Merge pull request #675 from fastmikey99/fix/pdf-import-past-page-clip

PDF import: clip art reaching past the page to the crop box

ROADMAP.md

@@ -4,7 +4,7 @@ VectorCraft is a clean-room, open-source, pure-Rust reimplementation of the Adob
 
 This file tracks **how far we are and what's left**. Time estimates are wall-clock hours of continuous Claude Opus 5.5 agent work (including builds and the CI gate), given both for **one agent** and for **4–6 parallel agents** on disjoint crates. They are counted from the remaining work (see [Parity estimate](#parity-estimate)), calibrated against measured throughput, and updated as work lands.
 
-_Last updated: 2026-10-09 (after placed documents (File › Place of a `.vectorcraft` file, linked), M4.14–M4.98, M8.1–M8.20, M14.4–M14.7, the MCP prompts, resource templates, completions and logging, the 2026-10-07 issue fixes, the Layers panel rework and the 2026-10-08 community contributions: Pen and shape-tool modifiers, path editing, Layers, Artboards, Swatches/Brushes and dialogs, Japanese composition and PDF text, Hebrew/Arabic type, interface languages (Spanish among them), the font menu, saved selections, MCP and EPS, the missing-font export warning (#319), open paths in the Shape Builder (#374), PDF/`.ai` import keeping strokes live and hidden and nested layers (#372), per-corner Live Corners (#381), SVG type opening at the size it draws at (#396), the Wayland clipboard with copied-file paste (#398), ruler guides that can be selected, moved and deleted (#414), two batches of Preferences options that were stored but unused (#394), skeleton reads with `document.find` for large documents over MCP (#416) and anchors dragged with Direct Selection snapping to points, segments and smart guides, clipped SVG layers opening as clipping layers (#420), PDF/`.ai` layers read on pages of tens of millions of operators (#419), Document Color Mode putting RGB greys on the black plate on request and converting pattern tiles (#421) and movable dialogs with Offset Path's Joins dropdown (#423) and numeric fields scrubbed by dragging their label (#400), stroke weight presets in the stroke unit (#426), the Appearance panel's Opacity popup (#430), type on a path brackets, flip and Type on a Path Options that keep the type in place (#429), Layers rows dragged with a ghost and dropped on the trash or New Layer (#435), tool flyouts opening on a right or corner press (#436), tool flyouts torn off into floating toolbars (#437), the Shape Builder's erase mode deleting edges too (#439), installed fonts registered outside the font folders and fonts installed while the app runs listed in the font menus (#443), alike circular corners on rectangles that aren't square (#442), tool cursors as OS cursors so the pointer no longer trails over the canvas (#444), artboards copied and pasted with their art (#446), guides dragged out of the rulers snapping to the art's side midpoints and the artboards, artboard guides (#451), the toolbar's double arrow and Window › Toolbars › Double Column switching between one and two columns in any window height (#455), the Smart Guides display options and Snapping Tolerance (#394), Font Size and Leading combo boxes with preset lists in the Character and Properties panels and the Control bar (#458), screenshots pasted from the clipboard on macOS, Windows and the web (#457), the Properties panel's Fill and Stroke matching the Control bar's, Illustrator's Window menu panel keys (Shift+F7 for Align…) and Alt-wheel zoom whatever the wheel preference (Discord feedback), the point/area type widget (Discord feedback), groups in legacy Illustrator .ai files (Discord feedback), a rotating desktop log file (#465), image controls in the Control bar and Properties panel: Image Trace with its presets, Mask, Crop Image, Embed/Unembed, and an Image Trace object's preset, Expand and Release (Discord feedback), the Pen working as the Anchor Point tool with Alt held over a selected path's handles and anchors, and anchor buttons that convert, connect and cut paths at the selected anchors (Discord feedback), Japanese hanging punctuation (burasagari, #463), Escape leaving Presentation Mode with no screen mode saved across restarts, documents dropped on the desktop app opening as tabs of their own and a note that an `.ai` file's art off its artboards isn't in the part read (#472), Add/Delete Anchor Point tool icons drawn from the Pen tool's (#479), EPS import running more of PostScript Level 3 (patterns, Type 3 fonts, mesh shadings, masked images, VM and resource probes), and a preview that stands in saying why (Discord feedback, #466, #469, #474), chords with modifiers recorded in Edit › Keyboard Shortcuts (#487), Type Object Selection by Path Only (#394), Anti-aliased Artwork (#394), Show Pixel Grid (#394), pen and touch taps pressed and lifted within one frame clicking where they tapped and the Linux/Wayland pen-tablet workaround (#491), bounding-box resizes landing on the other objects' anchors and centres with labels, and the Rotate, Scale, Shear and Reflect tools snapping their drags (#482), Shift-drag marquees that toggle the objects or anchors they reach and a Lasso whose Alt subtraction keeps the rest of the selection (#483), the mouse wheel stepping the focused numeric field, the dialogs' and preferences' drag values made such fields (#485), Show The Home Screen When No Documents Are Open (#394), Numbers Without Units Are Points (#394), Auto-Collapse Iconic Panels (#394), the Contextual Task Bar moved by its handle and its More Options menu pinning and resetting its position, a designed macOS DMG windoLine truncated
