test: Keck-AO acceptance checks 3-4 — astrometry + Einstein-radius invariance

[Jammy2211]

Overview

Completes the two keck_ao.md validation-anchor acceptance checks left open at the phase-4 merge (#11, PR #12): astrometric parity of the B1938+666 reduction against HST, and science invariance — a PyAutoLens fit reproducing the published Einstein radius (~0.45"; Lagattuta et al. 2012 / Vegetti et al. 2012).

Plan

• Survey MAST for archival HST imaging of B1938+666 (WFC3-IR preferred — reducible by our own phase-2 path for a like-for-like comparison); fall back to published positions.
• Audit the NIRC2 narrow-camera plate scale: raw header PIXSCALE 9.952 vs adapter native_scale 9.942 mas (literature spread 9.942/9.952/9.971) — a direct multiplicative bias on any measured Einstein radius; correct the adapter if the audit demands it.
• Fit the reduced dataset (10 mas, 39x180s, tier-A PSF) with PyAutoLens: SIE + external shear, Sersic lens light + Sersic source; compare theta_E with published values.
• Report here; parity appendix for docs/design/keck_ao.md and any adapter fix go to review before a PR (default present-and-wait — no --auto on this leg).

Prompt: PyAutoMind feature/pyautoreduce/keck_ao_acceptance_checks.md. Context: #9, #11, PR #12.

[Jammy2211]

Interim results + parked fit (laptop reset)

Check 3 — plate-scale audit (finding, fix gated on review): the adapter's narrow-camera native_scale (9.942 mas, from SHARP paper prose) is wrong for both epochs. Truth: 9.952 mas pre-2015 (Yelda 2010 — B1938 raw headers agree: PIXSCALE 0.009952) and 9.971 mas post-2015 (Service 2016). Effect on shipped phase-4 mosaics: output pixels are truly 10.010 mas, so a naive θ_E reads 0.10% low. Proposed fix: epoch-aware native_scale selected alongside the distortion solution.

Check 3 — geometry leg: no ACS/WFC3 archival imaging of B1938 exists (only 1997–98 WFPC2/NICMOS), so no own-pipeline HST comparison; and the ring is below direct azimuthal-profile detection in the 7,020 s stack (annulus medians at the noise floor) — the model fit carries the geometric test, as in Lagattuta+12.

Check 4 — fit CHECKPOINTED, resumable: Nautilus (SIE+shear+Sersic lens, SersicCore source, 20 mas Chen-style binning, plate-scale-corrected pixels, tier-A PSF) ran ~1 h on CPU; posterior not yet converged; 94 MB checkpoint.hdf5 on disk.

Resume after reset (Nautilus resumes from its checkpoint via the identity-hashed output path):

cd ~/Code/PyAutoLabs/PyAutoReduce
NUMBA_CACHE_DIR=/tmp/numba_cache MPLCONFIGDIR=/tmp/matplotlib \
PYAUTO_SKIP_WORKSPACE_VERSION_CHECK=1 \
~/venv/PyAuto/bin/python prototypes/b1938_lens_fit.py

Inputs it needs are all on disk under prototypes/output/b1938_keck/ (dataset regenerated post-merge) and output/b1938_keck_acceptance/ (checkpoint). Success criterion: θ_E consistent with published ~0.45″.