Add hacks/make_superbias.py: per-chip superbias / NMAD / bad-pixel mask

[nugent68]

Summary

Adds a single new script, hacks/make_superbias.py, that builds three Rice-compressed FITS products per LS4Cam chip from a stack of bias frames:

File	dtype	Header keyword	Contents
superbias_<CCD>.fits.fz	float32	MED_BIAS	per-pixel median across the bias stack
superbias_nmad_<CCD>.fits.fz	float32	MB_NMAD	per-pixel NMAD = 1.4826 × median(|x − median(x)|)
superbias_mask_<CCD>.fits.fz	int16	NMASKED, MASKTHR	1 where NMAD > 3 × 1.4826 × median(NMAD), else 0

Two input modes

• --input-mode multi-ext (default): each input is a multi-extension raw exposure with the bias section still present; the script calls LS4Cam.overscan_and_trim per chip extension internally.
• --input-mode per-chip: a flat directory of already-trimmed per-chip FITS files, grouped by the CCD_LOC header keyword. Different chips may have different N. Files lacking CCD_LOC are skipped with a warning.

Verified for bit-exact equivalence on a 31-frame test set: per-chip mode and multi-ext mode produce identical superbias, NMAD, and mask images for the same underlying frames.

Real-data shakedown

Run on 31 single-amp LS4Cam bias frames (992 chip-images):

• 32 chips processed in ~16 min on Docker Desktop / Intel Mac
• MED_BIAS clusters at 0 ± 2.5 ADU on every chip ✅
• MB_NMAD quantizes to {3.71, 4.45, 5.19} (= 1.4826 × {2.5, 3.0, 3.5}) — expected artifact of int16 raw + N=31
• 130,535 total bad pixels masked across the focal plane (mean 0.049%)
• NW_B and NW_C together account for 48% of all masked pixels — visually confirmed bad columns

Test plan

• Run with default --input-mode multi-ext against a directory of LS4Cam multi-ext bias frames; verify 3 outputs per chip and reasonable header values.
• Run with --input-mode per-chip against an already-trimmed per-chip directory; verify same outputs.
• --chips NE_G NW_B filter restricts processing to a subset.
• CI lint passes (ruff check hacks/make_superbias.py is already green).

🤖 Generated with Claude Code