+_Last updated: 2026-10-10 (after placed documents (File › Place of a `.vectorcraft` file, linked), M4.14–M4.98, M8.1–M8.20, M14.4–M14.7, the MCP prompts, resource templates, completions and logging, the 2026-10-07 issue fixes, the Layers panel rework and the 2026-10-08 community contributions: Pen and shape-tool modifiers, path editing, Layers, Artboards, Swatches/Brushes and dialogs, Japanese composition and PDF text, Hebrew/Arabic type, interface languages (Spanish among them), the font menu, saved selections, MCP and EPS, the missing-font export warning (#319), open paths in the Shape Builder (#374), PDF/`.ai` import keeping strokes live and hidden and nested layers (#372), per-corner Live Corners (#381), SVG type opening at the size it draws at (#396), the Wayland clipboard with copied-file paste (#398), ruler guides that can be selected, moved and deleted (#414), two batches of Preferences options that were stored but unused (#394), skeleton reads with `document.find` for large documents over MCP (#416) and anchors dragged with Direct Selection snapping to points, segments and smart guides, clipped SVG layers opening as clipping layers (#420), PDF/`.ai` layers read on pages of tens of millions of operators (#419), Document Color Mode putting RGB greys on the black plate on request and converting pattern tiles (#421) and movable dialogs with Offset Path's Joins dropdown (#423) and numeric fields scrubbed by dragging their label (#400), stroke weight presets in the stroke unit (#426), the Appearance panel's Opacity popup (#430), type on a path brackets, flip and Type on a Path Options that keep the type in place (#429), Layers rows dragged with a ghost and dropped on the trash or New Layer (#435), tool flyouts opening on a right or corner press (#436), tool flyouts torn off into floating toolbars (#437), the Shape Builder's erase mode deleting edges too (#439), installed fonts registered outside the font folders and fonts installed while the app runs listed in the font menus (#443), alike circular corners on rectangles that aren't square (#442), tool cursors as OS cursors so the pointer no longer trails over the canvas (#444), artboards copied and pasted with their art (#446), guides dragged out of the rulers snapping to the art's side midpoints and the artboards, artboard guides (#451), the toolbar's double arrow and Window › Toolbars › Double Column switching between one and two columns in any window height (#455), the Smart Guides display options and Snapping Tolerance (#394), Font Size and Leading combo boxes with preset lists in the Character and Properties panels and the Control bar (#458), screenshots pasted from the clipboard on macOS, Windows and the web (#457), the Properties panel's Fill and Stroke matching the Control bar's, Illustrator's Window menu panel keys (Shift+F7 for Align…) and Alt-wheel zoom whatever the wheel preference (Discord feedback), the point/area type widget (Discord feedback), groups in legacy Illustrator .ai files (Discord feedback), a rotating desktop log file (#465), image controls in the Control bar and Properties panel: Image Trace with its presets, Mask, Crop Image, Embed/Unembed, and an Image Trace object's preset, Expand and Release (Discord feedback), the Pen working as the Anchor Point tool with Alt held over a selected path's handles and anchors, and anchor buttons that convert, connect and cut paths at the selected anchors (Discord feedback), Japanese hanging punctuation (burasagari, #463), Escape leaving Presentation Mode with no screen mode saved across restarts, documents dropped on the desktop app opening as tabs of their own and a note that an `.ai` file's art off its artboards isn't in the part read (#472), Add/Delete Anchor Point tool icons drawn from the Pen tool's (#479), EPS import running more of PostScript Level 3 (patterns, Type 3 fonts, mesh shadings, masked images, VM and resource probes), and a preview that stands in saying why (Discord feedback, #466, #469, #474), chords with modifiers recorded in Edit › Keyboard Shortcuts (#487), Type Object Selection by Path Only (#394), Anti-aliased Artwork (#394), Show Pixel Grid (#394), pen and touch taps pressed and lifted within one frame clicking where they tapped and the Linux/Wayland pen-tablet workaround (#491), bounding-box resizes landing on the other objects' anchors and centres with labels, and the Rotate, Scale, Shear and Reflect tools snapping their drags (#482), Shift-drag marquees that toggle the objects or anchors they reach and a Lasso whose Alt subtraction keeps the rest of the selection (#483), the mouse wheel stepping the focused numeric field, the dialogs' and preferences' drag values made such fields (#485), Show The Home Screen When No Documents Are Open (#394), Numbers Without Units Are Points (#394), Auto-Collapse Iconic Panels (#394), the Contextual Task Bar moved by its handle and its More Options menu pinning and resetting its position, a designed macOS DMG windoLine truncated
 
