_sacc and the regression tests pass, apart from three failures that are already on the base commit (pure-E/B synthetic run, pure-E/B mapping, τ centring).
• Dry runs resolve for …_tomo.sacc, …_tomo_onecov.sacc and …_onecov.sacc; the cosmo_val paper DAG is unchanged.
• End-to-end: xi_onecov_sacc.py on a synthetic two-bin part with a type-major .dat recovers the expected pair-major covariance (to text precision) and keeps the blind stamp.

Not yet checked

• The reorder against a real tomographic OneCovariance run: it follows OneCovariance's source code, but there's no tomographic catalogue yet to produce one.

Out of scope

• Tomographic pseudo-Cℓ, COSEBIs, pure-E/B and ρ/τ parts (rest of Tomographic SACC writers: one tracer and n(z) per bin #376).
• 3x2pt (γ_t, w(θ)): deferred until there is a lens sample.

🤖 Generated with Claude Code

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[cailmdaley]

@LisaGoh i'm going to start working on getting #253 into shape now, in the meantime this can at least be synced up with all of the changes from develop!