[pierrevbc]

Problem

docs/gaps.md lists "PDF export rasterizes slide artwork" as a beta blocker: each slide is one
raster image at print resolution under an invisible text layer. Files are large (0.1 to 2 MB a
slide at the default 200 dpi) and blurry when zoomed or printed.

What changes

Four commits, rebased on main at 9859da3 (after the shared PDF writer, #708): the export, the
progress docs, tests for the drawing main gained since this PR was opened (pattern fills,
rectangle and shape path gradients, fill overlays, compound outlines, text shadows and glows,
gradient text, colour emoji, WMF/EMF pictures, WordArt warps), and tests for the newer bevels,
lit faces, group shadows, circle gradients, picture tiles and upright vertical text.

Slides are drawn by the same renderer through a vector canvas, so shapes, outlines (width, dashes,
caps, joins, arrowheads), solid and linear/radial gradient fills, transparency, rotation, flips,
groups, tables and charts become PDF paths and paints, and text is real text in the embedded font
subsets (searchable, selectable, mapped back to Unicode cluster by cluster). JPEG and PNG pictures
are embedded as they are (JPEG stays DCT), others as lossless pixels; crops and picture
transparency are kept. Backgrounds (solid, gradient, picture) are vector too.

What a PDF can't express is rasterized locally, at dpi, in place, never dropped: a shape with an
outer or inner shadow, glow, soft edges, reflection or a fill overlay (its geometry and effects;
its text stays vector), pattern fills, rectangle and shape path gradients, text shadows and glows
(over the glyphs' area; the text itself stays real text), colour emoji (over the glyph's em box;
the letters around stay text) and WMF/EMF pictures (decoded to pixels, as on screen, at print
resolution). Gradient text is drawn glyph by glyph as vector paths; compound outlines, arrowheads
and tables with file-defined styles are vector. Pictures the renderer can't decode show the
placeholder they show on screen. Glyphs from a variable font's non-default named instance are
drawn as outlines over invisible real text, since the embedded subset is the default instance.

vector: false (file.export, MCP export) keeps the raster mode, byte for byte. Links,
bookmarks, notes pages, handouts and hidden-slide handling are unchanged.

By crate:

• render (canvas.rs, lib.rs, paint.rs, table.rs, chart.rs, warp.rs,
  images.rs, colr.rs): a Canvas trait for every drawing call; vello_cpu's RenderContext
  implements it, so on-screen output is unchanged. draw_slide draws a slide on any canvas
  (shapes, backgrounds, tables, charts, text, warps). canvas::pixels runs drawing that only
  exists in pixels directly on a pixel canvas and rasterizes it locally on a vector one. Shapes
  with effects or a fill overlay are rasterized whole over the area extent::spill computes,
  with the view and a useBgFill background shifted so strokes stay aligned. paint::gradient
  returns a Paint (gradient text no longer needs a paint transform, which vector canvases
  lack); colr::has_color tells colour glyphs apart. Pictures hand the canvas the original file
  when it is an unchanged JPEG or PNG.
• text: TextLayout::texts, each paragraph's characters as laid out (fields filled in), for
  glyph-to-text mapping.
• pdf: krilla canvas (vector.rs, on craft_pdf::write::backend) with font and image
  caches; PdfOptions::vector (default true); dpi now also sets the resolution of local
  rasters. It keeps Use the shared PDF writer and report font embedding failures #708's contract: the export runs under write::guard, fonts go through
  write::font, and a font that can't be embedded fails the export with its error on vector
  slides too (never dropped or substituted text). New dependency on deckcraft-color (workspace
  crate, same layer rules); cargo xtask pdf-deps passes.
• render, drawing added on main since: chartex charts (waterfall, funnel, treemap, sunburst,
  box), radar and stock marks, chart trendlines, 3-D extrusion sides and contours take
  &mut dyn Canvas and stay vector; bevels and lit faces, path gradients and patterns go through
  canvas::pixels; the table painter's off-screen text skip uses the slide box on a vector
  canvas; a cropped picture is decoded for its whole size (Boolean intersect blurs raster image output #762) on both routes.
• engine, mcp: vector export parameter.
• Progress docs: the gap "PDF export rasterizes slide artwork" is removed from docs/gaps.md
  (top 14 now); "Export to PDF" goes from P to D in docs/parity-checklist.md (summary regenerated
  by cargo xtask parity); docs/file-format-parity.md, docs/architecture.md,
  docs/roadmap.md (share row and beta item 5 now cover printing only) and the ROADMAP.md
  progress log are updated.