 ## Where we are
 
@@ -166,7 +166,7 @@ grade by behaviour against `plan/illustrator/`, not by whether a menu item exist
 - **Fonts:** a font is found by any of its names (family and style in every language of its name table, legacy family/style pairs, PostScript name), so documents from other apps open with the font they name; Find Font and `text.fonts` say for each font whether it was found exactly, by another name, substituted or missing, and how many glyphs are missing. PDF and `.ai` import brings Japanese text back as live type: CID fonts without a ToUnicode map read through the standard CMaps, vertical columns (WMode 1) as vertical type, letter spacing as tracking; a glyph is live type only when the installed font (in the closest style of its family) draws the same glyph, and text in a font that isn't installed comes in as live type shown in the fallback font until the font is installed (exports that draw it — PDF, EPS, EMF/WMF, raster, SVG outlines or embedded fonts — warn that they wrote the fallback font; `textAs: outlines` keeps the PDF's embedded glyphs). Two installed versions of one family and style (a font library keeping an old file beside a new one) both load, the first staying the one the family and style resolve to; PDF and `.ai` import tries each version when the glyphs differ, and type set in a version other than that one names it (`CharStyle::font_version`), which layout, the substituted-glyph highlight, Find Font, Package and Document Info follow. Fonts outside the font folders are listed too: those font services such as Adobe Fonts load on Windows, through DirectWrite, and those apps and font managers register with CoreText on macOS (#579).
 - **Files (M4):**
   - `.vectorcraft` (lossless JSON, compressed, atomic saves, save down to v1), Save As with a format chooser, Save a Copy, Revert, templates, Data Recovery after a crash (browser tabs hold their copies with Web Locks too, so a background tab's copies are never taken for a crash's, and a tab repairs copies another removed when it resumes: #367), background save and export. The desktop app writes its warnings and errors (graphics adapter, a lost graphics device, panics the guard recovered from) to standard error and to `logs/vectorcraft.log` next to the preferences, the two previous runs kept and each file capped at 16 MiB, so a launch without a terminal still leaves something to attach to a bug report (#465).
-  - SVG/SVGZ in and out with SVG Options, Preserve Editing, symbols, filters, rich text and physical units (SVG type under a transform or `viewBox` scale opens at the size it draws at, the scale moved into its size, leading, baseline shift and strokes, a stretch, skew or rotation kept as its transform: #396; a top-level group with a clip path, such as a clipped Inkscape layer, opens as a layer with a clipping mask, its name kept: #420; hidden objects written hidden on Save and `convert`, and `data-*` attributes kept as the objects' own data: #522; top-level groups open as layers named from `data-name`, else the id, also when one more group (a transform on all the art, a `<switch>`) wraps them, hidden ones hidden, and a file whose art sits loose under a root `<svg>` with an id (or `data-name`), as one-layer exports write it, opens with its one layer named from it; nested groups stay groups, except the sublayers VectorCraft writes, marked `data-vc-layer`, and Inkscape sublayers, which open as sublayers in their layer's colour: #372); PDF and PDF-compatible `.ai` in and out (layers, with hidden ones kept hidden with their art and nested ones as sublayers, on pages of up to 64 million operators, art whose layer can't be told going to a hidden, non-printing layer rather than showing or printing: #419; strokes kept live, a fill and a stroke of one outline as one object; a type object's stroke written as each glyph's outline stroked back on the type, so it stays one editable object; lines wrapped in a frame (each wrapped line ending in a space, one left edge, leading and font) back as one area type object, left aligned or justified with its paragraph spacing, where laying it out breaks the lines where the file does; glyphs set apart by gaps wider than a space (a barcode's digit groups) keeping their places; layers stacked as the pages mark them even when one first paints on a later page, art ending a layer's marked content or a clip closing in a later layer kept in its own layer, and layers with no art kept empty: #508; masks, editable text with type set along a curve coming back as type on a path, each artboard's art on its own page only, an `.ai`'s white page under its layers left out rather than coming in as art, spot colours, presets, security, marks and bleed, ICC output intents, raster effects, subset fonts, PDF/X, PDF layers, overprint, page thumbnails, fast web view, PDF 1.3). CMYK images keep their ink amounts from PDF import (and from placed CMYK TIFFs) to PDF export: a CMYK JPEG passes through unchanged, others stay CMYK (DeviceCMYK, or ICC-based CMYK in PDF/X-3/X-4) unless the Output settings convert them; the canvas shows them through the working CMYK profile (#320). A masked CMYK image (`/SMask`, `/Mask` or a colour key) keeps its inks too, its alpha coming in as a greyscale opacity mask.
