* #1336: Markdown looks up the backtick run closing a code span instead of rescanning the text
Code spans and link brackets searched the rest of the text for the run
of backticks closing a code span. Every run nothing closes (one per
length) cost a scan to the end, and the bracket scan repeated those
searches for each `[` over them, outside its look-ahead budget: 2 MB of
such runs took 2-18 s. The runs are now indexed by length once per text
and found with a binary search; what reads as a code span or a link is
unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* #1452: Markdown brackets no longer turn later links in the paragraph into plain text
The link and image bracket scans share a look-ahead budget per text, so
hostile input stays linear, but it was charged more than they read: a
`[` nothing closes took the whole budget, and every bracket pair and
every link took the rest of the text, as if a destination had been read
to the end. An interval like `[0, 1)`, a dozen references in a long
paragraph or ten links were enough to make every later link plain text.
A scan now pays for the bytes it reads and records the pairs it passes,
and the brackets still open at the end of the text, so later brackets
are looked up. A destination is read, and paid for, only when `(`
follows the `]`; one that parses costs its own length.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
The mime_of helper is missing SVG (most common web vector image) and
ICO formats, causing SVG data URIs and HTML img references to be marked
as generic application/octet-stream. Add image/svg+xml and image/x-icon.
Replace from_utf8(h).unwrap_or("zz") with .ok() and let_chaining.
The fallback was unreachable because h is always valid UTF-8 (it comes
from slicing a valid UTF-8 input string). The magic string "zz" made
the intent unclear to readers.
* Harden against hostile-document DoS and non-web link export
A security pass found several small files that make the app allocate or
compute without bound when opened (GUI, wordcraft-cli, or the MCP server),
and link targets leaving on export without the scheme check the HTML and
Markdown writers already apply. None lets a document run code or read files;
all are denial-of-service or credential-leak-on-click. Each fix bounds an
attacker-controlled size and lands with a regression test, per the
never-crash standard.
- layout: a table repeats its header rows on every page it spills onto, so
many tall header rows paginate to O(headers x pages) placed items. Cap the
total header-row repeats per table (lib.rs). 2 KB docx: 1.1 GB -> 124 MB.
- formats/model, html: an HTML table of many <tr> with large rowspan expanded
to rows x 64 covered cells, and `model::table` pre-allocated a full
rows x cols grid only to clear it. Cap rows at MAX_ROWS before
insert_covered, bound its total work, and build the grid template with one
row. 640 KB html: >1.5 GB -> 214 MB (a plain large table is faster too).
- docx/read/freeform: a custom-geometry Bezier segment took any number of
points, making de Casteljau flattening quadratic. Cap points per segment.
3 KB docx: 12 s -> 0.15 s.
- docbin: footnote/header/text-box story tables and the piece table can name
the same backing text many times. Hold the total characters materialised to
twice the WordDocument stream size (a ceiling no real file reaches, since
each char costs >=1 stream byte). 45 KB doc: >1.2 GB -> 14 MB.
- render, pdf, formats/rtf: document pictures were decoded with no dimension
limit, so a tiny file declaring a huge image could allocate before the size
check. Decode through image::Limits with a side cap first.
- doc/link, formats/rtf, odt, latex: route exported hyperlink targets through
is_safe_href (as HTML/Markdown already do), and extend it to reject
protocol-relative (//host) and UNC (\host) references. Keeps file:,
javascript:, ms-msdt: and UNC targets from being written into exported RTF,
ODT, LaTeX and PDF where a reader would follow them (UNC leaks NTLM on
Windows).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* Align picture-decode limits with the bounds already in force
Review of the previous commit: `image::load_from_memory` and
`load_from_memory_with_format` already applied `Limits::default()`, whose
512 MB allocation cap bounded every picture decode. The PDF change had set
`max_alloc = None`, loosening that; restore the default. The render and RTF
side caps (16 384 px) newly rejected wide-but-short pictures that fit in
memory, so render now uses the pixmap's own `u16` side limit (what
`premul_pixmap` enforced after decoding, applied before it instead) and the
RTF re-encode keeps only the explicit allocation cap. Net effect: rejection
moves to header time with no change to the sizes accepted. Also fix a
comment: the header-row repeat cap is per table, not per document.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
* #778: .xls and legacy-encoding recipient lists; filter and sort recipients
- Recipient lists: legacy Excel 97-2003 .xls (BIFF8 reader on the cfb crate already in the
tree); Excel 5.0/95 and password-protected .xls refused with a clear message; a CSV named
.xls reads as text, an HTML one gets an error. .xls added to the picker filters (desktop,
web, built-in browser) and drop handling.
- Text lists (and plain-text import) detect legacy encodings when not UTF-8/UTF-16: GBK/GB18030,
Big5, Shift_JIS, EUC-KR, Windows code pages (encoding_rs + chardetng, MIT/Apache); an
`encoding` param overrides.
- mailings.filterRecipients / mailings.sortRecipients: field, comparison (equal, notEqual,
less, greater, lessOrEqual, greaterOrEqual, blank, notBlank, contains, notContains), value,
And/Or joins; sort keys. Preview, Finish & Merge and labels use only the matching records
in sort order. Edit Recipient List edits every record and keeps the filter.
- Recipient List dialog: Filter and Sort section and a merge-order column; catalog rows;
strings in every catalog; docs.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* i18n: Hungarian and French strings for this change
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Review fixes for #885: total order for recipient sort, Bulgarian strings
- compare_cells now ranks numbers (compared numerically) before text (compared
case-insensitively), so it is a total order; a column mixing 9, 10 and 1a
could make slice::sort_by panic. Regression test sorts mixed columns.
