raster_snapshot produces RGBA pixels through AnyRender/Vello CPU. pdf::pdf_snapshot consumes the same owned snapshot and writes PDF through upstream Krilla at 5eb4e1f. CDP Page.printToPDF and CLI --dump pdf use the PDF backend. CLI PDF exports include CSS backgrounds, preserving colored controls and white text on colored backgrounds. CDP continues to use its printBackground option, which defaults to false. Both paths apply print media styles unless explicitly overridden. The renderer owns the print request, restores media after snapshot capture, and returns PDF bytes to both clients without publishing interactive geometry. Vector print capture is independent of the full-document bitmap dimension limit.

The entry points and shared code are organized by responsibility:

moli/src/fetch_dump/pdf.rs starts CLI printing; screenshot.rs handles PNG exports.
moli-renderer-v8/src/runtime/page_print.rs owns PDF request/reply types and the print lifecycle. page_screenshot.rs owns screenshot/screencast types and capture.
moli-paint/src/validation.rs validates shared resource references, scope balance and work budgets; error.rs defines the public paint errors.
moli-paint/src/geometry.rs converts shapes and transforms; scene/ converts drawing primitives into AnyRender commands for both output backends.
moli-paint/src/raster/ allocates and renders pixel surfaces; pdf/ encodes native PDF content and invokes bounded raster fallbacks for effects that need them. Ordinary PDF text bypasses AnyRender to retain Unicode mappings.

The dependency uses a fixed Git revision with only raster-images enabled; simple-text remains disabled because Parley already performs shaping. Krilla's own pdf-writer Git pin (fdf0901) is recorded in Cargo.lock. This revision requires Rust 1.92; the repository toolchain is 1.96.1. Krilla and the adapter now share skrifa 0.44 with Parley, Vello and usvg. The adapter uses skrifa::raw instead of declaring a separate read-fonts dependency. Moli's root [patch.crates-io] also aligns subsetter with skrifa 0.44 and write-fonts 0.50, so subsetting shares read-fonts 0.41 and font-types 0.12 with the rendering stack. The patch pins our subsetter dependency update and can be removed when upstream publishes compatible fontations dependencies. The dependency update preserves the subsetting algorithm, variable fonts and CFF2 support. Krilla's imagesize 0.14 and base64 0.22 also coexist with usvg's 0.15 and 0.23; their version requirements do not overlap.

The PDF adapter embeds and subsets fonts, using Parley's existing glyph positions and the snapshot's shared logical text/UTF-8 cluster ranges. These mappings are collected only for print demands, keeping them out of the screenshot hot path. Ligature ranges are joined in logical order before arranging RTL clusters visually. Synthetic italic uses a glyph-local shear; synthetic weight uses PDF fill-and-stroke. Font instances are cached separately by font identity and normalized variation coordinates. The adapter reverses avar v1 and fvar normalization for Krilla's design-space variation API. Non-default avar v2 instances currently return an explicit error.

Paths, borders, decorations, basic SVG paths, images, sRGB linear and non-repeating radial gradients, clips, opacity and PDF blend modes use native PDF primitives. Existing AnyRender geometry lowering is shared for non-text drawing; text bypasses its glyph API because that API does not carry Unicode semantics. SVG scenes that cannot be replayed natively, shadows, other gradient interpolation spaces and filters use bounded CPU raster fallbacks. Ordinary filtered text gets one transparent text copy for extraction, subject to page ownership, clips and opacity. Fully transparent groups contribute no text. Curved/rotated clips use a bounded CPU coverage probe at two samples per CSS pixel to exclude invisible clusters; the actual glyphs remain native PDF text. Destination masks that leave some visible ink and contain body text currently return an explicit error: their per-cluster semantic visibility is not inferred from the result image's bounding box. Destination-dependent operators (including plus-lighter) also require a containing isolated layer. Shadow and clip-mask glyphs never add semantic copies. When visible lettering comes only from background-clip:text with a transparent foreground, it is rasterized mask artwork and is not searchable in this version.

krilla-svg was evaluated but not added: both its compatible release and the pinned upstream Git revision use usvg 0.47, while Moli owns usvg 0.48 trees. Basic SVG geometry reuses the existing AnyRender lowering; more complex SVGs use a bounded image fallback.

Pagination retains the previous fit-to-printable-width behavior and slices the continuous print layout. Each text cluster is assigned by the center of its visible, transformed ink bounds after document clipping. It has one semantic owner page; ink that crosses onto a neighboring page is outlined there. This prevents off-page runs and replayed slices from duplicating the extracted document. Device pixel ratio does not affect PDF page geometry. Page ranges still refer to original document pages.

