Fidelity

Each Word feature below is rated by how faithfully the current release of stetdoc keeps it, shows it and writes it. Four words cover every case:

• Exact. Kept byte for byte, or written the way Word writes it, with nothing changed.
• Normalized. Kept or written in an equivalent form that Word reads the same way. Each case below says what changes.
• Read-only. Shown to the agent, but an edit to it is refused (readonly_region), never silently dropped.
• Omitted. Not shown, not compared or not supported in this release. The package still keeps it exactly.

Two rules hold everywhere. An unedited round trip returns the input bytes, and an edited one changes only the elements that the edit touched (law L1 and L2). Anything stetdoc does not model is preserved exactly, because it is never rewritten.

By feature

An empty cell means the feature takes no part in that column.

Settled details

Each of these was decided in a design, with the evidence there.

• Dates are the caller's wall clock labelled Z, with no dateUtc. A revision's or a comment's w:date is the local time that the caller passes (the CLI's clock in the zone TZ names, or the user's zone in a host application), written as YYYY-MM-DDTHH:MM:SSZ, which is Word's own convention. No w16cex:dateUtc is written. See revisions 5.3 (on GitHub) and comments and notes (on GitHub).
• A tracked replacement is a cleaned word-level diff, not the character-minimal one. Dense edits become one replacement, deletion before insertion, as a careful reviewer would write it. See revisions 2.5 (on GitHub).
• Moves are written as a deletion plus an insertion, in edits and in compare. Writing true moves is the next release.
• A deleted paragraph that carries a section break keeps its mark and the break, so accepting leaves an empty paragraph that still ends the section. Word drops the break with the mark. See revisions 3.2 (on GitHub).
• A section break added with set is untracked only, because a w:pPrChange cannot hold it. A tracked attempt is invalid_payload. See document and ops 2.1 (on GitHub).
• A tracked deletion of an equation that already holds a tracked deletion is refused, because Word writes a form for it that the schema does not allow, and stetdoc never writes a nested deletion. See revisions 1.3 (on GitHub).
• A merge into a paragraph that begins a field spanning paragraphs is refused (readonly_region): the paragraphs a table of contents or another such field spans are one read-only block, and a merge would move the other paragraph's text into it. A merge out of such a field was already refused. See revisions 3.2 (on GitHub).
• Number formats that cannot show a value fall back to decimal, with a Note: a Roman numeral above 3,999, and the script formats (CJK, Hindi, Thai, Hebrew, Arabic, Cyrillic, bahtText, dollarText) until they are built. See styles and numbering 5 (on GitHub).
• Tinted and shaded theme colours are approximated within 2 in any channel of Word's value (89% exact on the curated Word documents). See styles and numbering 3.5 (on GitHub).
• Pattern shading is not read for hidden text. Only w:fill is read, so white text on a shading pattern whose colour is the pattern's own counts as on no background and is flagged, although it is readable. See styles and numbering 3.6 (on GitHub).
• Compare does not compare table, row, cell and section properties, style and numbering definitions, field codes, or the original's comments. See compare 8.3 (on GitHub).
• Compare excludes equations, charts and SmartArt in deleted content, because their changed form cannot be written the way Word writes it. See compare 19 (on GitHub).
• A header or footer present on one side only is skipped by compare, with a compare_story_unmatched warning in the receipt. See compare 4.7 (on GitHub).
• White space at the ends of a w:t without xml:space="preserve" is text, as Word shows it (checked in Word for the web, hand-check file 37): the views and the offsets count it, and what stetdoc writes or copies with such white space says preserve unless an ancestor does. LibreOffice drops it, so a document that relies on it reads one way in Word and stetdoc and another in LibreOffice; check flags such text (edge_space_without_preserve). See wordml 4.7 (on GitHub).
• A table placed directly inside a paragraph (w:p/w:tbl, which the schema does not allow) is left out of the views, and edits cannot reach it. Edits of the paragraph around it and of the blocks beside it keep it byte for byte.
• The text of a vertically merged continuation cell is flagged hidden. A cell with w:vMerge and no restart joins the merged cell that the cell above starts (ECMA-376 Part 1, 17.4.85, and ST_Merge: "all following cells which have a value of continue shall be merged into this merged cell group"), and the group is drawn as one cell with the first cell's content, so Word does not show the continuation cell's own text. The view keeps it, flagged hidden, so that an agent sees what the file holds.
• A text edit never deletes a note's own number (w:footnoteRef, w:endnoteRef): deleting a note's whole text keeps the number, as Word does. A note goes only when its reference in the body goes: the Final view of the notes leaves out a note whose reference is a pending deletion, and the Original view one whose reference is a pending insertion. A note's id (fn3), which refs and delete-note use, counts every reference and is the same in every view; the two views add the number Word shows as its mark once their changes are resolved, when that is another ({footnote 3 mark=2}, "mark").
• Quotes typed into a paragraph follow its convention, curly or straight, and the receipt says so. --quotes as-typed keeps them as typed. See document and ops 2.8 (on GitHub).

Never

Pagination, page numbers and layout; the legacy binary .doc format; executing macros.