+  - SVG/SVGZ in and out with SVG Options, Preserve Editing, symbols, filters, rich text and physical units (SVG type under a transform or `viewBox` scale opens at the size it draws at, the scale moved into its size, leading, baseline shift and strokes, a stretch, skew or rotation kept as its transform: #396; a top-level group with a clip path, such as a clipped Inkscape layer, opens as a layer with a clipping mask, its name kept: #420; hidden objects written hidden on Save and `convert`, and `data-*` attributes kept as the objects' own data: #522; top-level groups open as layers named from `data-name`, else the id, also when one more group (a transform on all the art, a `<switch>`) wraps them, hidden ones hidden, and a file whose art sits loose under a root `<svg>` with an id (or `data-name`), as one-layer exports write it, opens with its one layer named from it; nested groups stay groups, except the sublayers VectorCraft writes, marked `data-vc-layer`, and Inkscape sublayers, which open as sublayers in their layer's colour: #372); PDF and PDF-compatible `.ai` in and out (layers, with hidden ones kept hidden with their art and nested ones as sublayers, on pages of up to 64 million operators, art whose layer can't be told going to a hidden, non-printing layer rather than showing or printing: #419; strokes kept live, a fill and a stroke of one outline as one object; a type object's stroke written as each glyph's outline stroked back on the type, so it stays one editable object; lines wrapped in a frame (each wrapped line ending in a space, one left edge, leading and font) back as one area type object, left aligned or justified with its paragraph spacing, where laying it out breaks the lines where the file does; glyphs set apart by gaps wider than a space (a barcode's digit groups) keeping their places; layers stacked as the pages mark them even when one first paints on a later page, art ending a layer's marked content or a clip closing in a later layer kept in its own layer, and layers with no art kept empty: #508; art reaching past a plain PDF's page clipped to its crop box, as PDF viewers show it, each such object in a clip group of its own (a PDF-compatible `.ai` opens whole: #675); masks, editable text with type set along a curve coming back as type on a path, each artboard's art on its own page only, an `.ai`'s white page under its layers left out rather than coming in as art, spot colours, presets, security, marks and bleed, ICC output intents, raster effects, subset fonts, PDF/X, PDF layers, overprint, page thumbnails, fast web view, PDF 1.3). CMYK images keep their ink amounts from PDF import (and from placed CMYK TIFFs) to PDF export: a CMYK JPEG passes through unchanged, others stay CMYK (DeviceCMYK, or ICC-based CMYK in PDF/X-3/X-4) unless the Output settings convert them; the canvas shows them through the working CMYK profile (#320). A masked CMYK image (`/SMask`, `/Mask` or a colour key) keeps its inks too, its alpha coming in as a greyscale opacity mask.
   - EPS and DXF in and out, EMF/WMF in and out; PNG (with PNG-8), JPEG, WebP, GIF, TIFF, BMP, Targa and layered PSD export. EPS import runs a PostScript Level 3 interpreter written from the PostScript Language Reference: resource categories and `resourceforall`, local and global VM (`currentglobal`/`setglobal`/`gcheck`), `clipsave`/`cliprestore`, subarrays and substrings that share storage, the standard dictionaries in `systemdict`, `dictstack`, `internaldict`, executable filters and strings, `load` by any key, `//name` immediate names, `token`, `execform`, user paths, `cshow`, `rootfont`, `cachestatus`, colour rendering and halftone queries, file `status` and data read with `flushfile`, and `pdfmark` marks dropped (Discord feedback, #466, #469, #474). Masked images (`ImageType 3` in its three interleavings) and reusable streams are read, and an image's several data sources are read in turn, as the reference defines (procedures that read one file get their own rows: such images came in banded, #498). `restore` puts dictionaries back as they were at the `save`, so programs that count their saves in a dictionary read to their end (part of #505). Tiling patterns come in as pattern swatches, uncoloured ones too; Type 3 fonts' glyphs as their outlines, spaced by `ashow`/`widthshow`/`xshow`/`kshow`; mesh and function-based shadings (types 1, 