A percentile black/white stretch was tested first and made every sample
worse (these scans already use the full 0-255 range, so clipping just
blows out endpoints). A flat gamma=1.2 darken, fit against 12 finished
covers in singled/manual/, cut the error ~60% with no regressions - used
as the default here.
Automates the resize step of the Photoshop workflow. Always box-reduces
before the Lanczos resample (reducing_gap=3.0) instead of Pillow's default
single-pass filter, which is the source of moire on large downscales -
replacing the manual "blur if I see moire" judgment call.