The covered text walks every run of a covered line, and a run can draw
no words (`( ) Tj`), so a line now costs its word boxes or its runs,
whichever are more. A word followed by many empty runs can no longer
be walked once per cover outside the 50-million-comparison bound.
Review on #4198 (CodeRabbit).
TextLine::runs also says whether joining put a space before each run,
and covered_words uses it: a word shown in two runs with nothing
between them (`SE` `CRET`) is reported as one word, not `SE CRET`.
Runs with uncovered runs left out between them are still spaced.
The regression test writes its content streams on one line each.
Review on #4198 (CodeRabbit).
A text line joins runs painted on one baseline, and the audit reported
the line's whole text for a cover over part of it, including words
painted beside the cover that stay visible (`SECRETtail` for a cover
over `SECRET`).
TextLine now keeps each joined run's text with the end of its word
boxes (TextLine::runs), and the audit reports the runs whose words
the cover overlaps. A single run, a line reordered for reading
(right-to-left or visual-order Indic), or a cover that overlaps the
line's box but none of its words still reports the whole line.
Fixes#4164
The tests each branch appends to the audit, edit and organize test files are
kept side by side; prune.rs merges cleanly, with the form memo on top of the
budgeted page content.
The redaction audit read each page's text through text_lines, which walked
the whole page tree again for every page (quadratic in the page count) and
let /ActualText replace the glyphs, so an empty or substitute ActualText hid
covered text from the audit. Edit gains glyph_text_lines, which reads the
page dictionary it is given and ignores ActualText; the audit uses it.
The cover walk decoded every /Contents entry and every form draw with
a fresh 1 GiB allowance; the operator bound does not count whitespace,
so a large blank form drawn many times meant hours of decoding. The
walk charges a per-page DecodeBudget with the strict decoder, and
spending it fails the audit.
The audit returned the text under each cover whatever the document's
permissions, and the CLI and the automation tool printed it. A user
whose password withholds copying could read through the audit the text
that text extraction refuses them. The audit now reports every finding
for such a document with empty covered text; the owner password lifts
the restriction as it does elsewhere.
Every cover was compared with every text line of its page, and each
finding kept its own copy of the text it covered. Both grow with the
square of the content: a few megabytes of covers over one long line
meant hours of comparisons and gigabytes of report. The audit now
counts comparisons per page and covered-text bytes across the document
and fails with an error past either bound, never with a partial
report.
A text line joins compatible runs on one baseline, and the audit took
the line's last painting operator as its paint position. Text drawn
after a cover on the same baseline joined the covered words and moved
the line past the cover, so the covered words went unreported. The
audit now compares against the line's first painting operator: any
run painted before the cover is checked, and runs painted after it
in the same low-contrast color are unreadable on it too.
The annotation scan and the form XObject walk read /Subtype only as a
direct name. A /Subtype stored as a reference to the name read as
none, so a /Redact annotation or a covering form named that way was
skipped as an unrelated object and the page reported clean. Both now
follow the reference, and a reference that cannot be loaded fails the
audit.
The cover walk cleared the current path when it restored the graphics
state, but the current path is not part of the graphics state. A
rectangle built before Q and filled after it paints a box over the
text, and the audit missed it and reported the page clean.
reading_lines grouped the joined operators by the stream their first token
is in and walked each group on its own. A marked-content sequence opened in
one stream (its /ActualText dictionary, or the whole BDC) and closed in the
next lost its replacement text: the export read "Glyphs" where every reader
of the joined streams reads "Replacement text".
The page's operators are now walked once. Each shown text still counts as
the stream its operator starts in, and text from another stream still starts
a new line, as before.
Pieces::join separated a page's content streams with a newline, which ended
a comment left open at the end of one stream. hayro, MuPDF and Poppler join
them with a space, so the comment runs on into the next stream and hides its
first line. Edit text, object_list, the redaction audit, redaction and the
Office/HTML export therefore saw text and boxes that nothing draws, and
redacting unrelated text on the page made the hidden content visible: the
rewritten stream dropped the comment.
- Pieces::join uses a space, so every reader of the joined streams agrees
with rendering.
- The parser records, per piece, whether a top-level comment runs into it
and out of it (boundaries only, not every comment). A redaction rewrite
of a piece ends an incoming comment before its first operator and leaves
an outgoing one open, so hidden content stays hidden.
- Reading for export parses the streams joined, instead of one by one.
- The redaction tag check compares the joined streams the same way.
- Pieces::piece_of uses a binary search.