- Bulgarian catalog: the 24 Filter and Sort strings.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* vi: mail merge Filter and Sort strings
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Dmitrii Khlapov <19991457+dkhlapov@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Brandon Thomas <echelon@gmail.com>
* DOCX: merge settings.xml into the original instead of regenerating it (#1034)
A save regenerated word/settings.xml from the model, so every setting the
model doesn't hold was dropped: compatibility options (lost in 18 of 18
real-world Word files), rsids, document variables, proofing state and the
colour scheme mapping.
The reader now keeps the original part when it holds such settings, and
the writer merges into it: the original root and every unowned child stay
byte for byte; owned children follow the model in CT_Settings order; in
w:compat only compatibilityMode is replaced. The privacy options are the
model's since #1062, so they are owned too. Kept children that point at a
relationship (a mail-merge data source) are dropped. Anything unexpected
falls back to the generated part. Files WordCraft wrote itself keep
nothing extra.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* DOCX: carry unmodelled parts and keep comment ids, style defaults, latent styles, list ids and app props (#1034)
Builds on the settings.xml merge.
- Parts the package or main document relate to that WordCraft neither
models nor writes (webSettings, stylesWithEffects, people, add-in task
panes, the thumbnail, future parts) are carried verbatim with their
relationships and content types; relationship types the writer
produces, and any the body refers to by id, are not (HANDLED_RELS).
- commentsIds / commentsExtensible: durable ids and UTC dates are read
into the model and regenerated against the written paragraph ids.
- w:default on localized default styles, w:latentStyles, lists'
nsid/tmpl/multiLevelType, and app.xml Template/Company/Manager/
HyperlinkBase round-trip; counts are not carried, and Company/Manager
are dropped when the file asks to remove personal information.
- Comment reference marks use the document's "annotation reference"
style id instead of a hard-coded CommentReference.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: przemek-arch <263159043+przemek-arch@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Brandon Thomas <echelon@gmail.com>
ODT: every frame inflated its picture entry again, and the byte budget
was checked only after inflating, so frames pointing at an entry that
didn't fit (or at any entry once the budget was used up) inflated it
again each time. Each entry is now inflated at most once, and only when
the size it declares fits in what is left; the bytes inflated and every
later copy are charged, so an entry that inflates past its declared
size uses the budget up and nothing is inflated after it.
.doc: inline pictures remembered only successes, so a picture that
couldn't be read or didn't fit was located, decoded and copied again
for every anchor on it, and an inline shape's BLIP was decoded against
the whole media cap rather than what was left. Failures are remembered
too, bytes are copied and decoded only within what is left of the cap,
and nothing is read once it is used up.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* ODT: read and write tracked changes (gap 19)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
# Conflicts:
# ROADMAP.md
# crates/engine/src/tests.rs
# crates/formats/src/model.rs
# crates/formats/src/odt.rs
# docs/file-format-parity.md
# docs/gaps.md
* ODT tracked changes: deleted paragraph marks don't take the body's insertion; one region per deletion
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Dmitrii Khlapov <19991457+dkhlapov@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* #1381: WIP
* #1381: HTML import: bold <th>, cell text-align, region paper for HTML/MD/TXT
Header cells are bold unless their own CSS says otherwise. text-align on a
td/th, or inherited from its tr/row group/table, aligns the cell's
paragraphs; a paragraph's own alignment wins, and a table's align attribute
(which places the table) no longer counts. Files that give no page size
(HTML, Markdown, plain text) open on the region's paper, as a new blank
document does: file.open uses the session's region, wordcraft-cli reads the
system region the way the app does (Region::from_locale_vars, shared with
ui-egui). Tables size to that page.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* #1381: the CLI uses the system region everywhere; help mentions WORDCRAFT_REGION
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Dmitrii Khlapov <19991457+dkhlapov@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Brandon Thomas <echelon@gmail.com>
* #816: copied equations reach other programs as MathML
A copy or cut whose selection holds an equation now also puts HTML on the
system clipboard, with each equation as Presentation MathML (wrapped in
<semantics> with the linear format as an application/x-unicodemath
annotation), so Word, LibreOffice, Google Docs and browsers paste a
formatted equation instead of `y=2x_s`.
- formats: new mathml module, a writer from the equation tree (runs,
fractions, scripts and pre-scripts, radicals, n-ary operators, delimiters,
functions, limits, accents, bars, border boxes, group characters, equation
arrays, matrices, phantoms) and a bounded reader (our annotation, else the
presentation elements, else a TeX annotation). HTML export writes
equations as MathML; HTML import (open and paste) reads <math> back.
- engine: edit.copy/edit.cut (and column copy) return `html` and keep
Session::clipboard_html when the fragment holds an equation.
- ui-egui: Services::set_clipboard_html; copy sites put HTML with the text
through it, falling back to egui's text copy.
- wordcraft: arboard set().html() with the text alternative; on Wayland
without data-control the text-only path is kept.
OMML on the clipboard (Word's "Office Open XML") needs custom clipboard
formats arboard can't set; left as a follow-up.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* #816: honest clipboard docs; MathML reader fixes (token cap, function names, column vectors)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Dmitrii Khlapov <19991457+dkhlapov@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Brandon Thomas <echelon@gmail.com>
* Text boxes: VML text boxes keep their internal margins, fit to their text and alignment
A VML text box's inset ("l,t,r,b", ECMA-376 Part 4 §19.1.2.22) was ignored,
so its text sat Word's default 7.2 pt in and 3.6 pt down even where the file
says 0,0,0,0, which is how letterheads and form headers place address blocks
exactly. It's now read like DrawingML's lIns/tIns/rIns/bIns: each value left
out keeps the default, a bare number is in EMUs, o:insetmode="auto" leaves
the defaults. mso-fit-shape-to-text (VML's "Resize shape to fit text") is read
as auto-fit, like a:spAutoFit, and the shape's v-text-anchor as the text's
vertical alignment. In a group, a member's margins go with that member's
text.
