A/B: BLEND_HANDLING = uberseg vs noisefill — do the FLAGS=2 B-modes go away? #814

noisefill (the default) zeroes and noise-fills only pixels flagged in the mask/weight; light from neighbouring detections stays in the stamp. uberseg masks pixels the segmentation map assigns to a neighbour, plus sky pixels nearer to a neighbour than to the object; it needs the segmentation stamp from vignetmaker run 3, which should only run when uberseg is on.

Fabian's hypothesis: the B-modes seen in FLAGS=2 (deblended) objects come from unremoved neighbour light.

Test: both settings on the blended image sims (m-bias, shape noise) and on a data patch (FLAGS=2 subset B-modes). A blended-sims uberseg run from July exists, uninspected.

Before running: under uberseg, flagged pixels get zero weight but are not noise-filled (ngmix.py:1647–1665), so this A/B currently also switches off the defect fill. Keep noise-filling flag≠0 / weight==0 pixels under uberseg so the two arms differ only in neighbour handling.