How it was tested

cargo xtask ci passes on this branch, rebased on main at 9859da3 (all 7 steps: fmt, pdf-deps,
clippy -D warnings, 1,690 tests passed and 6 ignored, assets, layers, wasm32 check). On Windows
the deckcraft-ui-egui lib tests sometimes crash with an access violation when run in parallel,
on main as well; the run above used RUST_TEST_THREADS=1.

New tests, all on decks generated in code:

• crates/pdf: vector_slides_are_paths_and_text_without_a_page_image,
  vector_text_is_extractable, vector_pdf_is_much_smaller_than_raster,
  a_shadow_rasterizes_that_shape_only, pattern_fills_rasterize_their_area_only,
  jpeg_pictures_are_embedded_as_they_are (byte for byte), links_and_bookmarks_survive_vector_export
  (slides and handouts).
• crates/render: draw_slide_on_a_pixel_canvas_matches_render_slide (zero pixel difference), and
  vector_canvases_get_paths_and_text_and_effects_as_local_pixels, with a recording canvas that
  checks what reaches a vector canvas. Third commit, same recording canvas:
  vector_canvases_get_pixel_only_fills_as_local_pixels (pattern, rectangle path gradient, fill
  overlay over that element only; a triple compound outline as strokes),
  vector_canvases_keep_text_and_rasterize_its_effects (text shadow and glow as pixels, glyph
  runs kept; gradient text as paths), vector_canvases_get_metafile_pictures_as_pixels (a WMF
  built in the test), vector_canvases_get_colour_emoji_as_pixels (when a colour font is
  installed).
• crates/pdf (third commit): pixel_only_drawing_exports_as_local_images, a vector PDF with a
  pattern, a path gradient, shadowed text with an emoji and a WMF picture: no panic, local images
  only, extractable text.

Checked by hand with the built CLI on this branch: the sample deck exports to 0.14 MB vector
(1.01 MB with vector: false); its 8 pages render in poppler and pdftotext gives the slide text.
A deck generated with deckcraft-cli run (gradient, shadow and glow, pattern, bevel, chart,
table) gives 2 pages, extractable text including the chart labels, and 3 local images (the
shadowed shape, the pattern, the bevel). Earlier, on an older main: a 21-slide public deck from
10.4 MB to 0.73 MB.

• crates/pdf (Use the shared PDF writer and report font embedding failures #708 contract): unembeddable_font_returns_error_instead_of_a_pdf_missing_text
  now covers vector and raster slides; text_layer_links_and_bookmarks_survive_serialization
  covers both modes.
• crates/render/src/tests_vector.rs (fourth commit): replays the vector route into pixels and
  compares it with the pixel route (off-centre, tiled and rotWithShape=0 gradients, flipped and
  aligned tiles, bevel and lit face, group shadow, grpFill under a shadow, gradient lines and
  heads), and counts the local rasters.

Known limits

• Shadows, glow, soft edges, reflection, fill overlays, pattern fills, rectangle and shape path
  gradients, text shadows and glows, colour emoji and WMF/EMF pictures are still pixels (per
  element, at dpi).
• WordArt (warped), gradient and emoji text isn't selectable text in the PDF (paths or pixels).
• The text of a shape with effects stays vector, drawn over the shape's raster.
• Printing is still missing; the share row of docs/roadmap.md now covers only that.

Related PRs

• Use the shared PDF writer and report font embedding failures #708 (merged, shared PDF writer): this PR is rebased on it and uses its writer, guard and font
  errors; vector: false is Use the shared PDF writer and report font embedding failures #708's export unchanged.
• git merge-tree against main at 9859da3 is clean; the overlaps listed before this rebase
  were against an older main.
• Charts: every LibreOffice chart type, a data editor, and formatting for each #255 (every chart type, data and formatting) rewrites the chart painter on &mut RenderContext.
  Whichever lands second converts it to &mut dyn Canvas, as this PR does for the current
  painter and the chartex and radar painters.

[Harrison-Booth]

Thanks for this. Exporting vector slides with real text, instead of one raster image per slide, still looks valuable for file size and sharpness. I checked just now: the PR is open, GitHub reports it conflicts with main, and no one has commented on it. Do you plan to bring it up to date with main? If you would rather not, I can open a PR that credits you, keeps your commits and shows a git range-diff against this one, so you'll still get credit and have visible work.

[pierrevbc]

Thanks for picking this up. I don't plan to bring it up to date myself, so please go ahead and open the new PR with my commits and the credit. Feel free to close this one when yours is up.