The reader feeds layout, so the editor, PDF and page images all change.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Drawings: hidden shapes and pictures are kept but not shown
A drawing marked hidden (VML visibility:hidden, floating or inline;
DrawingML wp:docPr/@hidden, or a group member's cNvPr/@hidden) was drawn,
took room and wrapped text. Word for Windows shows it in no view; it's now
kept and saved hidden, but left out of layout: not placed, no wrap area, no
room in its line, nothing drawn, its text box's notes not placed. Exports to
other formats and copies leave it out, the accessibility checker no longer
asks for its alt text, and the selection pane says it's hidden.
Word for Mac draws hidden VML shapes; Windows follows the markup.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Naxvog <4963183+naxvog@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Brandon Thomas <echelon@gmail.com>
* Open RTF in its code page, and Word/RTF files whose extension is wrong
RTF import read every 8-bit byte as Windows-1252, so Cyrillic (and any
other non-Western) RTF came out as mojibake. Bytes now decode in the
current font's code page (\fcharset, \cpg), else the default font's,
else the document's (\ansicpg), for the single-byte Windows code pages
874 and 1250-1258.
A .doc, .docx or .rtf whose content is another of these formats now opens
by its signature, as Word does: RTF saved as .doc (hh.ru résumés and
accounting exports are commonly saved that way), a .docx renamed .doc, a
Word 97-2003 file renamed .docx.
* RTF: read the page size and margins, lay tables out on their cell edges
RTF documents opened on US Letter with 1" margins whatever they said, and
every table got equal columns and a full grid. Russian accounting and HR
exports (A4, 2 cm margins, borderless layout tables whose rows have
different \cellx edges) came out about twice as long as in Word.
- \paperw/\paperh/\margl…\margb (and the …sxn section forms), \headery,
\footery and \landscape set the document's section; out-of-range values
keep Word's defaults. Tables size to that page.
- A table's column grid is the union of its rows' \cellx edges; each cell
spans the grid columns between its left and right edges, as Word does.
- A table with no left/right cell borders is laid out without the grid
style (\clbrdrb rules alone don't make it ruled; \brdrnone is no border).
On 44 RTF files (mostly .doc-named exports) checked against Word's page
counts: total page-count difference 198 -> 30, exact matches 3 -> 17, none
worse.
* RTF: paragraph spacing and indents, RTF defaults, the default font
- \sb, \sa, \sl/\slmult (multiple, at least, exactly), \li/\lin, \ri/\rin
and \fi are read onto each paragraph (new model::Spacing); \pard resets
them.
- The document's paragraph defaults are RTF's (single spacing, no space
around paragraphs) instead of a new document's 1.15 and 8 pt after.
- Text that never selects a font is in the default font (\deff, font 0
when the file names none), as Word shows it.
Against Word's page counts on 251 real documents (44 RTF, mostly .doc-named,
plus .docx): exact matches 212 -> 216, within 10% 248 -> 249; a 68-page
Russian federal law RTF goes from 89 to 73 pages (the rest is the Word
installation's own default margins, which the file doesn't name).
* RTF: import headers and footers
\header, \headerr, \headerl, \headerf and the \footer… forms were skipped,
so letterheads, logos and "page N" lines were lost. Each group is now read
into its own output (the body's paragraph, table and row state is set aside
and put back when the group closes), built into the document with its
pictures (new model::part_blocks), and attached to the section as the
default, first-page or even-page header or footer; \titlepg sets the
separate first page. A header left open by a truncated file still ends.
40 of 41 RTF files in the corpus have headers or footers; against Word's
page counts exact matches go 216 -> 218 of 251, none worse.
* RTF: cell and paragraph borders
Tables either got a full grid (any left/right cell border) or none, and
paragraph borders were kept only on empty paragraphs (as a rule), so the
rules under letterheads and section titles were lost.
- Cells: \clbrdrt/l/b/r with their line style (\brdrs, \brdrdb, \brdrdot,
\brdrdash, \brdrth, \brdrtriple, \brdrwavy…, \brdrnone), \brdrw width and
\brdrcf colour become the cell's own borders (model::Cell::borders); RTF
tables get no grid style, so only the borders the file names are drawn,
as in Word.
- Paragraphs: \brdrt/b/l/r likewise become the paragraph's borders
(model::Para::borders), reset by \pard; an empty paragraph with only a
bottom border stays a rule.
40 of 41 RTF files in the corpus name cell borders; page counts against
Word are unchanged (218 of 251 exact).
---------
Co-authored-by: Магда М.Г. <mail@mgmagda.ru>
Co-authored-by: Brandon Thomas <echelon@gmail.com>
RTF import read every 8-bit byte as Windows-1252, so Cyrillic (and any
other non-Western) RTF came out as mojibake. Bytes now decode in the
current font's code page (\fcharset, \cpg), else the default font's,
else the document's (\ansicpg), for the single-byte Windows code pages
874 and 1250-1258.
A .doc, .docx or .rtf whose content is another of these formats now opens
by its signature, as Word does: RTF saved as .doc (hh.ru résumés and
accounting exports are commonly saved that way), a .docx renamed .doc, a
Word 97-2003 file renamed .docx.
Co-authored-by: Магда М.Г. <mail@mgmagda.ru>
Co-authored-by: Brandon Thomas <echelon@gmail.com>
Word's symbol fonts are proprietary and never bundled, so text set in them
showed the substitute's letters. New wordcraft_doc::symbol_fonts maps every
code (stored as U+F020-U+F0FF, e.g. w:sym, or as a plain 8-bit code) to its
Unicode equivalent: Symbol from Unicode's Adobe symbol.txt, the Wingdings
and Webdings sets from Unicode L2/12-368 (WG2 N4384), checked against UCD.
