prompt: record lacosmic CR dial + star-pass decoupling shipped (PyAutoReduce#66)

complete/2026/08/lacosmic-cr-option-and-star-pass-decoupling.md

@@ -0,0 +1,147 @@
+- issue: https://github.com/PyAutoLabs/PyAutoReduce/issues/61 + https://github.com/PyAutoLabs/PyAutoReduce/issues/62 (both deliberately OPEN — closure is the default-flip decision, not this ship)
+- library-pr: https://github.com/PyAutoLabs/PyAutoReduce/pull/66 (merged 2026-08-06, head 2168738e)
+- shipped: 2026-08-06
+- repos: PyAutoReduce
+- tests: 265 passed / 15 skipped, up from the 246/15 baseline (+19, numpy/astropy-only house style). Re-verified on main (9955812) 2026-08-18: kwargs/dial suite 34/34, star-pass decoupling 4/4.
+- record-note: reconciliation record (2026-08-18) — the implementation shipped from a `/wake_up` session branch (`claude/wake-up-q6ajyq`) without advancing this prompt out of `draft/`; this record closes that drift. No new code shipped with this record.
+
+Community-reported (@samlange04, PyAutoReduce#61 + #62): AstroDrizzle's unconditional
+`driz_cr=median=blot` stack rejection compares each frame against a blotted-median
+reference that reads systematically low at steep gradients, so genuine core flux can
+be flagged as CR — eroding deflector cores and holing field-star PSFs before
+`find_stars` ever sees them. Shipped as PR #66: a user-facing CR-method dial plus the
+decoupling of PSF-star measurement from the shipped science CR pass. **The science
+default is unchanged** — flipping it stays human-gated on the SLACS validation.
+
+## What shipped (PR #66)
+
+- **`TargetSpec.cr_method: "driz_cr" (default) | "deepcr"`** (#61). `driz_cr` keeps
+  byte-identical kwargs (pinned by regression test, including that no `resetbits`
+  key appears). `deepcr` = per-frame CR masks (reusing `package/cosmic_rays.py`, the
+  machinery frame products already run) written into each exposure's DQ as the
+  AstroDrizzle CR bit 4096 via the pure, idempotent `dq_with_cr_flags`, then a plain
+  weighted-mean drizzle (`driz_cr=median=blot=False`) — no blotted-median reference
+  anywhere. Fail-fast in `reduce_target` for non-astrodrizzle backends and
+  instruments without a registered deepCR model (e.g. `wfc3_ir`, whose CRs calwf3
+  already flags per-frame).
+- **The #61 `resetbits` trap is pinned by a named unit test**: the per-frame route
+  must emit `resetbits=0` — AstroDrizzle's default (4096) clears exactly the DQ bit
+  the masks live in, silently producing an unmasked drizzle that still looks
+  plausible (the reporter's first attempt scored a "flawless" core=1.000 that was
+  the no-CR image scored against itself).
+- **`TargetSpec.psf_star_pass: "auto" (default) | "science" | "no_cr"`** (#62).
+  `pipeline.py::_psf` draws star finding **and** stamp extraction from
+  `_star_pass_image(...)` instead of unconditionally from the science mosaic.
+  `auto` never adds a drizzle (uses the least-CR-rejected pass that costs nothing
+  extra); `no_cr` is the explicit opt-in second AstroDrizzle pass on the identical
+  grid with CR DQ flags treated as good (`final_bits | 4096`) and `resetbits=0` so
+  the science pass's flags survive in the inputs. The `reduction.json` `psf` block
+  **always** records `star_source_pass` (+ reason or star-pass kwargs) so the
+  coupling cannot silently regress — the #62 wiring requirement.
+- **Design doc**: `docs/design/hst_acs_pipeline.md` Stage 3 documents the dial as a
+  deviation-in-waiting (default stays STScI `driz_cr`; flip gated on SLACS
+  reference-bar validation vs *tuned* driz_cr, scale 1.5/1.2 per published STScI
+  reprocessing vs the 1.2/0.7 pipeline defaults we inherit); Stage 5 documents the
+  star-pass decoupling.
+
+## Traps and findings worth keeping
+
+- **Naming honesty: the dial is `"deepcr"`, not `"lacosmic"`.** The prompt title
+  says LACosmic (the reporter's method), but the per-frame masker this pipeline
+  already runs is deepCR — labelling it `"lacosmic"` would falsify provenance. Same
+  method class (per-frame, no median reference). An astroscrappy/LACosmic backend is
+  a clean follow-up **only if** validation shows a gap between the two.
+- **The reporter's follow-up grid (issue #61, 2026-08-09) overturned the headline
+  number.** The ~37% core erosion was an artifact of a dither-erasing TweakReg
+  misalignment in *their* pipeline (every frame aligned onto the first, collapsing
+  the stack to a single pointing — the degenerate condition the blotted-median
+  mechanism needs). With correct MAST alignment, driz_cr holds core ratio 0.998 vs
+  no-CR at **every** snr/scale/combine_type setting they tried; dither-erased,
+  driz_cr's incremental damage is ~23%. So driz_cr is fine on well-dithered,
+  correctly-aligned stacks; the failure regime is degenerate/small dither (or
+  erased dither). PyAutoReduce's stage-2 default (trust a-priori WCS, TweakReg only
+  on trigger) preserves dither, so the erosion likely never reproduces in our
+  default config.
+- **That correction reshapes the validation, and weakens the default-flip case.**
+  The discriminating test for deepCR-vs-tuned-driz_cr is a **poorly-dithered**
+  target (few exposures, small/degenerate dither, true single-exposure SNAP), not a
+  clean 4-dither box. The remaining case for per-frame as default is robustness
+  (clean core regardless of dither quality) rather than a fixed erosion on typical
+  data.
+- **`psf_star_pass="auto"` deliberately never doubles combine time**: on driz_cr
+  runs AstroDrizzle keeps no less-rejected intermediate (the median image IS the
+  biased reference), so auto = science mosaic + recorded reason; the dedicated
+  no-CR pass is an explicit opt-in.
+- **Honest gaps at ship time**: the deepcr route and no_cr star pass are exercised
+  through unit-tested pure functions and monkeypatched plumbing — no
+  drizzlepac/deepCR execution or real-data validation in that environment;
+  `no_cr` + `psf_backend="starred"` weights stamps with the science-pass noise map
+  (documented approximation); JWST image3's `outlier_detection` has the analogous
+  median-reference failure mode but `cr_method` there is rejected, not addressed.
+
+## Remaining (tracked elsewhere, not this task)
+
+- Default-flip decision: human-gated on the SLACS reference-bar validation
+  (`active/pyautoreduce_slacs1430_acs_comparison.md`) with a tuned-driz_cr arm, now
+  targeted at the poorly-dithered regime per the reporter's grid. #61/#62 close
+  with that decision, either way.
+- The reporter's 2026-08-09 correction comment on #61 is unanswered; a reply is
+  a `/community` (human-gated) action.
+
+## Original prompt
+
+# LACosmic per-frame CR masking option + decouple PSF-star pass from the science CR pass
+
+Type: feature
+Target: PyAutoReduce
+Difficulty: medium
+Autonomy: supervised
+Priority: high
+Status: draft
+
+Community-reported (@samlange04, PyAutoReduce#61 + #62, measured on SLACS
+ACS/WFC F814W): the unconditional `driz_cr=median=blot` route in
+`autoreduce/drizzle/combine.py::drizzle_kwargs_for` systematically erodes flux
+at steep gradients — the blotted-median reference reads low at any sub-pixel-
+shifted peak, so genuine core flux is flagged as CR. Reporter measured ~37%
+deflector-core flux loss, worse with looser thresholds (the reference model is
+the fault, not the cut), and full core preservation (peak 1.000, 1" 0.979)
+with LACosmic per-frame masking + plain weighted-mean drizzle. The same
+mechanism holes field-star cores before `psf/stars.py::find_stars` ever sees
+them (#62): rebuilding from a no-CR pass took their usable star count
+344 → 599 (+74%) and rescued 4 lens/filter pairs from model-PSF fallback.
+
+Scope (one task — the halves are coupled through the no-CR pass):
+
+1. **CR-method dial** on `TargetSpec` (sibling of `final_pixfrac` /
+   `final_kernel`): `cr_method: "driz_cr" | "lacosmic"`. LACosmic route =
+   per-frame Laplacian masking (van Dokkum 2001; `astroscrappy` or deepCR —
+   `frame_products` already runs deepCR per-frame, so there may be a lever to
+   reuse) written into the DQ array, then plain weighted-mean drizzle
+   (`median=False, blot=False, driz_cr=False`).
+   **Trap (reporter hit it, verify with a test):** the LACosmic drizzle pass
+   must set `resetbits=0` — the AstroDrizzle default 4096 clears exactly the
+   DQ bit the CR mask lives in, silently producing an unmasked drizzle.
+2. **Decouple star-finding from the science pass**: `pipeline.py::_psf`
+   currently measures stars on whatever `_combine` produced. Tier-1/1b star
+   extraction should draw from the least-CR-rejected pass available,
+   independent of which pass ships as the science mosaic (kernel shape is
+   pass-independent). Provenance records which pass fed the stars.
+3. **Default decision is human-gated**: whether `lacosmic` becomes the
+   default (reporter suggests so) is a documented deviation from STScI
+   defaults — justify against the SLACS reference-quality bar
+   (`docs/design/hst_acs_pipeline.md`) and validate on the
+   slacs1430+4105 comparison task (`active/pyautoreduce_slacs1430_acs_comparison.md`)
+   before flipping; landing it as an option first is acceptable.
+
+Validation: unit tests for the pure kwargs/decision functions (house style);
+real-data before/after on a SLACS ACS target measuring core-flux retention
+and usable-star count, mirroring the reporter's numbers. Include a tuned
+`driz_cr` comparison arm: our adapters set no `driz_cr_snr`/`driz_cr_scale`,
+so we run AstroDrizzle's aggressive pipeline defaults (scale 1.2/0.7) — the
+STScI-documented mitigation for bright-source flagging is raising
+`driz_cr_scale` (published reprocessing used 1.5/1.2), and the default-flip
+decision should compare LACosmic against *tuned* driz_cr, not only against
+our current untuned defaults.
+
+<!-- filed from /community triage of PyAutoReduce#61 + #62 (2026-07-31) -->

