Skip to content

Redaction: paths cut out of the areas, overlay text (part of #198) - #230

Merged
timzifer merged 6 commits into
mainfrom
claude/issue-198
Oct 14, 2026
Merged

timzifer merged 6 commits into
mainfrom
claude/issue-198

Conversation

@timzifer

Copy link
Copy Markdown
Owner

Part of #198 (two of its points; the issue keeps the rest). Stacked on #229 (on #227, #226, #225).

Paths are cut, not only clipped (redactcut.go)

Before, a path reaching into an area was clipped and its numbers stayed in the file. Now:

  • The path construction operators are read again from what is written (float64, as in the file, not the interpreter's float32 points) and each segment is split where it crosses the edge lines of an area (lines: linear; cubics: the cubic in t, bracketed at its turning points and bisected). A run of pieces in the area is replaced by the straight line between where it enters and leaves. The area is convex, so the line stays in it and the loop of run and line encloses only points of the area: what the path fills outside is unchanged. The clip of before stays as well.
  • Strokes: cut against the area narrowed by how far a stroke reaches — w/2·√2, or w/2·mitre limit with mitre joins (default: ten half widths) — so that the new joins stay in the area; what lies nearer the edge stays. A hairline counts as one unit. Dashed strokes are only clipped (a shorter path moves the dashes) — their numbers remain.
  • Filled subpaths are cut as loops (rotated to start outside); open stroked ones lose the runs at their ends; open subpaths both filled and stroked with an end in the area are left clipped only.
  • s, b, b* are written as S, B, B* with the subpath closed explicitly; a closed line there and back (x y m x y l s) is stroked as the line — stilus covers a loop that doubles back with extra points twice at its edges, which showed in the corpus (issue8798r.pdf, issue20232.pdf) before this was handled.
  • A clip path (W n) in an area becomes an empty clip; one reaching into an area is cut like a fill.
  • Split curves are flattened anew by a renderer: along their edge a pixel may differ by a few levels (test tolerance 24 for curves, 2 for lines).

/OverlayText (redactoverlay.go)

ApplyRedactions writes the overlay text in the annotation's rectangle: Helvetica (with the fallback fonts for what it lacks, embedded), size and colour from /DA (size 0: as large as fits, 12 pt lines when repeated), /Q alignment, /Repeat over and over, clipped to the rectangle. /RO is not used (what a viewer shows before applying).

Tests

  • TestCutGeometry: what the cut leaves has no subpath start, segment end or curve sample inside the (narrowed) area, for fills, strokes and both.
  • TestRedactCut: eight shapes (filled circle, stroked curve, open filled path starting inside, round-joined polyline with ends and a corner inside, clip, re … b, a there-and-back s line, a dashed line): renders outside the area equal the original (scale 1 and 3), the area is empty, and the numbers of points inside the area are gone from the file — but for the dashed line.
  • TestRedactOverlay: centred red text, repeated text over a black box, Ω from a fallback font.
  • Corpus API run (pdf.js, 3 pages): no differences, no failures. I also compared, locally, the pages redacted with and without cutting outside the areas: after the s fix no file differs (I did not add that check to the corpus run: redaction outside cutting already changes pages outside areas — glyphs whose boxes meet an area, widgets removed whole — in 334 files).

The /BBox of an annotation's appearance stream is mapped onto its /Rect,
and the border of a field lies on the edge of that box. With the box off
the pixels, the exact clip and the border's edge, both antialiased,
lightened the edge pixels twice: a field Editor.SetFields wrote was drawn
with other edges than the same value shown from a FormState, which is not
clipped. The clip of an annotation's appearance is now rounded outward to
whole device pixels when it stays on the axes, as PDFium, MuPDF, Poppler
and Ghostscript clip rectangles. The accuracy report against PDFium is
unchanged.

overlayFile puts its fields off the pixels, which the test held them to.

Fixes #222.
The first fix of #222 rounded the clip of every annotation's appearance
outward to whole pixels. That made the render of an appearance stream
differ from the page it is flattened to, whose content clips exactly: the
corpus API run reported 41 files of pdf.js whose flattened first page
differed, by up to 4.9 % of its pixels.

The other way round: the appearance cera generates for a widget is
clipped to the widget's box exactly, as the stream Editor.SetFields
writes for it and the form Editor.Flatten draws are. The render with a
FormState, the saved file and the flattened page are drawn alike at any
scale; the border of a field off the pixels is lighter on its outer edge
in all three, as the appearance streams of documents are.

The corpus API report and the accuracy report against PDFium are the same
as on main.
ImportOptions.Structure brings the elements of the source's structure
tree whose content is on imported pages, with the elements above them;
an element left without content is dropped. Where the result has a tree,
the elements follow its own in a Part under its Document element (or its
root): the source's Document becomes the Part, or one is made around its
elements, with the source's /Lang where it differs. A result without a
tree takes the first one brought as its own.