Layout shapes the mapped characters (fallback faces draw them); PDF text,
copy as text and TXT export read as the symbols; the stored text and docx
round trips are unchanged.
Co-authored-by: Dmitrii Khlapov <19991457+dkhlapov@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Brandon Thomas <echelon@gmail.com>
* ODT import reads page size, orientation and margins from styles.xml (#810)
* Review fixes for #829: clamp ODT page size and margins
Page width/height from styles.xml are kept to 36..=1584 pt (the DOCX
reader's floor, Word's 22-inch maximum) and each margin is capped at the
page's longer side, like w:pgMar. Adds a hostile-values test and collapses
a nested if for clippy.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Brandon Thomas <echelon@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Keep inherited complex-script size, bold and italic independent from Latin
properties and retain explicit values during DOCX round trips. Preserve
ordinary formatting-command behavior and add regression coverage.
Resolve eight existing Clippy warnings without changing behavior so all
six required CI gates pass.
Co-authored-by: ChatGPT <noreply@openai.com>
Co-authored-by: Brandon Thomas <echelon@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
- `ui.render` and `ui.screenshot` without a `path` return the PNG as base64 `png` in the
reply instead of writing a fixed name into the shared temporary folder, where another
user could have planted a link to one of this user's files; the MCP bridge's
`render_page` and `screenshot` use that and no longer round-trip through
`temp_dir()/wordcraft-mcp-*.png`. A given `path` still works as before (#678).
- One allow-list, `wordcraft_doc::link`: `is_web_or_mail_link` for Ctrl/⌘-click and
`is_safe_href` (web, mail, `#bookmark`, relative path; schemes read the way browsers
do, so `java\nscript:` is a script) for the HTML and Markdown exporters and the HTML
importer. Scripts and other schemes export as plain text (#679).
- Tests: the allow-list, exporters with hostile links, the MCP bridge against a fake app
(no `path` ever asked for), the control channel's in-reply PNGs.
Co-authored-by: petycomppety <30414255+petycomppety@users.noreply.github.com>
Co-authored-by: Brandon Thomas <echelon@gmail.com>
* Content controls: keep, show and edit w:sdt (#381)
Content controls are now part of the model instead of being unwrapped on open:
ControlStart/ControlEnd markers around their content (inline or around whole
paragraphs), with table, row and cell wraps for controls around tables,
repeating-section rows and cells. All sdtPr properties round-trip (title, tag,
id, lock, placeholder, temporary, rich/plain text, w14 check box, combo box,
drop-down list, date, picture, building block gallery, w15 repeating
sections); unknown children are kept verbatim. The writer plans per block list
which controls are written inline and which around blocks so the XML always
nests, and drops markers that lost their partner.
Editing: deletions keep a marker whose partner is outside the range, typing
replaces a shown placeholder, emptying a control brings it back, sdtLocked and
contentLocked refuse with a status message. Developer tab with insert
commands, Design Mode, Properties dialog; frames and title tabs on the page,
check boxes toggle on click, list and calendar pickers.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* #381: refresh catalog count and parity checklist after rebase
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Brandon Thomas <echelon@gmail.com>
Select Recipients › Use an Existing List… now reads .xlsx, .xlsm and .ods
workbooks besides CSV. A new read-only SpreadsheetML / ODF spreadsheet
reader in wordcraft-formats lists the sheets and turns one sheet into rows
of display strings: shared, inline and rich strings, numbers, booleans,
cached formula results, dates and times as ISO 8601 from the cell's number
format, sparse cells, ODS value types, repeated rows/columns and covered
cells. Zip entries, rows, columns, cells, string lengths and repeat counts
are capped, and nothing recurses.
mailings.recipients takes "data" (base64) and "sheet" and "headers"; a
workbook with several sheets and no sheet answers with the sheet names, and
the app shows a Select Table dialog with "First row of data contains column
headers". The desktop picker, dropped files and the web inbox route
spreadsheets to the recipient list.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Brandon Thomas <echelon@gmail.com>
- TXT: the engine opens and saves .txt through wordcraft-formats (BOM/unmarked
UTF-16 sniffing, Windows-1252 fallback; list labels and a tab on export)
instead of its own narrower decoder and body plain text.
- RTF: \uN and \ucN are handled in document info, bookmark names and field
instructions, not only in body text.
- ODT: common paragraph styles (and common parents of automatic styles) give
their text formatting, except the style that sets the paragraph kind;
table-cell-properties fo:background-color becomes cell shading.
- HTML: tables without style or visible borders export without cell borders;
white-space:pre paragraphs keep leading/trailing spaces.
- LaTeX: brace groups restore the paragraph alignment set inside them.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Brandon Thomas <echelon@gmail.com>
`crates/docx/src/xml.rs` parses untrusted document XML through
`quick_xml::reader::NsReader`. In 0.38.4 that is RUSTSEC-2026-0195 (CVSS 7.5): the
namespace resolver pushes a binding for every `xmlns` / `xmlns:*` declaration on a
start tag into an unbounded `Vec`, with no limit on the number of declarations.
None of the reader's existing caps bound that path - MAX_PART (256 MiB), MAX_TOTAL
(512 MiB), MAX_ENTRIES, MAX_DEPTH and MAX_ELEMENTS bound part size, entry count,
nesting and element count, not declarations per tag.
Measured before this change, `wordcraft-cli convert` on a crafted .docx whose
`word/document.xml` is one start tag with 21,845,333 `xmlns:a="u"` declarations:
508 KB on disk -> 262 MB decompressed (under MAX_PART) -> 1.19 GB peak RSS,
exit 0: the document opened and converted without any warning.
