Inkwright.Signing 0.1.3

Inkwright

A modern .NET 10 PDF library (Apache-2.0) with not one native dependency. The architecture is layered the way System.Text.Json is: COS primitives → streaming I/O → document model → add-ons that only the people who need them pay for.

What it does

• Reads broadly: cross-reference tables and streams, hybrid files, object streams, incremental updates. Recoverable structural damage is scanned and reported through LoadDiagnostics; unrecoverable input raises PdfFormatException. Across a corpus of 7,807 real files, 7,800 open and 7,555 eligible files rewrite and reload; five deliberately corrupt fixtures (no catalogue or header) are rejected, and two empty files never reach the parser.
• Lossless editing: an object the model never parsed is copied byte for byte on rewrite. Object numbers survive, and so do diagrams, annotations and everything the library did not understand.
• Incremental saving: changes are appended and not one original byte moves — the only way not to break signatures somebody else made.
• Report composer (Inkwright.Layout): two-phase Measure/Draw with SpacePlan, line breaking and justification, tables with repeating headers, running heads, "page X of Y", explicit breaks, and a LayoutException carrying the element tree instead of silently clipping.
• Text: built-in metrics for the fourteen standard fonts, TrueType/OpenType embedding with subsetting and Identity-H, letter extraction with coordinates. Checked against PdfPig on real files: mean similarity ≥ 0.90.
• PDF → Markdown: Tagged PDF becomes CommonMark/GFM headings, paragraphs, lists, links and tables; raster Figure elements come out as deduplicated PNG sidecars. Untagged layout inference for columns, headings, lists, and tables is opt-in; every loss of semantics lands in diagnostics.
• Images: JPEG is embedded as it stands, PNG without a single re-encode — its deflate stream is already what PDF wants — and alpha becomes a soft mask.
• Comments: notes, highlight/underline/strikeout over words found by search, FreeText and callouts, shapes, polylines, ink, stamps, attachments and links. Reply threads and review states per ISO 32000-1, cascading deletion together with the popup and the thread, flattening into page content, export to flat records with the quoted text, and a ready appearance stream for every type. Review rounds travel to other tools and back through XFDF (ISO 19444-1) and FDF (§12.7.7), merging on /NM so a second import does not duplicate anything.
• Forms (Inkwright.Forms): typed fields, filling with generated appearance streams, flattening, and [PdfForm] plus an incremental source generator for reflection-free mapping and Native AOT.
• Signatures (Inkwright.Signing): PAdES building blocks from B-B to B-LTA — signature and document timestamps (RFC 3161), the validation data store (DSS/VRI), DocMDP certification signatures and FieldMDP field locking — plus external keys for an HSM, built entirely on the base class library with no BouncyCastle. For B-LT and B-LTA the application supplies OCSP/CRL data it has already verified and sets the trust policy. Digests are SHA-2, and also SHA-3 and SHAKE256 per ISO/TS 32001 where the platform provides them; the Edwards curves of ISO/TS 32002 are verified by arithmetic of the library's own.
• Archival documents: Tagged PDF and eleven PDF/A levels — eight from parts 1 to 3 (A1B, A1A, A2B, A2U, A2A, A3B, A3U, A3A) and three from part 4 (A4, A4E, A4F) — with a generated sRGB ICC profile and a validator that names the ISO 19005 clause a file breaks. Reconciled clause by clause against the texts of parts 1–3 and part 4, and checked against 106 archival files from another implementation. PDF/A-4 is the only level a PDF 2.0 file can claim: parts 1 to 3 require a %PDF-1.n header and part 4 requires %PDF-2.n. Factur-X and ZUGFeRD attachments with the fx: schema description.
• Accessibility: PDF/UA-1 and UA-2 profiles (ISO 14289) with a validator that tells a broken requirement from an unmet recommendation and names the clause; conformance declarations per PDF Declarations, including the Well-Tagged PDF 1.0 levels.
• Encryption: reads RC4, AES-128, AES-256 and AES-GCM; writes AES-256 in cipher block chaining (revision 6) or Galois/counter mode (revision 7, ISO/TS 32003) with authentication. Certificate encryption to X.509 recipients (§7.6.5) both ways, on EnvelopedCms.
• Navigation: outlines (PdfOutline) are read and edited in the model and built by the composer, destinations to a page and to a structure element (§12.3.2.3), and output intents on the document and on a page.
• Integrity: the authentication code of ISO/TS 32004 — written and verified, as a standalone token or as a signature attribute; checked against 94 files from another implementation.

A quick look

using Inkwright;
using Inkwright.Layout;

// A report that flows across pages by itself
Report.Create()
    .Page(page =>
    {
        page.Size(PageSize.A4);
        page.Margin(45);
        page.Header().Text("Quarterly report", TextStyle.Default.FontSize(18).Bold());
        page.Content().Table(table =>
        {
            table.Columns(ColumnWidth.Fraction(4), ColumnWidth.Fraction(1));
            table.Grid();
            table.Header().Text("Item").Text("Amount");
            foreach (var line in lines)
                table.Row().Text(line.Name).Text(line.Total);
        });
        page.Footer().AlignCenter().Text(text => text.PageNumbers("{0} of {1}"));
    })
    .Save("report.pdf");

// Editing somebody else's file: a stamp on top, original bytes untouched
using PdfDocument document = PdfDocument.Load("invoice.pdf");

using (ContentCanvas canvas = ContentCanvas.ForPage(document.Pages[0]))
{
    canvas.Save()
          .Opacity(0.15)
          .Transform(PdfMatrix.Rotation(38) * PdfMatrix.Translation(140, 240))
          .DrawText("PAID", 0, 0, document.Fonts.HelveticaBold, 72, PdfColor.Parse("#1B7F3B"))
          .Restore();
    canvas.Commit();
}

document.SaveIncremental("invoice-paid.pdf");

// The text back out, with coordinates
using PdfDocument document = PdfDocument.Load("scan.pdf");

foreach (PdfLetter letter in document.Pages[0].ExtractLetters())
    Console.WriteLine($"{letter.Text} at {letter.Origin} {letter.FontSize:0.#} pt");

// Reviewing: highlight words, leave a note, reply, set a state
PdfPage page = document.Pages[0];

page.Annotations.MarkText("delivery dates", PdfAnnotationType.Highlight, author: "Priya");
PdfTextAnnotation note = page.Annotations.AddNote(
    new PdfPoint(505, 700), "Has the rate changed?", "Priya", PdfNoteIcon.Help);

note.Reply("No, unchanged", "Tom");
note.SetReviewState(PdfReviewState.Completed, "Priya");

document.SaveIncremental("reviewed.pdf");

// A review round travels between reviewers apart from the document, as XFDF or FDF
document.ExportReview("round-4.xfdf");

PdfReviewImportResult round = other.ImportReview("round-4.xfdf");
Console.WriteLine($"{round.Added} added, {round.Updated} updated");

// A form filled from a typed record
[PdfForm]
public partial record Invoice(
    [property: PdfField("customer.name")] string Customer,
    [property: PdfField("invoice.total", Format = "0.00")] decimal Total,
    [property: PdfField("invoice.paid")] bool Paid);

PdfForm form = PdfForm.Open(document)!;
form.Fill(new Invoice("Ada Lovelace", 1234.50m, true));
form.Flatten();

// A signature as an incremental update
byte[] signed = PdfSigner.Sign(document, new PdfSignatureOptions
{
    Certificate = certificate,
    Reason = "Approved",
    Appearance = PdfRectangle.FromSize(60, 60, 220, 70),
});

Performance

A 200-page report (8000 table rows) and a 90-page document, measured on an Intel Core i7-8700 under Windows Server 2022 with .NET 10.0.10 and BenchmarkDotNet 0.15.8. Method and comparison notes are in docs/benchmarks.md; numbers move with the machine, so treat them as a shape rather than a promise.

Documentation

• Getting started
• Architecture
• Reading and editing
• Navigation: outlines, destinations, output intents
• Layout
• Text and fonts
• Exporting PDF to Markdown
• Images
• Merging and splitting
• Comments and annotations
• Forms
• Signatures
• Tagging, PDF/A and Factur-X
• Accessibility and PDF/UA
• ISO 19005 conformance matrix, parts 1–3
• ISO 19005-4 conformance matrix (PDF/A-4)
• ISO 14289 conformance matrix
• Reconciliation with ISO 32000-2
• Verification baseline and regression policy
• Benchmarks
• Building the NuGet packages

Building

dotnet build Inkwright.slnx
dotnet test Inkwright.slnx
dotnet pack Inkwright.slnx -c Release --output artifacts/packages
dotnet run --project samples/Inkwright.Samples
dotnet run -c Release --project benchmarks/Inkwright.Benchmarks -- --filter *

Layout of the repository

src/Inkwright                  core: COS, filters, reader, writer, DOM, content, fonts, text,
                               colour, functions, image decoding, merging
src/Inkwright.Cjk              the 168 predefined CMaps of ISO 32000-2 Table 116
src/Inkwright.Layout           composer, Tagged PDF, PDF/A profiles
src/Inkwright.Forms            AcroForm: fields, appearances, flattening
src/Inkwright.Forms.Generator  Roslyn incremental generator (packed into Forms)
src/Inkwright.Signing          CMS/PAdES signatures on the base class library
tests/                         xUnit: round-trip over 7807 real PDFs, extraction checked against
                               PdfPig, layout, tagging, PDF/A, forms, signatures
benchmarks/                    BenchmarkDotNet against PDFsharp and PdfPig
samples/                       an invoice, low-level drawing, editing, a form, an archival document

The limits chosen on purpose are listed in architecture.md: no rasterisation, no shaping of complex scripts, JPX and progressive JPEG pass through untouched, and trust-chain policy remains the caller's responsibility.

Built on this

Quillwright.Pdf prints Word documents: it does the pagination, tables, running heads and tagging itself and writes the file through Inkwright. A useful example of what ContentCanvas gives you alongside font metrics and a structure tree.

Cellwright.Pdf prints Excel workbooks: the styled grid with number formats, two-dimensional pagination with scaling, repeating headers and running heads. The text layer it shares with Quillwright — Inkwright.Text.BidiLayout and ArabicShaper — lives here: the same UAX #9, only pointing the producer's way, logical order → drawing order.

Third-party code this project borrows from is listed in THIRD-PARTY-NOTICES.md.