OriginStack 2.20.0

Which download?

You have	Download
Windows, want it installed	OriginStack-2.20.0-setup.exe
Windows, no install (one file)	OriginStack-2.20.0-windows-x64-portable.exe
Windows, no install (folder)	OriginStack-2.20.0-windows-x64.zip
Mac with Apple silicon (M1 or later)	OriginStack-2.20.0-macos-arm64.zip
Mac with Intel	OriginStack-2.20.0-macos-x86_64.zip
Linux	OriginStack-2.20.0-linux-x64.tar.gz

The learned deconvolver now works on other cameras: it rescales to the star width it was trained
on, reads monochrome and tile-compressed FITS, and --auto picks it when it can measure the stars.
The desktop app compares any Phase 4 step before and after, and there is a per-frame report of why
each frame was kept or rejected.

Added

• Learned deconvolver on other cameras. --deconvolve nn measures the star width in your
  stack and, when it is outside the 2.3-5.2 px it was trained on, resamples the image to the
  nearest bound, runs the network and resamples back, so a different pixel scale no longer
  matters (checked on Seestar S50 and Las Cumbres Observatory stacks, against Pan-STARRS). Narrow
  stars (under about 4 px) keep their flux through a star-aware restore. Stacks in the trained
  range, such as Origin, are unchanged. --auto now chooses it off Origin too (at least 3
  accepted frames), and Phase 4 puts the previous settings back, with one log line, when the
  stack has no fitted star, fewer than 30 usable stars, a median width outside 1.1-14 px or a
  median ellipticity above 0.35.
• Monochrome, tile-compressed and cube FITS. Mono FITS (header says no Bayer pattern, or has a
  filter) is calibrated on its single plane and replicated to three channels, as mono SER/XISF
  already were; .fits.fz frames (Las Cumbres) and 3-plane FITS cubes load.
• --nn-deconv-blend FRAC sets how much of the deconvolver is kept against the learned
  denoiser (default automatic: 0.75, galaxies 0.35 under --auto). An explicit value now wins
  over --auto, including 0.75.
• --frame-report writes <stem>_frames.csv and .json: each light's verdict, the stage that
  decided it, a plain reason ("registration residual 4.93 px RMS > 1.50 px") and its metrics.
• Calibration masters and mixed folders. A larger master is cropped to the lights when the
  binning matches and the offset is known; other mismatches are dropped with a warning that says
  why, on every path (it was silent on some). Lights that differ in binning, frame size, gain/ISO
  or camera are split into separate stacks instead of mixed.
• Desktop app: a before/after window for any Phase 4 step (wipe slider, zoom, pan, hold to show
  "on"); a Copy diagnostics entry in the report-a-problem menu; the last input and output folders
  are remembered; a note under the light frames field when the camera is not an Origin.
• Star removal (--remove-stars-method nn, experimental): refills the cores left at the
  brightest unsaturated stars (bright-star leftover 0.20 -> 0.02 on Orion), and handles mono
  stacks (Las Cumbres M51 kept 0.92 of its faint structure, was 0.59).

Fixed

• Local contrast no longer leaves a dark ring round saturated stars.
• Stars within about 30 px of the frame edge now get the same star-aware restore as the rest.

Changed

• Heavy scipy and astropy modules load on first use, and CI gates clippy and a new lint rule.
• Python 3.12 or newer is the supported version (the installers ship 3.12, CI tests 3.12).
• astro_native 0.63.2: constants only, no behaviour change.