felt: ngmix delivery outcomes + ci-green ground truth (post constitution-sweep)

Author: cailmdaley

.felt/shapepipe/ci-green-on-develop/ci-green-on-develop.md

@@ -7,7 +7,7 @@ tags:
     - ci
     - constitution
 created-at: 2026-05-29T12:34:31.800189806+02:00
-outcome: 'Conda removed from ShapePipe (#733 merged); develop is green including image publish. The container (slim Python 3.12 + apt + uv lockfile) is the single source of truth and CI runs pytest inside the dev image. ci-release.yml + doc-tests.yml + install_shapepipe + environment*.yml deleted; cd.yml builds docs in the image; install docs rewritten. The conda path had rotted into a multi-layer install failure — deleted rather than repaired, so the production astropy bug is moot. Remaining for a clean/tight/shared dev environment: (1) commit a shared .felt in the repo + curate which fibers go public; (2) make repo CLAUDE.md a real onboarding doc; (3) port cluster job scripts off conda; (4) #729 dependabot actions (recreating); (5) verify docs deploy on master; (6) strengthen the thin test suite (#708).'
+outcome: 'Near-realized. Ground truth 2026-06-11: develop green incl. image publish; #708 closed (test scaffolding landed on develop: 5 structural test files + property/synthetic-FITS tiers); candide PBS scripts modernized to SLURM+apptainer (#737); CLAUDE.md already a solid onboarding doc (minor gaps: CONTRIBUTING pointer, first-PR walkthrough); docs deploy infra correct post-#738 — /latest/ live, switcher published, ROOT still serves 2022 v1.0.1 until the next master push (the one remaining trigger). Conda survives only in out-of-scope CANFAR/CC-IN2P3 scripts (job_sp_canfar.bash + init_run_exclusive_canfar.sh have ACTIVE conda paths; cc_{mpi,smp}.sh use ccenv anaconda) — needs its own cluster-aware pass to fully realize ''no conda anywhere''.'
 ---
 
 ## Desired State

.felt/shapepipe/ngmix-size-columns/ngmix-size-columns.md

@@ -7,7 +7,7 @@ tags:
     - shapepipe
     - ngmix
 created-at: 2026-06-05T22:30:50.004535516+02:00
-outcome: 'Realized + fresh-eyes reviewed, no code defects: shapepipe fix/ngmix-size-columns (honest r50 + dedupe, pushed), cs_util feat/size-conversions (size web, 24/24, on cailmdaley fork — no upstream push rights), sp_validation fix/psf-leakage-fwhm (T_to_fwhm bug fixed — moves α(size); pushed). PRs + merge order (cs_util release → sp_validation) are Cail''s gesture; caution: sp_validation CI is green-but-vacuous on the cs_util.size import, ordering is discipline-enforced. See report.html.'
+outcome: 'Delivered end-to-end 2026-06-11: cs_util#65 merged + released as v0.2.1 on PyPI (size web, 26/26 — first release carrying cs_util.size); shapepipe#743 (honest r50 + dedupe, targeting ngmix_v2.0 per merge-base) merged, #741 CI green after; sp_validation#198 (T_to_fwhm fix + galaxy smoke test killing the green-but-vacuous CI hole, pin cs_util>=0.2.1) OPEN — blocked by shear_psf_leakage''s poetry cap cs-util ^0.1.0; relaxation branch fix/relax-cs-util-cap pushed to CosmoStat/shear_psf_leakage, PR creation is Cail''s click. Deferred (stated in #743 body): wiring honest r50/r50psf into make_cat so paper-convention r_h reaches the science catalogue.'
 shuttle:
     kind: oneshot
     host: candide

.felt/shapepipe/ngmix-weights-ivar/ngmix-weights-ivar.md

@@ -1,13 +1,14 @@
 ---
 id: 01KTCQPE3JGEYN7NQS8HW1AT6B
 name: 'ngmix weight map: fix v2.0 regressions + inverse-variance (#604)'
-status: open
+status: closed
 tags:
     - constitution
     - shapepipe
     - ngmix
 created-at: 2026-06-05T22:30:49.970813955+02:00
-outcome: '#604 implementation is wired through code and active example configs in /tmp/pr740-wt through commit 22e31e51: SExtractor emits BACKGROUND_RMS, vignetmaker stores background_rms_vignet*.sqlite, ngmix templates point BKG_RMS_VIGNET_PATH at that sqlite, ngmix builds per-pixel 1/RMS² inverse-variance gated by weight/flag/valid-RMS masks with scalar fallback when no RMS file is supplied, and test_ngmix now also guards the optional seventh RMS sqlite input contract. Focused RMS/weight tests and py_compile passed in the dev image; full test_ngmix.py still has the pre-existing ngmix.fitting.Fitter API failure.'
+closed-at: 2026-06-11T02:37:33.556855912+02:00
+outcome: 'Realized, fresh-eyes reviewed (no blockers), and delivered to PR #741 (CosmoStat/shapepipe, ngmix_v2.0): per-pixel 1/RMS² inverse-variance weights gated by weight/flag/valid-RMS masks with scalar sigma_mad fallback — feat f466c987, mask binarization 6ddea5b2, NaN-edge guard 4aa3b2a1, integration test pinning spatially-varying RMS through the full observation chain 158986a4, config wiring 7a44abda + all-or-nothing semantics documented 0bc6016e. Martin''s #604 weight-map thread answered with the commits. BKG_RMS_VIGNET_PATH is all-or-nothing per tile (missing file → FileNotFoundError; fallback only when option absent). Container gotcha: test from a scratch-wf worktree needs --bind /automnt/n17data and /automnt-prefixed paths (plain /n17data symlink invisible inside the sandbox); dev sandbox ships ngmix 1.3.6 so test_metacal_is_reproducible_with_fixed_seed only passes in CI''s v2 image.'
 shuttle:
     enabled: true
     kind: oneshot