refactor: post-phase-3 consolidation (behaviour-preserving)

[Jammy2211]

Overview

Consolidate structure debt from three fast instrument phases. Behaviour-preserving: identical products/provenance (modulo timestamps/version stamps) for identical inputs; public API unchanged. Plan approved in-session (default present-and-wait run; refactor cap is safe but no --auto was given).

Plan

• Decompose pipeline.reduce_target into per-stage functions (record layout byte-identical).
• Extract autoreduce/validation/parity.py (sub-pixel registration + ratio stats) — de-triplicates the integration scripts.
• Extract drizzle/_common.py (chdir-scratch context manager + shared provenance fragment) used by both combine backends.
• Vectorize psf/stars.reject_crowded (identical selection results; randomized equivalence test).
• Hoist cheap function-local imports; tidy exptime coercion.

Out of scope: _fwhm_of upgrade, dial defaults, tier-2 PSF (behaviour-visible).

Witnesses

1. 83-test unit suite per commit (+1 equivalence test).
2. Offline cached end-to-end re-runs, ACS (slacs0008-0004) and JWST (F444W ring): products byte-compared, reduction.json compared modulo written_at/software. Baselines captured from main (f91d79c) before any edit.

RefactorDecision: API guard none-expected; witness registry gap overridden with the above evidence.

[Jammy2211]

Witness-baseline capture immediately caught a phase-3 regression: the footprint filter's WCS(header) fails on ACS FLC headers (CPDIS lookup-table distortion needs the open HDUList; JWST cal files don't carry it, so the four-band integration never tripped it). One-line fix (fobj=hdul) landed as a labelled fix: rider commit ahead of the refactor — declared here so the PR diff reads honestly. Baselines re-capturing.

[Jammy2211]

Second rider caught by baseline capture: ACS ePSF build crashes at maxiters=10 in photutils 2.3 (recentering drift → malformed overlap_slices) — the star set changed since phase 1 via the saturation-cap and finite-window fixes, exposing it. Fix: extraction window +10 → +20 pad (photutils then warn-and-skips the drifting star as designed), edge margin 36 → 46 to match. JWST ePSFs shift marginally with the wider window — the refactor's byte-identity baselines are being captured with both riders in, so the refactor itself remains provably behaviour-preserving.

[Jammy2211]

Shipped: PR #10 is up. Witness verdict: all eight product files byte-identical across both observatory paths; reduction.json identical modulo timestamps — the refactor is provably behaviour-preserving. Two fix riders (CPDIS/fobj, ePSF window +20) disclosed above and in the PR body. 85/85 tests. Merge is yours.

[Jammy2211]

PR #10 merged as 371721f (squash) — human-directed in the keck-ao session (user authorized the merge in-conversation; the Keck-AO implementation builds on the consolidated layout). Main checkout returned to main; branch deletion left to the owning session / repo_cleanup.