AMORPH 1.0.0

The first release, with executables and a graphical interface for Windows,
Linux and macOS. Since the
AMORPH of the paper (Rowe and Brewer, 2018), the fixes of two bugs of the
sampler change the results of every analysis, and those of the
post-processing change some; see "Results change".
The pull requests are those of
github.com/ij3tt/amorph.

Results change

• The Student-t likelihood is normalised: it had log(nu/2) where
  lgamma(nu/2) belongs and compared r² with the variance instead of
  ν times the variance, which pushed ν of the noise to the top of its prior
  and made the noise Gaussian in practice. Its scale is now
  sqrt(sigma0² + sigma1*model), as in the paper; σ₀ and σ₁ came out about
  30 times too large (#2).
• The outer knots of the background are at the ends of the data. They were
  at x = 0, so above the upper control point the background was
  extrapolated towards x = 0 and could become negative (#2).
• Posterior samples without a wide component count as 0% amorphous.
  display.py dropped them, so the fraction came out too high for nearly
  crystalline materials, and differed from that of export_pgf.py (#3).
• The amorphous fraction and the shape of the wide component are integrals
  with trapezoidal weights, right on a non-uniform grid too (#9).
• The resampling of the post-processing is seeded with seed of
  config.yaml (or 0), so the same run gives the same posterior samples
  every time (#9).

Added

• A graphical interface, src/amorph_gui (PySide6, in Russian): the data
  with the range and the background control points dragged on the plot,
  every setting of config.yaml with its checks, runs of several seeds
  with their progress and a live plot, the results with the diagnosis and
  every figure, and the comparison of run archives (#13). Packaged for
  Windows, Linux and macOS with the sampler inside, ready to run without
  Python (#15).
• A faster sampler, with the same model: each peak is evaluated only within
  20 widths of its centre, and the curves are updated by the peaks that
  changed instead of computed anew. In the validation it saved 1.6 to 4.3
  times as many samples per minute (#14).
• Version numbers: AMORPH prints its version in the first line of its
  output, and run archives record it.
• Checks of config.yaml and of the data file before the sampler starts,
  with messages that name the file, the line and the setting, and exit
  status 1. A data file may have comments, blank lines, a byte order mark,
  CR LF line ends and commas. AMORPH.exe started by a double click waits
  for Enter after an error, so that it can be read. A warning when the
  control points lie outside the data (#4).
• inference_assumptions.narrow_peak_nu in config.yaml, the bounds of
  the prior on the shape ν of the narrow peaks, and seed (#5).
• scripts/diagnose_posterior.py: whether the amorphous fraction of a run
  is well determined, per band of the fraction; --range LOW HIGH also
  gives the fraction within a range of x (#5, #10).
• The report of showresults.py: log(Z), the information and the effective
  sample size, the fraction and the shape of the wide component, the
  diagnosis, and 11 figures (#11).
• Run archives (.amorph): the results of a run, its inputs and an HTML
  report with the figures in one zip file of a few megabytes, written by
  showresults.py; amorph_archive.py writes, reads and shows them. The
  results as data in amorph_results.py, for programs (#11).
• Datasets: G30RAW_amorph.txt and N3RAW_amorph.txt (26–46°), two
  samples whose amorphous fraction is ambiguous (#5), and their whole scans,
  5–70°, with the parts 20–50° and 15–70° (#10).
• README: the shape of the crystalline peaks, the recommended settings, what
  the calibration standards show and how to report results (#5–#9), and a
  Russian translation, README.ru.md (#8).

Fixed

• display.py crashed with NumPy 2 when a posterior sample had no wide
  component (#1).
• The post-processing reads the data file by the rules of the sampler; it
  stopped at a blank line and failed on a byte order mark or commas (#9).
• The histogram of the number of narrow peaks lost the last two counts (#9).
• export_pgf.py reported 0% when no sample had either component; the
  fraction is undefined then, and it stops with an error (#9).
• A missing or empty data file, a missing key or a wrong type in
  config.yaml crashed AMORPH or ran it without data (#4).

Build and CI

• CMake builds AMORPH with DNest4 and yaml-cpp 0.9.0 vendored in
  external/, so the build needs no network access (#1).
• CI on GitHub and GitVerse with shared scripts in scripts/ci/: the Linux
  build, the unit tests of the model, the input checks and the Python
  post-processing, and a smoke test; AMORPH.exe cross-compiled with
  mingw-w64, without MinGW DLLs, and run on Windows (#1–#4); a macOS build
  on Apple Silicon (#11).
• On-demand workflows: validate.yml compares two versions of AMORPH on the
  bundled datasets (#1, #2), experiment.yml runs settings on several seeds
  in parallel (#5–#7, #10).
• Releases: release.yml builds, tests and packs the three systems on a
  pushed tag v*, with SHA-256 checksums, and the GUI packaged with
  PyInstaller on each system, whose self-test runs on the package (#15). The Linux executable needs glibc
  2.34 and the libstdc++ of GCC 13 or newer, the macOS one macOS 11 or
  newer. The actions of the GitHub workflows run on Node 24.