v6.6.0

• A paragraph states its own font, so a run that names none of its own is read
  at that font instead of at nothing. Slides whose text was there but invisible
  now show it.
• An empty paragraph keeps the height its font implies, whether or not the file
  names a size for it, and copying across one yields a blank line.
• A sheet's cells read in the font the file names, falling back to the sheet's
  own only where it names none.
• An image stays inside its frame on a page read without the shipped stylesheet,
  rather than covering the whole page.
• A frame that names a side instead of an offset sits on that side, so a centred
  image in an odt is centred. The side it names is GraphicStyle's new
  horizontal_position, carried by the python, java and objc bindings.
• A pdf that comments its content stream renders: % to the end of the line is
  white space wherever it stands, not the start of a token.
• Text that decodes to half a surrogate pair costs that character a replacement
  mark instead of the whole document.
• A pdf's text sits where the file puts it: a line was placed against the
  browser's default font rather than its own, which dropped small text by
  several points.
• A pdf's cmyk colours are converted as Adobe converts them, so a process cyan
  reads as one instead of as pure #00ffff.
• A pdf page is the size of its crop box and shows what is on it and no more,
  as a viewer shows it.
• A pdf's JPEG 2000 images render. New dependency: openjpeg.
• A frame anchored to the page sits where the page says, whatever the text
  around it wraps like - a letter's address and date boxes land in their fields
  instead of in the running text.
• A text document is laid out on the master page it names, so a letter template
  keeps the margins that leave room for its letterhead.
• A docx or odt that chains its styles deeply opens instead of taking the
  process down with it: the w:basedOn / style:parent-style-name chain is
  walked onto a stack rather than recursed, so its length costs no stack.
• A document's xml parts are read once instead of buffered twice on the way into
  the parser, which lowers the memory opening a large one takes.

---

Fixes

• Put a frame where its style says, and keep its image inside (#684)
• Give a paragraph its own font, and an empty one a line break (#687)
• Read a comment as white space, and half a pair as one mark (#688)
• Place text by its own font, decode jpeg 2000, and clip to the crop box (#689)
• Put a page-anchored frame on the page, and lay a document out on the master page it names (#690)
• Walk a style's parent chain instead of recursing it, and read an xml part once (#691)

Refactoring

• Name the boxes a frame anchors against, not every element (#685)
• Aim the sheet's cell styling at the sheet's own cells (#686)

Publishing v6.6.0

• ✅ conan — success · conan package (conan.yml)
• ✅ maven — success · app.opendocument:odr-core-java — Maven Central + GitHub Packages
• ✅ android — success · app.opendocument:odr-core-android — Maven Central + GitHub Packages
• ✅ python — success · pyodr wheels — PyPI
• ✅ apple — success · OdrCoreObjC.xcframework — the release asset and its manifest
• ✅ wasm — success · @opendocument/odr-core — npm