Hidden (/F bit 2) and NoView (/F bit 32) annotations are not drawn on
screen, so the text under them is in plain view; the audit now skips
them the way the renderer does. A Print-only square still covers.
A cover counted only when it hid 40% of the line's words or box, so a
black box over one number in a long line, and an unapplied redaction
mark over it, audited clean while the number was still extractable. A
cover that hides half of any one word now counts as well.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RUNQSPVXx3dDBabRZHAnjz
* fix(audit): an applied redaction at the end of a line is not a faux redaction
After Apply Redactions removes text at the end of a line, the line keeps a
TJ adjustment where the text was, and the box of the line still spans it.
The audit compared the redaction box with that box: at the end of a line
it covered more than 40% of it and the words left beside it were reported
as covered text, on a correctly redacted file. In the middle of a line the
gap was usually too small to cross the threshold.
Text lines now also give the boxes of the words they draw (`ink`: glyphs,
without spaces and TJ adjustments), and the audit measures how much of
those words a cover hides. A real cover over text that is still there is
found as before.
doc_audit_redactions also reports its rectangles in top-left-origin points
like comment_list and doc_info; they were in PDF user space.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(audit): leave the rectangles of doc_audit_redactions in user space here
Reporting doc_audit_redactions rectangles in top-left-origin points is a
separate subject from the false alarm this branch fixes, and on its own it
made the tool and `pdfcraft-cli audit` disagree on the same file: the CLI
still prints PDF user space. Take it out of this change; the tool and the
CLI can move together in their own change.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(edit): word boxes span whole words when glyphs run backwards
With a negative font size or a negative horizontal scale (Tz), each glyph
advances backwards along the text matrix, and the box of a word kept only
its first glyph: the end of the word was compared with a start that had
already moved past it. A word box is now the union of its glyphs whichever
way they go, and a TJ adjustment ends a word past 0.2 em of the scaled size,
whatever its sign.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(audit): a cover over a short value after a long label is still found
Measuring a cover on the words of a line alone missed faux redactions main
found: a long label followed by a short value hidden under the box is less
than 40% of the line's words. A cover now also counts, as before, when it
hides 40% of the line's box and at least half of one word still drawn; a
cover over the gap an applied redaction left, with no word under it, still
does not. The audit README states both rules.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(audit): retain partial-value cover detections
Co-Authored-By: OpenAI Codex <noreply@openai.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: OpenAI Codex <noreply@openai.com>
- Pin FindingKind stable IDs (covered-text, unapplied-mark) with a test
and document the no-rename contract on id()
- Golden tests pinning the full AuditFinding struct for both finding
classes (page, kind, exact rect, covered text)
- Depth-limit boundary tests: 32 nested form XObjects audit clean,
33 fail closed with TooDeep
- Fix stale crate docs: m/l axis-aligned rectangles count as covers (#1510)
- Document AuditFinding.rect normalization and covered_text joining
Tests + docs only, no behavior change.
Co-authored-by: Erick Barraza <ejbarraza1994@gmail.com>
The redaction audit turned an re rectangle into an unknown path on a
redundant h, and its fill branch left out the close-fill-stroke operators b
and b*, so those opaque covers over extractable text produced no finding.
Keep a closed rectangle closed under h and treat b/b* like the other fills.
Claude-Session: https://claude.ai/code/session_01RUNQSPVXx3dDBabRZHAnjz
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* fix(audit): a rectangle drawn as an m/l path is a cover like re (fixes#1510)
* docs(parity): a rectangle drawn as an m/l path counts as a cover (#1510)
* Rework faux-redaction audit per maintainer review
- Paint-order aware: a rectangle counts as a cover only when painted
strictly after the text (positions from the joined content stream).
- Colour-aware: WCAG contrast between text and cover must be below 3.0,
so page backgrounds, table shading and highlights are not flagged.
- Fail closed: audit_redactions returns Err on undecodable content
streams (decoded_strict) and malformed /Annots; the operator bound is
an error, not a silent truncation.
- Graphics state carries across a page's content streams (one joined
walk); form XObjects still isolate their state at Do.
- Audit-only: the sanitize-for-sharing half is dropped (upstream #162).
- Parity entry marked partial, not shipped.
* Audit: register its layer, add a README, fail closed past 32 nested forms, document what it doesn't check
---------
Co-authored-by: Erick Barraza <ejbarraza1994@gmail.com>
Co-authored-by: Brandon Thomas <echelon@gmail.com>
Gabriel Luís da Silva