This does not implement CSS fragmentation (break-inside, repeated table headers, @page), tagged PDF or document outlines. HTML anchor links are retained as PDF annotations, including resolved relative URLs, wrapped lines, transforms, overflow and clip-path clipping, embedded frames and page boundaries. Same-document IDs and legacy named anchors become PDF destinations; omitted destination pages fall back to the original URL. Curved clip boundaries use half-CSS-pixel strips; rectangular and transformed rectangular clips use polygon intersections. JavaScript handlers, image maps and links inside SVG resources are not exported. PDF reading order still follows paint order, and complex-script extraction depends on the reader's handling of ToUnicode/ActualText. SVG text that takes a raster fallback has no separate text mapping. This is not an accessibility or PDF/UA implementation. Krilla uses ActualText for ambiguous/multiple-glyph clusters and Type3 fonts for some color glyphs. A reader that ignores these can report incorrect text even when MuPDF/Poppler extract it correctly; track upstream issues 411 and 425. In the generated shaping sample, PDF.js 5.6.205 reproduces incorrect extraction of ordinary a after the same glyph was mapped to the a + combining acute cluster. MuPDF 1.28.2 and Poppler 25.03.0 extract those strings correctly. These results were rechecked with the pinned Git revision. This is an observed compatibility limit, not a claim of cross-reader complex-text conformance. Apple Preview and interactive browser selection/copy still require manual checks.

Run the PDF regressions with:

cargo nextest run -p moli-paint -E 'test(pdf::)'
cargo nextest run -p moli-protocol -E 'test(print_to_pdf) | test(domains::page::pdf::)'

tests/fixtures/moli-pdf.ttf is original MIT/Apache-2.0 test data derived from the repository's generated Ahem face. It adds fi/ffi ligatures and a wght axis with a non-linear avar map, combining-mark positioning, Hebrew coverage and COLR color glyphs. Regenerate it with uv run scripts/generate-pdf-test-font.py from the repository root.

Reader-level regression and artifact generation:

uv run --locked --script scripts/check-native-pdf.py --moli target/debug/moli
cd moli-cdp-smoke
uv run moli-cdp-smoke --group pdf

The reader script checks the real CLI with MuPDF, including mixed RTL, multi-glyph clusters, repeated base letters, color glyphs, opacity, shadows, filters and a 200,100 CSS-pixel document. It retains PDFs, rendered pages, extracted text and per-process timings/sampled peak RSS in target/native-pdf-check/. The Rust tests use lopdf for PDF objects and ordinary text; its extractor does not interpret ActualText and Type3 text correctly, so it is not the complex-text oracle. The reader script also compares rendered solid and gradient button backgrounds against CLI screenshots; extracting a white label alone does not establish that the label is visible in the PDF. It checks the white label's visible pixels as well as its extracted text. A local fixture server supplies relative URLs, wrapped links, rotated and clipped links, an iframe with an SVG icon, and internal ID/named anchors. MuPDF checks both targets and hit regions, including PDF QuadPoints and destination page numbers. The font fixtures and page resources are bundled; these checks do not depend on public websites or system fonts.

CI runs this script against the exact HEAD release artifact in the Native PDF rendering and links job. Its locked MuPDF dependency and Python 3.12 match the local command above. Failure blocks the job; PDF files, raster previews, screenshots, extracted text and link JSON are uploaded even when a check fails. The CI job does not substitute PDF parsing or nonempty text for visual checks.

Validation on 2026-10-12 (Linux x86_64, Rust 1.96.1): cargo fmt --all, workspace Clippy with all targets/features and warnings denied, and the full nextest run passed (20,012 tests; 17 skipped). The Python smoke unit suite passed 43 tests. The real PDF smoke group passed with both debug and release binaries. The visual comparison covered borders, rounded clips, an sRGB gradient, alpha image, SVG paths, transformed text, opacity, shadow and blur. Interior gradient samples differed by at most one 8-bit channel value from the raster screenshot. The Git revision also passed the real CLI/MuPDF checks and CDP PDF smoke group; its rendered visual sample was pixel-identical to the crates.io baseline.

Release sizes measured with Rust 1.96.1 on Linux x86_64, using cargo build -p moli --release --locked with the repository's unchanged release profile and default features. Packaging used scripts/release.py with version 1.1.16, which strips the binary and creates a gzip level-9 archive:

Revision	Release binary bytes	Stripped binary bytes	Archive bytes
Before native PDF (`d9928760`)	156,266,960	107,667,480	43,425,106
Native PDF, official Krilla Git (`7a52ca98`)	158,167,208	109,363,224	44,157,871
Same Krilla, root `subsetter` patch (`8d38152`)	157,856,280	109,084,696	44,031,505

The subsetter dependency update removes the duplicate skrifa, read-fonts and font-types versions. It saves 272 KiB in the stripped binary and 126,366 bytes in the archive. The stripped binary's native-PDF increase relative to d9928760 is now 1.35 MiB (1.32%), down from 1.62 MiB. The four CLI reader samples below (82 pages total) produced byte-identical PDFs before and after the dependency update; their extracted text and rendered pixels also matched. The updated subsetter passed all 174 tests, and 14 static, variable, CFF and CFF2 font-subset comparisons were byte-identical to the registry version.

One observed release CLI run with local data URLs and fixed fonts:

Input	Pages	PDF bytes	Whole CLI time	Sampled process peak RSS
Font Text 123	1	3,290	95 ms	53.3 MiB
Shaping / font variants	1	9,119	32 ms	52.9 MiB
Opacity / shadow / blur	1	7,070	52 ms	56.1 MiB
286 rows, 200,100 CSS px	79	45,215	144 ms	59.3 MiB

These include process startup, font loading, layout and encoding. RSS is sampled from /proc every 5 ms, not an allocator-level peak; timings are observations, not a performance guarantee. The long document contained no Image XObjects, and all 286 row labels were extracted once. The effects sample contained two local raster images, with one semantic copy of each visible text run.