Seed the jackknife k-means and drop the patch-centre file

[cailmdaley]

calculate_2pcf builds the jackknife patches for ξ± with TreeCorr's k-means, which picks a fresh random seed on every unseeded call. Develop hides this by drawing the centres once per output tree and reusing {ver}_patches_npatch=N.dat, but a fresh tree, a rerun elsewhere, or two jobs racing to write the file all split the sky differently, so the ξ± jackknife covariance changed from run to run. It is the same class of bug that moved ρ/τ statistics by up to ~3σ between reruns (fixed in shear_psf_leakage #43).

A seed alone is not enough: TreeCorr's automatic k-means starts from a ball tree whose depth follows the machine's OpenMP thread count, so for non-power-of-two patch counts (e.g. 100) a seeded Catalog(npatch=..., rng=...) still splits the sky differently on 12-, 24- and 48-CPU nodes. This PR takes the centres from a seeded k-means on a tree of pinned depth (statistics.jackknife_patch_centers: min_top=6), the same fix as shear_psf_leakage #47. The patches are then a pure function of the catalogue's positions and weights, so the centre file has no job left and goes too.

What changes

• calculate_2pcf and calculate_aperture_mass_dispersion build their catalogue through one helper that passes the pinned-depth centres as patch_centers; the patch-centre file, its reuse and its write are removed.
• The per-patch ρ/τ covariance path in rho_tau.py takes its centres from the same helper.
• The thread-count test becomes a reproducibility test: two fresh output trees with 100 patches on simulated 4- and 16-CPU machines give identical patch labels, per-patch-pair counts, ξ± and jackknife variance.

Effect on results

• ξ± values: essentially unchanged. Patches enter the covariance, and otherwise only decide which cell pairs TreeCorr's bin_slop approximates; not measured against develop's draws.
• ξ± jackknife covariance: reproducible across trees, reruns and thread counts. It differs from any earlier unseeded draw.
• Existing _patches_npatch=*.dat files are no longer read. They can be deleted.
• A variant that reweights the catalogue gets its own patches. One that keeps positions and weights shares them.

Cost

The k-means runs on every measurement. On 16 threads it takes about 1 s per 10⁶ objects (≈10–15 s at 10⁷ for 20–200 patches), which is small next to the pair counting.

Verified

• Reproducibility test: fails on the seed-only commit (2f70c93; 59% of objects change patch), passes on the pinned-depth helper.
• Container suite (-m "not slow"), with develop merged in: 307 passed, 1 xfailed.

— Claude (Opus) on behalf of Cail.

🤖 Generated with Claude Code

[sachaguer]

Wow a subtle thing found here. I validate the strategy and jackknife seeding but hint again at the conflict between the parallel PRs that we have to figure first I believe (see PR #366)

[cailmdaley]

Wow indeed, Claude will stop at nothing to vanquish irreproducibilities... yes pausing this until we figure out the API conflicts