subtract_dark(scale=True) promotes float32 inputs to float64; scale=False does not

[mwcraig]

This issue was written and opened by Claude Code at the direction of Matt Craig (@mwcraig).

Environment: ccdproc 2.4.3, astropy 7.1.0, numpy 2.3.5.

With a float32 image and a float32 master dark, subtract_dark(ccd, master, scale=True)
returns float64 data; scale=False returns float32. Cause (ccdproc/core.py): the
scale=True branch does

master_scaled = master.copy()                                          # line 722
master_scaled = master_scaled.multiply(data_exposure / dark_exposure)  # line 725

data_exposure / dark_exposure is a 0-d astropy.units.Quantity, treated as float64 rather
than a weak scalar; multiplying the float32 master by it upcasts to float64, so
ccd.subtract(master_scaled) then returns float64 even though ccd is float32. The
scale=False branch calls ccd.subtract(master) directly with no intermediate multiply, so it
stays float32. The master.copy() on line 722 is also unnecessary: multiply() already
returns a new object.

Reproducer and output:

import numpy as np
from astropy.nddata import CCDData
import astropy.units as u
import ccdproc

rng = np.random.default_rng(0)
light = CCDData(rng.random((50, 50)).astype(np.float32), unit=u.adu, meta={'exposure': 30.0})
dark = CCDData(rng.random((50, 50)).astype(np.float32) * 0.1, unit=u.adu, meta={'exposure': 10.0})

r_scaled = ccdproc.subtract_dark(light, dark, exposure_time='exposure',
                                  exposure_unit=u.second, scale=True)
r_unscaled = ccdproc.subtract_dark(light, dark, exposure_time='exposure',
                                    exposure_unit=u.second, scale=False)
print(r_scaled.data.dtype, r_unscaled.data.dtype)

Output: float64 float32

Suggested fix: cast the ratio to master.dtype (or use a plain Python float) before
multiplying, e.g. master_scaled.multiply(np.array(data_exposure / dark_exposure, dtype=master.dtype)), and drop the .copy() on line 722.

Cost in practice: on a 4096x4096 float32 light with a scaled dark, this promotion (extra copy
plus float64 intermediate) measured 336 MB peak vs 201 MB when the ratio is pre-cast to float32
before scaling.