PSF leakage normalises by a 2023 pre-SOM n(z) for every catalogue version

psf_systematics.py:291 builds dndz_path from the global nz block rather than the per-version entry:

params_in["dndz_path"] = (
    f"{self.cc['nz']['dndz']['path']}_{self.cc[ver]['pipeline']}_{self.cc['nz']['dndz']['blind']}.txt"
)

Every other field in set_params_leakage_scale is self.cc[ver][...]. Since pipeline: SP for all SP versions, this resolves to /n17data/mkilbing/astro/data/CFIS/v1.0/nz/dndz_SP_A.txt (Sep 2023) for every version — confirmed in all ~18 cosmo_val/output/leakage_*/stats_file_leakage.txt, fiducial included. The fiducial version's own n(z) (.../nz/v1.4.6.3/nz_SP_v1.4.6.3_A.txt, Mar 2026, on a different z grid) sits unused in the same config.

dndz_path reaches get_theo_xi at shear_psf_leakage/run_scale.py:1232, i.e. the theoretical xi_pm forming the denominator of the xi_sys/xi_theory leakage-significance ratio. So the "leakage is X% of the cosmological signal" number is normalised by the wrong n(z) for every version. xi_sys itself is unaffected and the rescaling is common across versions, so version-to-version comparisons hold; the absolute fraction does not. The hardcoded config-level blind: A also bypasses self.blind here.

Blast radius: no submitted paper depends on this. The only consumer of dndz_path is LeakageScale.do_xi_sys (shear_psf_leakage/run_scale.py:1232), whose ratio output is xi_sys_ratio.pdf. Papers I, III and IV all quote a leakage-as-fraction-of-signal number, but each normalises by a measured denominator, not a theory one: Paper I's plots/xi_sys_over_xi_plus.pdf is a different figure from a different script (papers/catalog/2025_10_02_xi_sys_rho_tau.py, denominator read from a TreeCorr xi_* file); Paper III's 10% scale-cut criterion is stated against the total measured signal; Paper IV's C_ell^sys/C_ell is built from measured harmonic spectra. Paper II quotes no leakage fraction at all. On disk this path has produced xi_sys_ratio.pdf for 19 catalogue versions across the two output trees checked (17 under comprehensive/output, 2 in a working checkout), none of them cited in any paper — and none for the fiducial SP_v1.4.6.3_leak_corr, whose leakage directory holds only stats_file_leakage.txt in both trees. So this is worth fixing before anyone reruns do_xi_sys and reads the ratio off it, but it is not a correction to anything published.

Start at psf_systematics.py:291 and inspect set_params_leakage_scale, then follow dndz_path into shear_psf_leakage/run_scale.py:1232 and get_theo_xi. Verify that each catalogue version uses its own configured n(z) and blind value, and check the resulting xi_sys_ratio output for the affected versions.