feat: JWST frame products — implement the 7 feasibility deltas

[Jammy2211]

Overview

Implements the GO-phased verdict of docs/design/jwst.md §"Per-exposure frame products — feasibility" (#24, merged dfa0d96): extend the frames → registration → PSF chain (#16/#19/#21) to JWST/NIRCam. Prompt: feature/pyautoreduce/jwst_frame_products.md; user launched "--auto" 2026-07-10.

Plan (the 7 deltas, plus what implementation scoping added)

1. _crf capture — jwst_combine runs image3 with outlier_detection.save_results=True, captures the *_crf.fits paths into drizzle provenance; _package_frames uses them for JWST (tweakreg-updated WCS + outlier flags), falling back to _cal with recorded absence.
2. Units — _units_to_cps gains a surface-brightness branch: MJy/sr kept native, recorded; manifest data_units becomes derived (loud on heterogeneous inputs) instead of hardcoded e-/s.
3. Sky — read BKGLEVEL (skymatch level) when serialized; subtract + record, 0.0 recorded otherwise — mirroring the HST MDRIZSKY convention. Verified against real crf files during validation.
4. CR/DQ semantics — the meaningful divergence: JWST ramps remove CRs during slope fitting, so JUMP_DET pixels are good data; only DO_NOT_USE (bit 0, which outlier_detection also sets) means bad. JWST masking = dq & DO_NOT_USE, vs HST's any-nonzero-bit; cr_method = "ramp-jump (stage 1) + image3 outlier_detection (crf)"; no deepCR (none exists, none needed); dq_semantics switches to the JWST flag table. This lands the per-observatory dq_bad_bits refinement the roadmap anticipated.
5. WCS — v1 anchors target_pixel through the FITS-approx SIP the footprint filter already uses (gwcs upgrade recorded as follow-up); registration block's absolute-fit keywords record as absent (JWST keeps tweakreg fit info in datamodel meta, not FITS); the measured relative residuals carry over unchanged.
6. Per-frame ePSF — peak_max=None for surface-brightness units (the mosaic convention); exposure time via EXPTIME→XPOSURE→EFFEXPTM fallback (also weights psf_from_frames); STPSF stays the tier-2b roadmap item.
7. Guard — frame_products/psf_from_frames allowed for observatory in ("hst", "jwst"); the loud non-supported case moves to ground-based (existing tests updated from nircam to nirc2).

Validation: COSMOS-Web ring F115W (undersampled SW) end-to-end with frame_products=True, psf_from_frames=True from the warm cal cache; LW spot-check if runtime allows.

Supervised: proceeds to implementation; ship sign-off parks per contract.

[Jammy2211]

Ship sign-off (supervised run parks here) — JWST frame products implemented + validated on COSMOS-Web F115W

Implementation complete in ~/Code/PyAutoLabs-wt/jwst-frame-products/PyAutoReduce (branch feature/jwst-frame-products, uncommitted pending sign-off). Suite 194 passed (15 new JWST + reliability tests).

All 7 deltas landed, plus four facts the implementation/validation added (now in jwst.md):

• DQ policy divergence is load-bearing: JWST masks dq & DO_NOT_USE only — a JUMP_DET pixel is good data (ramps remove CRs); the ePSF estimator likewise patches only DO_NOT_USE (patching informational bits would smooth large chip fractions).
• Frame identity from filename stem (JWST has no ROOTNAME; the old derivation collided across a visit).
• Manifest schema v2: derived data_units (loud on mixed), sky_subtracted/sky_keyword (BKGLEVEL generalises MDRIZSKY), source (_crf vs _cal fallback).
• Registration reliability guard — the real-data catch of this validation: COSMOS-Web dithers put the target near detector edges (off-chip fractions 0.32–0.63), and correlating mostly-masked cutouts produced ~200 px mask-geometry artifacts. Now: best-covered frame is the reference, pairs >20% masked are flagged residual_reliable: false, and the headline max needs a clean pair — on this dataset it honestly reports UNMEASURED (only one clean frame exists).

End-to-end validation (F115W, undersampled SW, 12 exposures from warm cache): image3 ran with the new _crf capture (✓ saved + consumed: source: image3 _crf products); 6 chips packaged, 6 correctly skipped (target off-detector); real BKGLEVEL values (0.01–0.05 MJy/sr) read + subtracted; native MJy/sr units throughout; per-frame tier-1 ePSFs viable on 3/6 frames (9–12 stars; failures are genuine 4–7-star fields, recorded + loud); psf_from_frames built the delivered mosaic PSF from those 3 frames — the first combined-from-frames JWST PSF, with a moment-FWHM diagnostic (the DAO-based estimate can't see undersampled SW kernels).

On approval: commit, push, pending-release PR (API: additive — JWST support in frame_products/psf_from_frames, guard message change, manifest v2 field renames disclosed as Changed Behaviour), merge stays per your call. Known follow-ups for the record: STPSF tier-2b (would lift the 3/6 PSF coverage to 6/6), gwcs-anchored target_pixel, and the duplicate frame-ePSF compute when both flags are on.