The keys of the parent tree are numbered anew after the result's: pages
(/StructParents), their annotations (/StructParent) and the form XObjects
they draw. A page imported twice is tagged once, its first copy; a form
only pages that did not come draw as tagged content loses its keys. Marked
content and object references are mapped to the copies. Role maps are
merged, a type the result maps otherwise written as what the source means
by it; a class the result defines otherwise is renamed Cls_2, an
identifier the result has id_2; namespaces are joined. /MarkInfo says
/Marked.

The corpus API run gets a step: the pages looked at, imported with their
structure into the document itself, keep their marked content in the
tree. Over the pdf.js corpus it finds no difference.

Fixes #183.
PDF 2.0 lets a page carry /OutputIntents of its own (14.11.5), which
take precedence over the document's on it. Such a page is now drawn with
a view of the document (Page.cmykDoc) whose DeviceCMYK converts through
the profile of the page's intent, read as the catalogue's is: the first
intent of kind GTS_PDFX, GTS_PDFA1 or ISO_PDFE1 with a CMYK profile.

Document is split into what all views share (docShared: the reader,
fonts, the form, the profiles and their budget, transfer functions) and
what stands on DeviceCMYK (cmykView: colour spaces, shadings and
patterns by reference, decoded images under an id of their own). Pages
with one profile share a view; forms and patterns drawn on a page take
its view. NaiveCMYK, CMYKProfile and IgnoreOutputIntent decide for the
pages as for the document. A page profile counts against
maxCMYKProfiles, and one cera cannot read is counted as icc-lut, the page
then converting as the document does.

The corpus API report and the accuracy report against PDFium are the
same as before.

Fixes #186.
SignOptions.Timestamp takes a TimestampFunc that gets a token (RFC 3161)
from a timestamp authority for the SHA-256 digest of the signature value;
the token goes into the signer info as the unsigned attribute
signature-time-stamp (PAdES B-T). cera does no network I/O: the function
makes the round trip, and may answer with the token or with the whole
TimeStampResp (cms.Token). A token that does not stamp the digest, or
whose authority did not sign it, is an error (cms.Stamps). The room for
/Contents grows by 16 KB; Editor.SignContext passes a context.

Editor.Timestamp writes a document timestamp (/Type /DocTimeStamp,
ETSI.RFC3161) over the update, in an invisible field. Editor.AddValidation
records certificates, OCSP responses and CRLs the caller fetched, written
into the document security store (/DSS with /Certs, /OCSPs, /CRLs and a
/VRI entry per signature, keyed by the SHA-1 of its /Contents) by Update,
Sign and Timestamp; what the store has stays and is not stored twice.
With a document timestamp after it the signatures are B-LTA. Both are no
change by the rules of a certification (PDF 2.0, 12.8.2.2.2), so a
document certified to allow none may get them, and Verify does not count
a document timestamp as a change.

Two faults of Verify's comparison of revisions, found on the way: a
reference to a page was compared by the page's content, so that a link
to a page whose /Annots a later signature changed made the signature
report annotations changed; and fields were matched by name, so that a
form with two fields of one name reported a value filled in. Pages are
now compared by their place, fields by their objects.

The corpus API step "sign" signs with a timestamp, adds validation data
and a document timestamp, and checks both signatures.

Fixes #193. Revocation from /DSS when verifying, and the smaller points
#143 left, are #228.
A path that reached into an area was clipped, and its numbers stayed in
the file: a curve half under the area could be read back whole. Now its
segments are split where they cross the edges of the area (lines and
cubic curves against the lines of a convex quad), and each run of pieces
in the area is replaced by the straight line across it; the clip stays.
The area is convex, so what the path fills outside it is the same. A
stroke is cut against the area narrowed by how far it reaches (half its
width, the mitre limit with mitre joins), so that its new joins stay in
the area; a dashed stroke is only clipped, since a shorter path moves its
dashes. A filled subpath is cut as a loop; a stroked open one loses the
runs at its ends. s, b and b* are written as S, B and B* with the path
closed itself, and a line there and back is stroked as the line, as the
renderer draws it. A clip in an area becomes an empty one, a clip that
reaches into one is cut too.

ApplyRedactions writes /OverlayText over the rectangle of a redaction
annotation, in Helvetica (fallback fonts for what it lacks) at the size
and in the colour of /DA, aligned by /Q, with /Repeat over and over.

Part of #198: the other points of its list stay open.
@timzifer
timzifer merged commit d2690e6 into main Oct 14, 2026
38 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant