v2.6.0

Fixed

• Saves write every xref entry with the object's real generation number (issue 80). Every entry
  was written with generation 0 even when the object header said N G obj with G > 0. Our own reader
  and pypdf tolerated it; veraPDF and PDFBox did not. Affects any document that carried a
  non-zero-generation object through a save.
• Strings written as hex in the source are written as hex on save (issue 57), so byte-level
  conformance checks (PDF/A 6.1.6) see the same form after a round-trip.
• Real-number operands keep their written precision on save instead of being re-formatted
  through a double, which could change low-order digits.
• Page-tree traversal is bounded: a /Kids cycle no longer recurses without limit.
• Transparency-group resource cycles are guarded during rendering and content walks.
• Atomic-file-write retry no longer races its own budget on a transiently locked destination (issue 55).
• CID-to-GID mapping is split by consumer (issue 42): rendering keeps the identity mapping the font
  program requires; conformance uses the strict map.
• Glyph selection for a CIDFontType2 descendant whose embedded program is a CFF-outline (OTTO)
  sfnt (issue 36): the charset lookup is keyed on the descendant's own /Subtype rather than the
  program's outline flavour, which previously scrambled every glyph in such a font.
• XMP: structured properties are no longer destroyed on save. Setting any document property
  (PdfDocumentEditor.Metadata.Title and friends) re-serialized the XMP packet through a model with
  no struct representation, flattening xmpMM:History, xmpMM:DerivedFrom, xmpTPg:Fonts and
  similar into a single concatenated text blob — irrecoverably. XmpPacket is now backed by the
  recursive XmpNode model and emits structs and arrays-of-structs faithfully. Shapes the model
  still cannot express (a general XMP qualifier — rdf:value plus a non-xml:lang qualifier field
  or attribute) are preserved verbatim rather than dropped.

Changed

• The XMP format layer moved to its own assembly, PdfLibrary.Xmp, bundled in the package as with
  ICCSharp and FontParser. Namespaces are unchanged and PdfLibrary forwards the four public
  types, so this is source- and binary-compatible. The forwarders are transitional and are due for
  removal at 3.0.0, which will require a recompile for consumers still binding to the old
  assembly.
• XmpProperty's read projection of a struct-shaped property changed. The old flat parser
  returned the concatenated text of a struct's descendants as Value; now that structs are modelled
  faithfully (see Fixed, above), a struct has no scalar value and XmpProperty.FromNode projects it
  to Value = "" instead. This is deliberate and more correct, but it is a read-side behaviour
  change on a public type, and it reaches conformance: UaTitleRule reads dc:title through this
  projection, so a struct-shaped (qualified) title now reads as empty rather than as its
  concatenated text. Note also that the verbatim-preservation fallback this release adds is narrow —
  it only preserves an rdf:value-plus-qualifiers shape, not any unfamiliar XMP in general; see
  Docs/Architecture.md for what it does and does not cover (e.g. rdf:type on a struct is still
  dropped).
• Preflight reaches full verdict parity with veraPDF on every profile with a committed snapshot:
  986/986 whole-file agreement on the PDF/A-2b corpus, 22/22 on PDF/A-2u, with zero false positives
  across all 1,316 corpus files. The clause-by-clause table is
  PdfLibrary.Tests/Conformance/parity/PARITY-REPORT.md; clauses at full parity — PDF/A-2b (35/40):
  6.6.2.3.1, 6.2.4.3, 6.2.10, 6.3.3, 6.3.2, 6.6.2.3.3, 6.1.13, 6.5.1, 6.3.1, 6.1.7.1, 6.1.9, 6.2.4.4,
  6.4.1, 6.2.2, 6.2.5, 6.6.4, 6.2.11.3.3, 6.2.11.6, 6.2.3, 6.2.8.3, 6.2.9, 6.1.12, 6.1.3, 6.2.6, 6.2.8,
  6.2.11.3.2, 6.2.4.2, 6.6.2.1, 6.1.10, 6.1.4, 6.1.6, 6.1.8, 6.2.11.4.2, 6.4.2, 6.5.2; PDF/A-2u (3/3):
  6.2.11.7.2, 6.2.11.3.1, 6.6.4.
• Font scans no longer parse every charstring eagerly; a conformance scan of a font-heavy document
  is roughly half the time it was.
• The core package is 5.4 MB, up from 3.0 MB: bundled Adobe CID-to-Unicode CMaps, Standard-14 AFM
  metrics, and the CMYK soft-proof profile.

Added

• XmpPacket.SetStruct / SetStructArray and XmpField, for authoring nested XMP structs and
  arrays-of-structs (e.g. pdfaExtension:schemas).
• XmpPacket.Serialize()'s name/value contract is now explicit. An illegal character in a
  property value (a NUL/control character, an unpaired surrogate) is sanitized to U+FFFD rather
  than rejected, so a save is never blocked by garbage data the library did not write. An illegal
  property or struct-field name — only ever caller-supplied, via SetSimple/SetArray/
  SetStruct/SetStructArray — throws ArgumentException instead, since there is nothing to
  repair it into. Callers authoring structs with dynamic field names should validate them first.
• System font substitution. SystemFontLocator.Resolve(FontRequest) returns a FontMatch
  (bytes plus the face index inside a .ttc); EnumerateFaces() lists what the locator can see as
  SystemFontFace records (family, PostScript name, bold/italic, source path, face index);
  SystemFontLocator.Default scans the platform font directories. ISystemFontProvider gained
  both members with default implementations, so existing implementers keep compiling.
  BundledStandard14Provider wraps any provider with the Standard-14 AFM metrics.
• Font inventory. FontInventory.Read(PdfDocument) lists every font in a document as a
  FontInventoryEntry (kind, embedding, program holder, used codes); FontInventory.Find locates
  one by object number. FontProgramClassifier, ClassifiedProgram, FontProgramFormat,
  FontDescriptorValues, FontDescriptorMetrics, FontId, FontKind support it.
• CID-to-Unicode for text extraction: CidCMap and AdobeCidToUnicode (bundled Adobe UCS2
  tables for Japan1, Korea1, GB1, CNS1), ToUnicodeCodespace, and ToUnicodeCMapWriter for
  authoring a ToUnicode CMap.
• Metadata: PdfMetadata.Language and PdfMetadata.Trapped (PdfTrapped), synced to XMP.
• Editing: PdfDocumentEditor.SetFileId, ConsolidateOutputIntents, ReplaceOutputIntentProfile,
  PdfPageCollection.SetAnnotationFlags, PdfFormFields.HasXfa, PdfAnnotationInfo.Flags, and
  PdfDocumentEditor.Document.
• PdfDocumentEditor.PreviewDocumentProof() classifies signature byte ranges and catalog
  permissions without mutating (DocumentProofPreview).
• Conformance reads: ConformanceClaim.Read (the declared PDF/A profile, with an unreadable-XMP
  flag), OutputIntentProfileValidator.Validate, XmpConformance.ClassifyProperties and
  ModernEquivalentOf (XmpPropertyVerdict, XmpModernEquivalent), and
  Preflighter.Check(document, profile, sourceBytes) for byte-level rules.
• Render SPI: ImageCommand.ProofCmyk, PdfGraphicsState.ResolvedFillProofCmyk /
  ResolvedStrokeProofCmyk, DeviceCmykConverter.Naive, and
  RecordingRenderTarget.Record(page, scale, fontProvider). The nine-parameter ImageCommand
  constructor and Deconstruct from 2.5.2 are kept.