At the MAX_PART/MAX_TOTAL ceiling that is roughly 2.4 GB of resident memory from
about a megabyte of download, i.e. an OOM kill with unsaved work open. Nothing
panics, so the `Session::run` guard does not catch it.
0.41.0 caps namespace declarations per start tag at 256
(`DEFAULT_MAX_DECLARATIONS_PER_ELEMENT`, raisable through
`NamespaceResolver::set_max_declarations_per_element`) and returns
`NamespaceError::TooManyDeclarations` beyond it. The same fixture now fails in
about a second with a clear error and no namespace allocation:
wordcraft-cli: malformed XML: start tag declares more than 256 namespace
bindings; raise the limit with NamespaceResolver::set_max_declarations_per_element
Two API adaptations for 0.41 (0.x bumps are allowed to break):
- `resolve_attribute` moved off `NsReader` on to the resolver:
`r.resolve_attribute(key)` -> `r.resolver_mut().resolve_attribute(key)`.
- `Attribute::unescape_value` is deprecated in favour of `normalized_value`, which
also applies XML attribute-value normalization as the spec requires; attribute
values in a Word part are UTF-8 here (UTF-16 parts are transcoded first).
Regression test: `crates/docx/src/xml.rs::too_many_namespace_declarations_is_error`
asserts a tag with 10,000 declarations is an error. It fails on 0.38.4 and passes
on 0.41.0.
This also clears the second quick-xml advisory, RUSTSEC-2026-0194
(quadratic duplicate-attribute scan). That one was not reachable here - the reader
already passes `with_checks(false)` at xml.rs:257, and a 60,000-attribute tag parses
in the same 65 ms as a 20,000-attribute one - but `cargo audit` is clean afterwards.
Verified: `cargo xtask ci`; `cargo audit` (no vulnerabilities left, the two
unmaintained-crate warnings are `krilla`'s `rustybuzz`/`ttf-parser` and are
unrelated); 32 docx tests plus the malformed-input suite; and 21 crafted hostile
documents through `wordcraft-cli convert` with zero panics before and after.
One residual, unchanged by this commit: the fixture above still costs about 509 MB,
because MAX_PART lets one part decompress to 256 MiB and every part is held in memory
while the package is read. That is the existing, intended bound; what this commit
removes is the namespace allocation on top of it.
* Keep charts, SmartArt and OLE objects when saving DOCX (#319)
Charts and SmartArt diagrams were drawn since #292 but the writer skipped
them, and OLE objects kept only their picture, so opening and saving a
document silently dropped them.
The reader now keeps each object's own markup (the a:graphic of a chart or
diagram, the whole w:object of an OLE object) on the model object, and every
part it reaches through relationships (chart part, embedded workbook, colour
and style parts; diagram data, layout, quick style, colours and drawing; OLE
embedding and picture) as opaque bytes in Document::passthrough, with their
content types and relationships in a manifest, as the VBA project already is.
The writer emits the object with fresh relationship ids pointing at the
re-added parts, keeping each part's own relationship ids and content types.
A chart or diagram gets its frame from the object's current size and
position; an OLE object gets its new size written into its VML shape. A part
whose name the writer uses itself is renamed, and an object pasted more than
once gets its own copies of every part (sharing past a size budget). Part
counts, relationship counts, markup size and copy bytes are all capped.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Keep charts, SmartArt and OLE objects when saving DOCX (#319)
Charts and SmartArt diagrams were drawn since #292 but the writer skipped
them, and OLE objects kept only their picture, so opening and saving a
document silently dropped them.
The reader now keeps each object's own markup (the a:graphic of a chart or
diagram, the whole w:object of an OLE object) on the model object, and every
part it reaches through relationships (chart part, embedded workbook, colour
and style parts; diagram data, layout, quick style, colours and drawing; OLE
embedding and picture) as opaque bytes in Document::passthrough, with their
content types and relationships in a manifest, as the VBA project already is.
The writer emits the object with fresh relationship ids pointing at the
re-added parts, keeping each part's own relationship ids and content types.
A chart or diagram gets its frame from the object's current size and
position; an OLE object gets its new size written into its VML shape. A part
whose name the writer uses itself is renamed, and an object pasted more than
once gets its own copies of every part (sharing past a size budget). Part
counts, relationship counts, markup size and copy bytes are all capped.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Merge fix: #340's rotation round-trip test builds an Image with #333's ole field
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Brandon Thomas <echelon@gmail.com>
A few hundred KB of `*[`, `, `<` (for `>`) and `&` (for `;`). Bracket and destination scans now
share a per-paragraph look-ahead budget, tags are searched within 4 KB and entity names
within 33 bytes.
- HTML `unescape` searched the whole remaining text for `;` after every stray `&`; it now
looks only at the first 33 bytes.
- The Markdown/HTML-to-document step appended every inline through `insert_text`, which
rescans all earlier runs, so a paragraph with thousands of runs took quadratic time on
its own. It now appends runs directly and normalises once.
Ordinary links, images, tags and entities read as before.
Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>
A chart or SmartArt diagram in a DOCX was dropped on import: it took no
room, so the text after it moved up (five of the six chart documents in
docxide-pdf's fixtures came out a page short), and nothing was drawn.
They are now an InlineObject::Graphic: it takes its room, floats and
wraps like a picture, and draws a list of neutral items (shapes, paths,
one-line texts, pictures) that layout scales to the object's box.
- Charts (read/chart.rs) are drawn from the values the chart part
caches: clustered, stacked and 100% bar and column charts, line, area,
pie, doughnut and scatter charts, with title, legend, gridlines, axis
labels and Excel number and date formats.
- SmartArt (read/diagram.rs) is drawn from the drawing Word stores
beside the diagram: preset shapes, theme colours, text and pictures.
- Theme colour handling is shared (read/drawing_color.rs).
- Each chart or diagram part is read once, only through a relationship
of its type, under a per-file budget of items.
- Tagged PDF tags them as figures with their alternative text, and the
accessibility check reports charts without one.
Not yet: charts and diagrams aren't written back when saving a DOCX
(they weren't before either), so the editor doesn't let you move or
resize them. Radar, stock, 3-D and combination charts draw their first
supported chart type only; SmartArt pictures aren't clipped to their
shape.
docxide-pdf bench (365 documents, vs main + #135): page counts right
320 -> 326 (all five chart documents and one SmartArt document), Jaccard
passes 208 -> 215, SSIM passes 76 -> 79; eight documents better, none
worse.
Co-authored-by: Oliver Zieschang <4963183+naxvog@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Brandon Thomas <echelon@gmail.com>
* M5.5: Word equations (OMML) import, lay out and save like Word
Equations were flattened to plain text on import. Now OMML is read into a
math tree (fractions, scripts, radicals, n-ary operators, delimiters,
functions, limits, accents, bars, boxes, equation arrays, matrices,
phantoms), laid out with the OpenType MATH table (Cambria Math when
installed, else STIX Two Math and other open math fonts), drawn to screen
and PDF, and written back as the original OMML. Display equations sit on
their own centred line. insert.equation parses the linear format into
structure.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* M5.5: equation editor like Word's — Equation tab, in-place editing, # numbers
Insert › Equation now opens an equation for in-place editing, with the
Equation tab: ready-made equations, Unicode/LaTeX input and conversion
(professional ↔ linear), Normal Text, eight symbol sets, and the eleven
structure galleries (fraction, script, radical, integral, large operator,
bracket, function, accent, limit and log, operator, matrix) with tiles
rendered by the math engine. Typing builds up like Word (a/b, x^2, √, ∑…
then space), Math AutoCorrect turns \alpha, \sqrt… into symbols, Tab moves
between empty slots, arrows/Backspace work through structures, and the
right-click menu has structure commands (fraction type, radical degree,
limit position, matrix rows/columns, remove brackets…). A trailing #(n)
numbers a display equation with the number at the right margin; # alone
numbers it automatically. DOCX output keeps numbers Word-compatible.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* i18n: equation editor strings in Spanish, Brazilian Portuguese and Ukrainian
The catalogs that landed on main after #191 was opened lacked the Equation
tab and equation command labels.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Equations: our own names and order for the built-in equations
Clean room: the ready-made equation list used another product's gallery entry
names and order. Same formulas (and ids), with our own descriptive labels,
grouped by subject (geometry, algebra, calculus, trigonometry, probability).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* ROADMAP: equations row matches the landed equation editor
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fmt after merge
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Brandon Thomas <echelon@gmail.com>
* D0+D1: open Word 97-2003 .doc — container, FIB, piece table text (#28)
New wordcraft-docbin crate (L2) reads the OLE2 container via the cfb
crate (MIT) and parses the FIB and piece table from the [MS-DOC]
specification: text pieces decode as UTF-16LE or Windows-1252 (table
moved to wordcraft-doc::encoding), subdocument CP ranges, and Word's
special characters map to the model's (tabs, breaks, hyphens). .doc/.dot
registered in OPEN_EXTS and the I/O dispatch. Encrypted and Word 6/95
files give clear errors; every length is checked and proptest
truncation/flips never panic.
* D2: .doc formatting — stylesheet, fonts, CHPX/PAPX FKPs, sprms (#28)
Sprm decoder with operand-size rules (incl. sprmTDefTable/sprmPChgTabs
exceptions and the Prm0 isprm table), bin tables + ChpxFkp/PapxFkp page
lookup, STSH styles (StdfBase layout validated against real Word files:
bchUpe == cbStd, sti 0 = Normal) and the SttbfFfn font table in both
extended and self-delimiting flavours. The CP walk now splits runs on
formatting changes and resolves paragraph styles: bold/italic/strike/
underline/caps/hidden, colours (Ico palette + COLORREF), half-point
sizes, super/subscript, fonts, highlights; alignment, indents, spacing,
keep options, outline levels. Unknown sprms are skipped leniently;
an env-gated corpus test opens real files (styles + text asserted).
* D3: .doc structure — sections, headers/footers, tables, lists (#28)
- Sections: PlcfSed/Sepx with 16-bit measurement sprms (sprmSXaPage & co),
explicit sprmSBOrientation with a width/height fallback; even/odd/first
header and footer stories from Plcfhdd become document parts.
- Tables: sprmTDefTable grids (operand length fixed to cb+1 per the spec),
horizontal merges (sprmTMerge), vertical merges (sprmTVertMerge), cell
shading (sprmTSetShd / sprmTDefTableShd80), cell widths, valign, row
height/header/cant-split, and table borders (sprmTTableBorders80 Brc80s).
- Lists: PlfLst (LSTF + appended LVLs) and PlfLfo (LFO + LFOLVL overrides)
resolved into the numbering model; sprmPIlvl/sprmPIlfo mark paragraphs.
- Combined fixture (intro + table + list + 2 sections + header) round-trips
through Document JSON; hostile proptests now mutate the structured fixture.
* D4: .doc notes, fields, bookmarks and inline pictures (#28)
- Footnotes/endnotes: PlcffndRef/PlcfEndRef reference CPs paired with the
PlcffndTxt/PlcfEndTxt story ranges; each story walks into a Footnote or
Endnote part and the 0x02 anchor becomes a NoteRef object.
