initial commit

2026-09-21 13:41:40 +09:00
commit 855c7328df
411 changed files with 85352 additions and 0 deletions
@@ -0,0 +1,73 @@
+# Extended synthetic PDF corpus
+
+These five original documents contain nine pages. Their drawing source and
+explicit expected geometry are in `../extended_generator.py` and `manifest.json`.
+They are reproducible test fixtures, not approved golden images.
+
+| Document | Pages | Covered content |
+| --- | ---: | --- |
+| japanese-layout.pdf | 2 | Embedded Japanese TrueType subset, horizontal text, explicit ruby coordinates, vertical glyph positions, text matrices at 0/90/180/270 degrees |
+| graphics-compositing.pdf | 1 | Original RGB and RGBA images, soft mask, 50% vector transparency and overlap, circular clipping, axial shading, CMYK swatches |
+| page-geometry.pdf | 4 | Different MediaBox dimensions, offset CropBox, Rotate values 0/90/270/180, four known color markers per page |
+| unembedded-font.pdf | 1 | Standard-14 Helvetica, Times-Roman and Courier substitution baseline |
+| icc-linear-rgb.pdf | 1 | Embedded linear-sRGB matrix/TRC profile in ICCBased vector and raster colors, plus identical DeviceRGB control values |
+
+The Japanese text uses BIZ UDPGothic Regular. The font is subset-embedded, and
+its SIL Open Font License 1.1 is included in `FONT-OFL.txt`. The font file hash
+and upstream license URL are recorded in the manifest. The full font file is
+not redistributed here. The license permits embedding and redistribution;
+these documents do not modify the original font's reserved name.
+
+Generate with Python containing ReportLab, pypdf and Pillow:
+
+```sh
+python3 tests/fixtures/pdf/extended_generator.py
+```
+
+The default font location is
+`/usr/local/share/fonts/BIZ_UDPGothic/BIZUDPGothic-Regular.ttf`.
+Set `DOCVIEW_JAPANESE_FONT` to another installation of that font with its adjacent
+`OFL.txt` when necessary. A changed font file changes the fixture hashes.
+
+After building DocView's `pdf-worker-evidence` target and installing Poppler command-line tools:
+
+```sh
+python3 tests/compare_pdf_extended.py
+```
+
+The comparison generates full-resolution PNGs, difference images, contact sheets,
+text extraction and a machine-readable report under `tests/results/pdf-extended`.
+It assembles pages from the production sandboxed PDF worker's 512-pixel tiles, using FontBroker for unembedded fonts, and
+compares them with independently rendered Poppler pages at 144 dpi. CropBox and
+Rotate are honored. Deterministic size, marker, alpha and Japanese text/search
+assertion failures produce a nonzero exit status; image-difference metrics do
+not have an invented acceptance threshold.
+
+The script defaults to system `/usr/bin` Poppler tools when present. Override
+with `DOCVIEW_PDFTOPPM`, `DOCVIEW_PDFTOTEXT` and `DOCVIEW_PDFFONTS`, or choose a
+different build directory with `DOCVIEW_BUILD_DIR`.
+
+Explicit limitations: vertical text uses individual positioned glyphs, not native
+vertical CID metrics or OpenType shaping; ruby positions are authored directly;
+CMYK has no calibrated ICC golden; unembedded font coverage is Standard-14 Latin
+only. This corpus does not establish arbitrary-document, Windows, accessibility,
+print-production or full PDF compatibility.
+
+## ICC profile and additional scale checks
+
+`../icc_generator.py` creates `linear-srgb.icc` from D65 white, sRGB primaries
+and gamma 1 through public LittleCMS APIs. It uses no downloaded or borrowed
+profile. The generated copyright tag reads `No copyright, use freely`; the
+LittleCMS MIT license is retained in `LCMS2-LICENSE.txt`. The manifest records
+the numeric parameters, profile hash, library version and upstream source.
+The profile's creation timestamp is fixed, and its optional ID is zero, for
+reproducibility. The same profile bytes are embedded for vector and image colors.
+
+Run `python3 tests/fixtures/pdf/icc_generator.py` to add/rebuild only this fixture
+without requiring the Japanese font. The full generator includes it as well.
+`python3 tests/compare_pdf_icc.py` compares actual worker scales 0.5, 1 and 4
+against Poppler at 36/72/288 dpi. It checks 54 analytic color probes, page
+dimensions and original SHA-256. Linear 0.5 gray must differ materially from
+DeviceRGB 0.5; this prevents an ignored ICC profile from passing. Evidence is
+in [pdf-icc](../../../results/pdf-icc/README.md). This one RGB profile does not
+establish arbitrary ICC LUT, CMYK press calibration or print color proofing.