Corpus v1.0.0

Corpus 1.0.0

[1.0.0] - unreleased

First release aimed at reproducibility and reviewability rather than new
measurement features.

Added

• corpus package — the scientific core, importable and testable without
  Electron: calibration, segmentation, measurement, metrology,
  review, io, validation, dev.
• Test suite — 364 tests over synthetic fixtures, with no dependency on
  private TEM images. Covers calibration, particle measurement, core-shell
  metrology, watershed, filters, metadata schema, COCO export, the dataset
  pipeline, the segmentation backend contract, and end-to-end reproducibility.
• Run provenance — measurements.json now records corpus_version,
  segmentation_backend, and a run_fingerprint binding the image checksum to
  the measurement settings. Two runs with the same fingerprint are asserted to
  produce identical measurements.
• Demo dataset — Examples/demo_dataset/: six CC0 synthetic TEM phantoms
  with exact ground truth, COCO masks, provenance metadata, and versioned
  expected results for regression testing. Regenerates byte-for-byte from a
  fixed seed.
• Validation harness — python -m corpus.validation compares Corpus
  against a manual/ImageJ reference in one command, emitting CSV, JSON,
  Markdown and publication-ready figures with MAE, RMSE, bias, relative error,
  R² about the identity line, Pearson r, Bland–Altman limits, and explicit
  recall/precision. Uncomparable rows are counted and listed, never dropped.
• Synthetic self-check — a validation run against the demo phantoms'
  ground truth, committed under docs/validation/synthetic_selfcheck/.
• Segmentation backend contract — SegmentationBackend /
  SegmentationResult with ClassicalBackend and ManualBackend. Every
  measurement records which backend produced it. No ML runtime is imported by
  the classical path, and the tests assert it.
• Continuous integration — GitHub Actions running tests, lint, Python and
  Node syntax checks, demo-dataset integrity, and the smoke test on Linux and
  Windows, plus an Electron build smoke.
• Release automation — tag-triggered workflow building the Windows
  installer/ZIP and the Linux AppImage, with checksums, gated on the test
  suite and on the tag matching package.json.
• Canonical smoke test — python -m corpus.dev.smoke exercises
  measurement, demo integrity, the demo baseline, the dataset pipeline, split
  determinism, validation, and the backend contract.
• Attribution — AUTHORS.md distinguishing original author, current
  maintainer, third-party components, the derivative-work position, and data
  provenance.

Changed

• Repository metadata — package.json description, homepage, repository
  URL, bug tracker, keywords and maintainer now point at RxWhizz/Corpus. All
  obsolete migration references and hard-coded local development paths are gone.
• README — rewritten around what a new visitor needs in the first minute,
  with the classical and AI subsystems clearly separated and limitations stated
  plainly.
• measurement_modes.py — reduced from ~1070 to ~520 lines by moving the
  scientific logic into corpus/. It keeps the same CLI contract and
  re-exports every name it previously defined, so existing imports still work.
  Verified byte-identical output across all four presets before and after.
• Dataset manifest — now records file_sha256, annotation_review,
  skipped_review_labels and calibration_state per image.
• Dataset audit — missing licence, source and checksum are surfaced for
  every image rather than only for the public demo layer; a missing manifest is
  now an error; a public-demo row without an accepted licence now fails instead
  of warning.
• measurements.json — written with sorted keys, so repeated runs produce
  a byte-identical file.

Fixed

• Scale-bar detection crashed on OpenCV 5. cv2.HoughLinesP changed its
  return shape from (N, 1, 4) to (N, 4); Corpus now handles both. Before
  this fix, automatic calibration raised an unhandled IndexError on any image
  where the Hough detector found a line.
• Synthetic dataset generation crashed on small canvases. A fine nm/px on a
  small frame produced a particle larger than the image and raised an opaque
  low >= high from the placement draw. Particle size is now clamped to the
  frame, and a genuinely unusable canvas raises a clear message.

Verifying downloads

90c9f838d81cb1072100225566b2dd2113901172412cce4a8af34a8f1298dca1  ./Corpus-1.0.0-linux-x86_64.AppImage
54b619972dd1c3de407eabb7f0770f2f6175b49714780db60c1dd18f37e69359 *./Corpus-1.0.0-Setup.exe
ceb72c6f672372446219cc9b35138b85ee9aead774d8f5c63b83f813f57ef663 *./Corpus-1.0.0-win-x64.zip