4–7) as gradient meshes; sampled functions (type 0) and Flate/LZW predictors are read. It is tested with programs written from the specification in the ways common generators use these features; Inkscape/cairo EPS files open as vectors. When a file's PostScript still can't be read, it opens as its TIFF preview, its Windows metafile preview (as vectors) or its EPSI preview, and the warning (in the status bar, `document.open`/`app.open`/MCP `open_file` replies and `vectorcraft-cli info`) names the error, the operator that raised it and the procedures it ran in. Legacy PostScript `.ai` files (Illustrator 3–8, written with the prolog that defines their operators) keep their groups and nested groups (`u` … `U`), between and inside clipping groups; EPS and EMF/WMF import nest clips (and groups) at most 128 deep, so a hostile file can't exhaust the stack. DXF TEXT justified Fit or Aligned spans its two points (stretched, or scaled as a whole). PDF export writes type as real, searchable text in embedded subset fonts by default, as Illustrator's PDF presets do (Convert Text to Outlines is an option; printer's marks stay shapes), so text-heavy PDFs stay small (#733). DXF export writes each fill as its outlines followed by its hatch, so laser and cutter software, which skips hatches, gets filled art and type too (#546).
   - Affinity documents in (AF.1–AF.4): `.af` (Affinity 3) and `.afdesign`/`.afphoto`/`.afpub` (Affinity 1 and 2, container versions 8–12) open and place as editable art: layers and sublayers, groups (pass-through or isolated), artboards (with their names and backgrounds) and Publisher pages side by side, curves with live corners, parametric shapes (rectangles with corner radii, ellipses, polygons, stars, pies, triangles, trapezoids), compound shapes, solid/linear/elliptical gradient fills and strokes (weight, dashes, alignment, scale-with-object), clipping by nesting, vector and pixel masks, artistic and frame text with fonts, sizes, tracking, leading and alignment, placed images (their original JPEG/PNG) and pixel layers (8/16-bit RGB, CMYK), opacity, visibility, locking and blend modes, at the document's resolution in points. Effects, adjustments, brushes and pressure profiles, transparency gradients, master pages, symbol instances (imported as their cached picture), special shapes (cloud, heart, cog…, imported as ellipses) and Affinity-only blend modes are named in the import warning. A file whose native data can't be read opens as its embedded preview with the reason, and that preview can't be placed. Save never writes back over the Affinity file and there is no Affinity export. AF.4 adds Affinity 3 artboard recognition and actual-outline clipping, source-backed JPEG/PNG raster tiles, Unicode text-run recovery, stored fallback/exact installed fonts, OpenType feature flags and compound-geometry limits. 51 public CC0/MIT/Apache-2.0 files (33 current `.af`), pinned and sha256-verified (`cargo xtask corpus --affinity`), test native retention and every-board native/SVG/PDF/PSD output for current files; selected files compare their saved thumbnails. No Affinity reopen or independent paired PSD oracle was available; [source audit and complex-font gaps](docs/affinity-validation.md); [scope, provenance and limits](crates/affinity/README.md).
   - Placed documents: File › Place of a `.vectorcraft` file links to it by default (`file.place`'s `link`): one locked object showing an artboard of the document (or its art's bounds), read the way placing it without Link reads it, its symbols, patterns and images kept apart from the document's; without Link (or from bytes, not a file) it is an editable copy as before. It stays vectors in every output (SVG, PDF, EPS, EMF, DXF), and so do placed documents inside it, read from their own files up to 8 levels deep (documents that place each other stop there). On screen they draw from bitmaps cached per file and zoom step, shared by every copy and made in the background (a placeholder until ready), so a sheet of them redraws in about a millisecond; exports and prints draw them exactly. Like linked images, a save keeps only a preview of each (JPEG when opaque) unless Include Linked Files; output, and zooming past the preview, read the file again, falling back to the preview with a warning when it is gone or changed. Saves for older versions write them as plain art. Links check, update, relink, Update Links on open, Link Info (the artboard's size), Placement Options, Break Link (also Embed, Object › Expand: the editable copy placing without Link gives, its resources joining the document), Edit Original (the document in a new tab) and the Links panel cover them; saving a document updates the open documents that place it. The native format (v4) keeps data of 8 KB or more after the JSON instead of base64 in it.