complete/index.md

@@ -6,7 +6,7 @@ Token-light navigation over the finished-work records (schema:
 only then grep a dated bucket. Curators: edit the band between the CURATED
 markers; everything below GENERATED is rebuilt.
 
-1015 records across 7 buckets.
+1016 records across 7 buckets.
 
 <!-- CURATED:START -->
 ## Highlights
@@ -76,6 +76,7 @@ _(curate hard-won records here — survives regeneration.)_
 - [intra-family-dep-floors](2026/08/intra-family-dep-floors.md)
 - [jax-grad-smoke-timeout-budget](2026/08/jax-grad-smoke-timeout-budget.md)
 - [jax-likelihood-mass-sensitivity](2026/08/jax-likelihood-mass-sensitivity.md) — Audit + fix for median-literal mass blindness across all jax_likelihood/datacube scripts (phase 2 of the findi…
+- [lacosmic-cr-option-and-star-pass-decoupling](2026/08/lacosmic-cr-option-and-star-pass-decoupling.md) — both deliberately OPEN — closure is the default-flip decision, not this ship
 - [llms-txt-census-fixes](2026/08/llms-txt-census-fixes.md)
 - [message-prior-xp-jax-trace](2026/08/message-prior-xp-jax-trace.md) — made the remaining message and compound-prior array-namespace paths
 - [messages-xp-stack-jax-trace](2026/08/messages-xp-stack-jax-trace.md) — replaced ten fixed-shape `xp.array([...])` message constructors

dashboard.md

@@ -11,13 +11,13 @@ Tasks only — the organism's health lives with the Heart (`/health`), not here.
 | [In flight](#in-flight) (`active/`) | 3 |
 | [Parked](#parked) (`parked.md`) | 1 |
 | [Planned](#planned) (`planned.md`) | 7 |
-| [Backlog](#backlog) (`draft/`) | 136 |
+| [Backlog](#backlog) (`draft/`) | 135 |
 
