yfedoseev/pdf_oxide v0.3.76

v0.3.76 | Redaction/editor persistence and rendering-accuracy release: DOM edits and `add_text()` overlays on source-loaded pages now survive `save()` without leaving dangling `/Contents` references; multi-input (DeviceN) Type 0 tint transforms, non-CCITT 1-bit images, and inline images render correctly; CCITT `/ImageMask` XObjects decode instead of misreading compressed data as raw stencil rows; spatial table-cell ownership is corrected at singleton-span and boundary cases; and repeated `search()` calls on the same document are no longer `O(searches × full extraction)`. The Java binding gains PDF/A conversion, completing coverage across every first-party language binding.

on GitHub

Added

• PDF/A conversion exposed in the Java binding, with idiomatic Kotlin/Scala/Clojure facades (PdfAConverter, ConversionResult/ConversionAction/ActionType/ConversionError in fyi.oxide.pdf.compliance) — completes PDF/A conversion coverage across every first-party binding (#948).

Fixed

Editor / redaction persistence

• PdfDocument.save() silently discarded DOM edits (set_text(), remove_element(), erase_header()/erase_footer()/erase_artifacts()) made to a page loaded from an existing PDF — the overlay-merge path that reconciles DOM edits back into a page's /Contents on save had a defect that dropped the edit entirely; edits made via the DOM API now persist through save()/save_page() (#940).
• A page whose original /Contents was already a multi-entry array (ISO 32000-1 §7.7.3.3) kept a dangling reference to its other original streams after destructive redaction — the merge step that replaces /Contents with the redacted stream assumed a single original stream at array position 0; it now replaces every original content reference, keeping only genuine overlay/addition streams. The same defect class was independently hit through two different triggers: set_text()/remove_element() on a page whose /Contents was already an array, and add_text() combined with apply_redactions_destructive() on the same page (#940, #941, #799).
• add_text() overlay font registered under the raw font name while the content stream emitted the map_font_name()-transformed name, leaving a dangling /Tf resource for bold/italic overlays, generic family names (Arial, sans-serif), and Symbol/ZapfDingbats — registration now keys off the exact name the Tf operator emits, and /Resources//Resources/Font are resolved through indirect references (common for pages loaded from existing documents) (#941).
• save_page() discarded overlay text staged by a prior save_page() call on the same page — overlay_additions now accumulates across calls instead of being overwritten (#941).

Rendering

• CCITT Group 3/4 /ImageMask XObjects were treated as raw stencil rows instead of decoded — compressed fax data is now actually decompressed; DecodeParms/filter-chain handling, K < 0 Group 4 semantics, and allocation on oversized/malformed input are also corrected (#935, #939).
• Separation/DeviceN colour spaces with a genuinely multi-channel (N>1) Type 0 (sampled) tint transform rendered black or dropped shapes entirely — the sampled-function evaluator only ever handled a single input dimension, forwarding just the first scn operand and silently falling back to gray = 1 − components[0]; it now performs full N-dimensional multilinear interpolation across the sample grid (ISO 32000-1 §7.10.2's general algorithm, of which the prior 1-D case is the N=1 special case), gated by a MAX_SAMPLED_FUNCTION_DIMS = 8 bound against a pathological /Size array (#849, #859).
• Non-CCITT 1-bit /DeviceGray images (uncompressed or FlateDecode) were force-fed through the CCITT decompressor and silently dropped when decompression failed — the CCITT path is now gated on the XObject's filter actually being CCITTFaxDecode; the non-CCITT case is unpacked directly, folding /Decode [1 0] inversion the same way the CCITT path already does (#860).
• Inline images (BI…ID…EI) were parsed and classified as a paint operator but never actually painted — the renderer's operator dispatch had no match arm for Operator::InlineImage; it now expands the abbreviated dictionary keys and routes through the same render_image/render_image_mask path used for Do-invoked image XObjects (#860).
• Spatial table-cell ownership misattributed text near cell boundaries — stale per-glyph offsets outside a text span's bounding box are now rejected for singleton spans, and exact half-open internal grid intervals keep superscripts and boundary-adjacent text in their correct geometric cell; outer-edge tolerance near the table boundary is unchanged (#937, #938).

Performance

• Repeated search()/search_page() calls on the same document re-extracted and re-postprocessed every page's full spans on every call — a multi-pattern scan over one document cost O(searches × full extraction) with no benefit from repetition, since the existing span cache is bounded to 8 entries. An unbounded, lighter-weight per-page search index (page text + span bounding boxes only, no font/glyph data) is now built lazily on first use and reused across calls; prepare_search()/clear_search_index() give callers control over eager population and memory reclamation. Measured: repeat search() calls on a 60-page document dropped from ~30ms to ~20-50µs per call after the first (#936).

Changed / Dependencies

• Test/CI stability: two ocr-feature-gated CCITT diagnostic test files still called fax 0.2's u16 decode_g4 signature after the crate's 0.3 API bump (u32), breaking cargo clippy --all-targets --features rendering,barcodes,signatures,ocr — exactly the combo CONTRIBUTING.md's recommended pre-commit hook uses (#945).

Contributors

Rendering fixes from @Goldziher — CCITT /ImageMask decoding (#935, #939) and spatial table-cell ownership (#937, #938). Redaction/overlay persistence from @thomnico — the add_text()/destructive-redaction interaction (#941, #799, superseding an earlier iteration). Thank you!

Issues reported by:

• @lightedlogic — #940 (save() doesn't persist DOM edits made via set_text()/remove_element())
• @ankursri494 — #936 (repeated search() calls on the same document don't get faster)
• @ultrasaurus — #945 (recommended pre-commit hook fails on main)
• @bfchiheb — #947 (PDF/A conversion missing from the Java binding)

Installation

Rust (crates.io)

cargo add pdf_oxide

Python (PyPI)

pip install pdf_oxide

JavaScript/WASM (npm)

npm install pdf-oxide-wasm

CLI (Homebrew)

brew install yfedoseev/tap/pdf-oxide

CLI (Scoop — Windows)

scoop bucket add pdf-oxide https://github.com/yfedoseev/scoop-pdf-oxide
scoop install pdf-oxide

CLI (Shell installer)

curl -fsSL https://raw.githubusercontent.com/yfedoseev/pdf_oxide/main/install.sh | sh

CLI (cargo-binstall)

cargo binstall pdf_oxide_cli

MCP Server (for AI assistants)

cargo install pdf_oxide_mcp

Pre-built Binaries
Download archives for Linux, macOS, and Windows from the assets below. Each archive includes both pdf-oxide (CLI) and pdf-oxide-mcp (MCP server).

Platform Support

Changelog

See CHANGELOG.md for full details.