crates/pdf/src/import.rs

@@ -135,6 +135,9 @@ fn group_layer(
 pub const OFF_ARTBOARD_NOTE: &str =
     "only the PDF-compatible part of this file was read: art outside its artboards is kept in the editor's private data alone, so it doesn't open";
 
+/// Warning: art reaching past the page box was clipped to it.
+pub(crate) const PAST_PAGE_NOTE: &str = "art reaching past the page was clipped to it, as PDF viewers show it";
+
 /// Import a PDF, returning the document plus warnings about content that was approximated or skipped.
 pub fn import_with_report(bytes: &[u8], opts: &ImportOptions) -> Result<ImportReport, PdfError> {
     let original = crate::pages::open(bytes, opts.password.as_deref())?;
@@ -227,6 +230,29 @@ pub fn import_with_report(bytes: &[u8], opts: &ImportOptions) -> Result<ImportRe
             }
             parts.retain(|(_, art)| !art.is_empty());
         }
+        // Viewers show a page through its crop box, so art reaching past it (a mock-up backdrop
+        // around the page, say) never showed. A clip to the page box that the file states is
+        // dropped as redundant (by hayro and by `push_clip_path`), so restore the box here:
+        // each object reaching past it keeps its own clip, which Fit to Artwork Bounds and
+        // `CropTo::Bounding` then respect.
+        // The crop box is what viewers show, whichever box the artboard takes. Not in a
+        // PDF-compatible `.ai`: its PDF part stands in for the editable document, whose art
+        // crossing the artboard edge isn't clipped, so it opens whole (#646).
+        let shown = xf.transform_rect_bbox(crate::pages::page_box(page, CropTo::Crop));
+        let page_box = shown.inflate(0.01, 0.01);
+        let mut clipped = false;
+        for (_, art) in parts.iter_mut().filter(|_| !ai) {
+            for n in art.iter_mut() {
+                if n.visual_bounds().is_some_and(|r| !contains(page_box, r)) {
+                    let clip = Arc::new(b.clip_node(&shown.to_path(0.1), FillRule::NonZero));
+                    *n = Arc::new(Node::new(b.id(), NodeKind::Group { children: vec![clip, n.clone()], clip: true }));
+                    clipped = true;
+                }
+            }
+        }
+        if clipped {
+            b.warn(PAST_PAGE_NOTE);
+        }
         let children: Vec<Arc<Node>> = parts.iter().flat_map(|(_, v)| v.iter().cloned()).collect();
         placeholder &= ai && only_text(&children);
         let mut right = ab.x1;