- Fields: 0x13/0x14/0x15 state machine in the walker — instruction text is
hidden, the cached result kept; HYPERLINK fields keep their result as
linked runs, everything else becomes one Field object.
- Bookmarks: SttbfBkmk names + Plcfbkf/Plcfbkl CPs become BookmarkStart/
BookmarkEnd objects at their anchors.
- Pictures: sprmCPicLocation points into the Data stream; the PICF header
gives the display size (dxaGoal/dyaGoal scaled by mx/my) and the payload
is format-sniffed (PNG/JPEG/GIF/BMP; metafiles are dropped with a warn).
- Hostile proptests now truncate/flip a fixture with all of the above.
* D5: .doc import hardening, caps and roadmap (#28)
- Cap inline pictures at 5 000 per document.
- Engine test: .doc/.dot bytes dispatch through the Word 97-2003 reader
(garbage must error, never panic or open empty).
- ROADMAP: Word 97-2003 import row (60% — text, formatting, styles,
sections, headers/footers, tables, lists, notes, fields, bookmarks,
raster pictures).
Real-corpus validation stays opt-in: WORDCRAFT_DOCBIN_FIXTURES=<dir>
cargo test -p wordcraft-docbin corpus_files_open.
* Corpus fixes: compressed-piece FC space, PAPX query point, none-level labels (#28)
Validated against two real files (kept locally under plan/word/fixtures,
gitignored): a 1 MB Lorem-ipsum sample and a 2003 styles/lists/table
sample.
- FKP bin tables and pages address the WordDocument stream in true byte
offsets; a compressed piece stores its base fc doubled (byte = fc/2),
so fc_of_cp now unwinds the doubling. Before, every PAPX/CHPX lookup in
compressed (cp1252) files landed on the wrong entry — tables collapsed
to one row, lists and heading styles were lost.
- The paragraph-properties query now uses the paragraph mark's own FC
(inside the paragraph) instead of the FC after it, which picked up the
next paragraph's FKP entry.
- Fixture papx entries re-keyed to paragraph-start CPs, the real rgfc
convention.
- Counters::next_label returns nothing for NumFormat::None levels (Word's
numFmt none shows no label; also fixes a latent docx-reader case).
- Symbol/Wingdings bullet chars map to Unicode glyphs (• ▪ ○ ●).
With this, both corpus files render headings, styles, nested and numbered
lists, the tables and the PAGE field correctly; metafile-wrapped images
and OLE objects stay dropped by policy (logged).
* docbin: bound the text walk by real text; dedupe and cap pictures (#28)
Hostile CP numbers no longer drive the work:
- header/footnote/endnote story ranges from the Table stream are clamped to their
subdocument (ccpHdd/ccpFtn/ccpEdn) with checked base additions, and every walk stops at
the last piece's CP; a story spanning [0, 4e9) used to push billions of U+FFFD;
- characters are found by binary search over the pieces (PieceTable::char_at), and an
unreadable stretch (gap, truncated or out-of-stream piece) is skipped in one step rather
than one U+FFFD per CP; the old per-character scan was O(chars x pieces);
- the piece CPs are made non-decreasing at parse, so inverted pieces are empty and binary
search stays sound; surrogate pairs now decode;
- note references are looked up by binary search; the style's CharProps is borrowed, not
cloned, per character;
- pictures are interned by fc_pic (5000 anchors on one picture share one media entry and
one copy) and the bytes copied per document are capped at 256 MB.
Each fix has a crafted-bytes regression test; against the previous code the story,
piece-claim, last-piece and many-pieces tests hang (killed at 60 s) and the shared
picture test copies the picture once per anchor (80 GB).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* docbin: checked offset arithmetic in lists, sections and table rows (#28)
- list.rs: PlfLst/PlfLfo/LVL offsets (at + 2 + cLst*28, at + 4 + lfoMac*16, off += len,
...) saturate, which matters on 32-bit targets (wasm32) where fc + n can overflow usize;
the LFO override path no longer indexes abstracts directly.
- sections.rs: the Sepx range (fc_sepx + 2 + cb) is checked; Plcfhdd is read only as far
as the stories a capped section count can use (it used to reserve lcb/4 entries).
- table.rs: sprmTDyaRowHeight = i16::MIN negated in i16 (overflow panic in debug builds);
now -(v as f32).
Regression tests: the row height test failed before (overflow panic with debug assertions, a 0 pt height without); the list test pins the saturating
offsets (it only overflowed on 32-bit targets).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Format the merged OPEN_EXTS list
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Brandon Thomas <echelon@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* HTML import: pictures referenced by a relative path are embedded (#98)
An img whose src was a path next to the HTML file was dropped, leaving
only its alt text; only data: URIs loaded. The HTML importer now takes
an optional image loader (the formats crate stays free of file access
and wasm-safe), and opening an HTML file from disk loads relative
sources from the file's folder. URLs and absolute paths are not
followed, and files that aren't pictures are ignored.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* HTML import: relative pictures stay in the folder, read only regular files, within a budget (#98)
Review follow-up for #113. Pictures named by a relative path are now
loaded by LocalImages:
- Only regular files are read. The path is checked before opening (a
named pipe would block the open), the opened handle is checked again,
and the read goes through take(limit + 1), so the per-picture limit
holds even for files that report a length of 0, like /dev/zero.
- The path is canonicalized and must stay inside the HTML file's
folder, so `..` and symlinks can't reach other files.
- Each file is read once and shared between references, and every
reference counts against a 1 GB budget for the document, so one large
picture referenced many times can't make the import read or copy
without bound. Bytes read from oversized files count too.
The image loader now returns Arc<Vec<u8>>, and make_img takes either a
Vec or an Arc.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: lloyd <lloyd@stackworx.io>
Co-authored-by: Brandon Thomas <echelon@gmail.com>
* LaTeX (.tex) import and export (#7)
Open and save LaTeX source as a document format, next to Markdown and
HTML. Export writes a standalone article (fontenc, ulem, xcolor,
multirow, amsmath, hyperref, loaded only when used): unnumbered sections,
itemize/enumerate, quote, center/flushright, verbatim, tabular with the
document's column widths and merged cells, equations as $..$ / \[..\],
character formatting, colours and links; metadata goes to \hypersetup.
Import reads the everyday subset: preamble \title/\author, sectioning
(\part/\chapter shift levels), paragraphs, lists with labels, quote,
alignment environments, verbatim-like environments, tabular with
\multicolumn/\multirow/\cellcolor and p{} widths, inline and display
math, formatting commands and declarations, colours, links, refs,
citations, footnotes, accents, TeX ligatures and special characters.
Unknown commands keep their arguments' text; layout and definition
commands are dropped with their arguments. Nesting is depth-limited and
input is capped, so hostile files can't overflow the stack.
The flow model gains an Equation inline so equations survive LaTeX round
trips; the other text formats write their linear text as before.
Wired into open/save dispatch, the file dialog filter, Backstage Export,
the web app, CLI/MCP help, docs and the Linux desktop entry.
Fixes#7.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019FczntrDNYd6DupyBdpbL5
* #7: bound LaTeX column specs and equation export (review)
- spec_widths: `*{n}{…}` repeats recurse with a depth argument and stop at
MAX_NEST, so a deeply nested column spec no longer overflows the stack or
exhausts memory. Column counts saturate and widths stop at MAX_COLS.
- Column specs read from the source are capped at MAX_SPEC (4096) chars.
- balanced() looks groups up in a table of closers built in one stack pass,
so linear_to_latex no longer rescans to the end of the string for every
unmatched `(` (quadratic: 50k took seconds). spec_widths uses it too.
- env_name restores the position when `\begin{` / `\end{` has no `}`
within 64 chars, so the `{` is no longer lost.
- Regression tests: the 100k-deep `*{1}{` spec and a 200k-`(` equation in
deep_nesting_is_bounded; unclosed_begin_keeps_its_brace and
spec_widths_repeats_are_bounded.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LyVbT6Vv4BwWZ4hY8973qD
* #7: linear-time labels, brackets and \begin{; no commands from the document in exports
Applies @HaringIbon's review suggestions:
1. Parser::anchor keeps at most MAX_ANCHOR_RUN (64) bookmarks in a row with
no text between them; text() looked back past every one, so a file of
labels imported in quadratic time. Labels separated by text are all kept.
2. linear_to_latex_at checks for the `/` of a fraction before copying the
group out, so deeply nested matched brackets export in linear time.
3. env_name searches only the next 65 bytes for `}`, so a run of unclosed
`\begin{` imports in linear time.
4. An equation's `\name` is written as a command only when it is a known
symbol, function or font/spacing command (MATH_COMMANDS); anything else
becomes `\backslash name`. Two carets are never written side by side in
equations or anchor names, since TeX reads `^^5c` as a backslash. Compiling
an export can no longer `\input` files or `\write18` named by the document.
Regression tests: three cases added to deep_nesting_is_bounded, plus
latex_export_carries_no_commands_from_the_document and
latex_labels_in_a_row_are_read_in_linear_time.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LyVbT6Vv4BwWZ4hY8973qD
* Merge follow-ups: FTable.borderless (from #102) in the LaTeX importer; rustfmt
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Brandon Thomas <echelon@gmail.com>
* Lists: a numbered list after a bulleted list keeps its numbering in ODT, RTF and Markdown (#83)
* Import: a top-level list item of the other kind starts a new list
The writers now split a bulleted list from a following numbered one, but the shared importer merged consecutive list paragraphs back into one list, so ODT and RTF still read the numbered items as bullets.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Brandon Thomas <echelon@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
A w:num can redefine any of its abstract list's levels with a whole
w:lvl inside w:lvlOverride. We kept only the start value from it and
numbered with the abstract list's levels, so documents that rely on it
showed the wrong labels: a construction specification (SpecLink) read
"A. PART 1 GENERAL" and "I. SECTION INCLUDES" where Word shows "PART 1
GENERAL" and "1.01 SECTION INCLUDES". Saving such a document also
dropped the definitions, so Word showed the wrong labels afterwards.
Num::level_overrides keeps them; Numbering::level and the list counters
prefer them over the abstract list's levels; the .docx reader reads them
with the abstract levels' parser and the writer writes them back (one
w:lvlOverride per level with its startOverride and/or lvl).
In docxide-pdf's Word reference corpus 6 of 382 documents carry such
definitions; the specification's SSIM goes 13% -> 41%, the others
restate their abstract lists and are unchanged.
Tests: counters use a list's own levels ("" and "%2.%3" in decimalZero
give "1.01", "1.02", "2.01") and fall back to the abstract list for the
rest; level overrides round-trip through .docx beside start overrides.
Co-authored-by: Oliver Zieschang <4963183+naxvog@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* HTML import: keep body text when </head> is omitted
Close an open head implicitly on body-content start tags, before skipped raw-text handling. Preserve head metadata and title capture while allowing body paragraphs and inline text to import.
Co-Authored-By: Codex <noreply@openai.com>
* RTF import: read pictures inside shppict wrappers
Recognize shppict as a transparent destination so nested pict groups are decoded instead of inheriting Skip. Continue skipping nonshppict legacy duplicates and pictures under other skipped destinations.
Co-Authored-By: Codex <noreply@openai.com>
---------
Co-authored-by: Codex <noreply@openai.com>