 Live on GitHub: [open issues](https://github.com/search?q=org%3APyAutoLabs+is%3Aissue+is%3Aopen&type=issues) · [open pull requests](https://github.com/search?q=org%3APyAutoLabs+is%3Apr+is%3Aopen&type=prs)
 
 ## Start here
 
-**Highest priority** (filed as `high`) — showing 12 of 31
+**Highest priority** (filed as `high`) — showing 12 of 30
 
 - [TRIAGE: needs manual review before routing](draft/triage/jax_zero_contour.md) — medium · safe · high
 - [Decide whether the clipper belongs in the search identifier](draft/feature/autofit/clipper_in_search_identifier.md) — autofit · medium · human-required · high
@@ -26,11 +26,11 @@ Live on GitHub: [open issues](https://github.com/search?q=org%3APyAutoLabs+is%3A
 - [Optimize pixelized Prodigy settings on the laptop GPU](draft/research/autolens_workspace_developer/pixelized_prodigy_laptop_gpu_phase_2_settings.md) — autolens_workspace_developer · medium · human-required · high
 - [Release does not sync __version__ stamps and workspace pins back](draft/bug/pyautobuild/release_version_sync_back_to_main.md) — pyautobuild · medium · supervised · high
 - [Make draft/ staleness detectable — `intake reconcile` measured, and the](draft/feature/pyautomind/draft_staleness_detection_signals.md) — pyautomind · medium · supervised · high
-- [LACosmic per-frame CR masking option + decouple PSF-star pass from](draft/feature/pyautoreduce/lacosmic_cr_option_and_star_pass_decoupling.md) — pyautoreduce · medium · supervised · high
 - [Cluster package: point-source-default narrative + extended-source follow-up feature](draft/docs/workspaces/cluster_regime_narrative.md) — workspaces · medium · supervised · high
 - [multi_galaxy package: new regime package in autolens_workspace](draft/docs/autolens/multi_galaxy_package.md) — autolens · large · supervised · high
 - [Tune the JAX multi-start optimizers into a standard option (MGE](draft/experiment/autolens_profiling/jax_optimizer_settings_tuning.md) — autolens_profiling · large · supervised · high
 - [Profile and speed up JAX likelihood-function compile times (all use](draft/feature/autolens_profiling/jax_compile_time_profiling.md) — autolens_profiling · large · supervised · high
+- [Optimize MultiStartProdigy for pixelized meshes on the laptop GPU](draft/research/autolens_workspace_developer/pixelized_prodigy_laptop_gpu.md) — autolens_workspace_developer · large · human-required · high
 
 **Quick wins** (small enough, and safe enough to run unattended)
 
@@ -82,7 +82,7 @@ Scoped but not started; some are not yet prompt files. Full detail in [`planned.
 
 ## Backlog
 
-**136** filed prompts, not started. Each section is sorted most-pickable first (priority, then size).
+**135** filed prompts, not started. Each section is sorted most-pickable first (priority, then size).
 
 <details>
 <summary><b>bug</b> — 36</summary>
@@ -127,11 +127,10 @@ Scoped but not started; some are not yet prompt files. Full detail in [`planned.
 </details>
 
 <details>
-<summary><b>feature</b> — 28</summary>
+<summary><b>feature</b> — 27</summary>
 
 - [Decide whether the clipper belongs in the search identifier](draft/feature/autofit/clipper_in_search_identifier.md) — autofit · medium · human-required · high
 - [Make draft/ staleness detectable — `intake reconcile` measured, and the](draft/feature/pyautomind/draft_staleness_detection_signals.md) — pyautomind · medium · supervised · high
-- [LACosmic per-frame CR masking option + decouple PSF-star pass from](draft/feature/pyautoreduce/lacosmic_cr_option_and_star_pass_decoupling.md) — pyautoreduce · medium · supervised · high
 - [Profile and speed up JAX likelihood-function compile times (all use](draft/feature/autolens_profiling/jax_compile_time_profiling.md) — autolens_profiling · large · supervised · high
 - [Give the Profiling Agent a compile-time axis — the arc](draft/feature/profiling/profiling_agent_jax_compile_time_scope.md) — profiling · large · supervised · high
 - [Which other searches need prior-support handling — coverage audit after](draft/feature/autofit/prior_exit_sampler_coverage.md) — autofit · medium · supervised · medium