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).

This PR seeds the k-means (rng=np.random.default_rng(0)). Seeded, the patches are 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: seeded k-means; the patch-centre file, its reuse and its write are removed.
• calculate_aperture_mass_dispersion builds its own patched catalogue and gets the same seed.
• The thread-count test becomes a reproducibility test: two fresh output trees at 4 and 16 threads give identical per-patch-pair counts, and matching ξ± and jackknife variance. With the seed removed, the test fails.

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

• Container suite: 306 passed, 1 xfailed.

— Claude (Opus) on behalf of Cail.

🤖 Generated with Claude Code