crates/pdf/src/tests_import_fidelity.rs

@@ -446,3 +446,21 @@ fn a_stroke_under_a_stretched_transform_stays_a_stroke() {
     let b = objs[0].geometric_bounds().unwrap();
     assert!((b.x0 - 20.0).abs() < 1e-6 && (b.x1 - 80.0).abs() < 1e-6, "{b:?}");
 }
+
+#[test]
+fn art_past_the_page_clip_keeps_the_clip() {
+    // A page-sized clip, then a backdrop far larger than the page and a square on it: the square
+    // needs no clip, the backdrop keeps it, so the art's bounds stay on the page.
+    let r = open_with(&one_page("0 0 100 100 re W n 0 g -500 -500 1100 1100 re f 1 0 0 rg 10 10 20 20 re f", "", &[]), |_| {});
+    let d = &r.document;
+    let art = d.art_bounds().expect("art");
+    assert!(art.x0 >= -0.01 && art.y0 >= -0.01 && art.x1 <= 100.01 && art.y1 <= 100.01, "the art stays on the page: {art:?}");
+    let clips: Vec<Node> = all(d).into_iter().filter(|n| matches!(n.kind, NodeKind::Group { clip: true, .. })).collect();
+    let [c] = clips.as_slice() else { panic!("one clip group, the backdrop's: {clips:?}") };
+    assert_eq!(c.children().map(Vec::len), Some(2), "the clip path and the backdrop");
+    assert!(r.warnings.iter().any(|w| w == crate::import::PAST_PAGE_NOTE), "{:?}", r.warnings);
+    // Art inside the page gets no clip and no warning.
+    let r = open_with(&one_page("0 0 100 100 re W n 1 0 0 rg 10 10 20 20 re f", "", &[]), |_| {});
+    assert!(!all(&r.document).iter().any(|n| matches!(n.kind, NodeKind::Group { clip: true, .. })));
+    assert!(!r.warnings.iter().any(|w| w == crate::import::PAST_PAGE_NOTE), "{:?}", r.warnings);
+}

crates/pdf/src/tests_import_layers.rs

@@ -367,6 +367,25 @@ fn art_past_the_scan_limit_of_shown_layers_stays_shown() {
     assert!(r.warnings.iter().any(|w| w.contains("went to its page's layer")), "{:?}", r.warnings);
 }
 
+/// Art reaching past the page keeps a clip to it in a plain PDF (as viewers show it), but not in
+/// a PDF-compatible `.ai`, whose art crossing the artboard edge opens whole and editable (#646).
+#[test]
+fn art_past_the_page_is_clipped_in_a_pdf_but_not_in_an_ai_file() {
+    for ai in [false, true] {
+        let page = PdfPage {
+            entries: if ai { "/PieceInfo << /Illustrator << /Private << /AIPrivateData1 7 /NumBlock 1 >> >> >>".into() } else { String::new() },
+            ..PdfPage::new(100.0, 100.0, "0 0 100 100 re W n 1 0 0 rg 50 50 200 200 re f")
+        };
+        let r = import_with_report(&pdf_with_catalog(&[page], &[], "", None), &ImportOptions::default()).unwrap();
+        let mut clips = 0;
+        r.document.walk(|n| clips += usize::from(matches!(n.kind, NodeKind::Group { clip: true, .. })));
+        let past = r.warnings.iter().any(|w| w == crate::import::PAST_PAGE_NOTE);
+        assert_eq!((clips, past), if ai { (0, false) } else { (1, true) }, "ai: {ai}: {:?}", r.warnings);
+        let art = r.document.art_bounds().unwrap();
+        assert_eq!(art.x1 > 100.5, ai, "ai: {ai}: {art:?}");
+    }
+}
+
 /// The editor's private data of a `.ai` is its `AIPrivateData` streams, joined in order.
 #[test]
 fn an_ai_files_private_data_is_its_streams_joined() {