A/B validation of off-by-default shape-measurement options #813

Several ngmix-module options ship off by default, pending real-data evidence — the design contract is that production defaults flip here, on data, never in the implementing PR. Each sub-issue is one A/B question with a named metric and a named decider.

Current options awaiting a verdict:

• BLEND_HANDLING = uberseg (ngmix: config-selectable UberSeg neighbour masking (BLEND_HANDLING) #770 / UberSeg neighbour masking (Erin Sheldon) — config-selectable, validated against noise-fill #776) — UberSeg neighbour masking vs the historical noise-fill.
• Weight symmetrisation (4-fold) — follow-on PR to ngmix: config-selectable UberSeg neighbour masking (BLEND_HANDLING) #770, per @aguinot's recommendation on UberSeg neighbour masking (Erin Sheldon) — config-selectable, validated against noise-fill #776.
• Metacal reconvolution kernel azgauss vs fiducial fitgauss (A/B: azgauss vs fitgauss metacal reconvolution kernel #830) — noise-robust variant of gauss, selectable via the METACAL_PSF knob added in Selectable metacal reconvolution kernel + azgauss; bump ngmix v2.4.1 #829 (mechanism tracked separately as Update ngmix v2.4.0 → v2.4.1; expose metacal reconvolution kernel as config knob #819).

The decisive statistics (B-modes, leakage) need real area, which is why this lives in the First run on data (Nibi) milestone: the A/B runs are consumers of the P3 footprint (#801, #808) — pipeline run twice (or more) with config variants only, matched catalogues compared through the sp_validation machinery. Complementary evidence from image simulations (calibrated m/c truth) rides UNIONS-WL/MultiBand_ImSim#1.

Methodology (area needed, patch choice, matched-object vs matched-area comparison) needs group design — to discuss on a tomography call before the runs are scheduled.

— Claude, on behalf of Cail