The PDF print spike#

Superseded

This note records the experiment that chose how Scenet prints, as it was written before the print emitter existed. The decision it reached, a PDF drawn from Panel Core and not converted from the SVG, is the one that shipped. Its details are not: what print does is in Print a comic and in Design decisions, which win wherever this note differs. Three differences matter:

  • The PDF carries no date. The ReportLab column below writes a fixed CreationDate; the shipped writer writes none, and takes its /ID from its content.

  • Greys are tinted on black. The prototype lettered a plain cyan tint; a grey now prints on the black plate at the percentage that prints its own lightness under the printing condition.

  • The ink limit is kept by the emitter. The prototype left ink limits to manual prepress; the night foreground now prints C40 K100, within any printer’s limit, and a tinted palette prints its CMYK twins.

Decision#

Keep PDF emission as a third emitter over resolved Panel Core. Do not convert the SVG: the SVG has already discarded the semantic distinction between lettering, balloon borders, figure outlines and large dark masses, and it cannot express PDF overprint. The existing Page Core supplies page dimensions and resolved frames; trim and bleed should be added to page geometry rather than inferred by the emitter.

This note records the design decision and prototype findings, not a claim of PDF/X conformance. A real PDF/X claim requires a PDF/X-capable preflight run and recorded evidence.

Outcome: the comparison on a real panel#

The prototype below drew a hand-written fixture. The comparison the plan called for was run on a real document, the gallery’s 24-page.scene.yaml (a night page of three panels with lettering, a mass at every depth and no supplied picture), by drawing the same marks (scenet.emit.marks, with the same inks and overprint) two ways: through ReportLab 5.1.0, and through the dependency-free writer that became scenet.emit.pdf.

ReportLab 5.1.0

Direct writer

Dependencies

ReportLab, Pillow, charset-normalizer: an optional extra, and vendored into the playground to show the PDF there

None

Code over the marks

146 lines, without the even-odd inset holes or the clearance rings

scenet.emit.pdf, page composition included

Bytes

136,269

132,994

Two runs, two hash seeds

Byte-identical, with rl_config.invariant = 1

Byte-identical

Dates

A fixed CreationDate (the invariant date)

None

Overprint

setFillOverprint, setStrokeOverprint and setOverprintMask per paint; a stroke painted under its fill needs two drawing calls

Fill and stroke apart in one ExtGState, written directly

Both are deterministic and both control overprint. The direct writer was kept: it adds no dependency, so the PDF exists wherever the compiler runs, the browser playground included, and every byte is the project’s own. It is scenet.emit.pdf, and Print a comic says how to use it.

Candidate comparison#

Concern

ReportLab (optional print extra)

Small direct PDF writer

Licence/dependencies

BSD; acceptable under the current licence gate, but adds a runtime dependency outside the base install

No runtime dependency

CMYK and separations

Has colour-space and overprint APIs; mature PDF object handling

Must correctly write colour spaces, graphics state, page boxes, metadata and cross-reference data

Overprint

Supported with graphics-state controls; still needs inspection of the emitted PDF

Can emit /OP, /op and /OPM directly, but correctness is our responsibility

Geometry

Path API is straightforward; map each Core primitive to paths

Core primitives map to PDF path operators, but every operator and number format is ours

Determinism

Requires controlling metadata, document IDs, object ordering and compression behaviour; verify byte identity experimentally

Easier to control object ordering and metadata, but byte identity still needs tests for every serialization choice

Maintenance risk

Lower PDF-format risk; higher dependency and API coupling

No dependency, but larger security/correctness burden for a format with complex object rules

Atmosphere and conformance#

The SVG atmosphere veil uses feTurbulence, which has no direct PDF vector equivalent. Do not pretend that it is reproduced faithfully. For the first implementation, omit the veil in print and document the limitation; keep already-resolved vector streaks and flecks. Revisit rasterising the veil only if a deterministic, licence-compatible raster path is available.

Target a constrained PDF suitable for print first, but do not label output PDF/X-1a or PDF/X-4 until a preflight tool validates it. PDF/X-1a would require CMYK/spot colours and flattened transparency; PDF/X-4 permits live transparency and requires an output intent. The first release should state its exact guarantees and limitations in the CLI and print how-to.

Runnable prototype and findings#

The first prototype, a script drawing one hand-written fixture with both backends, is in the project’s history; scenet.emit.pdf replaced it. Its findings, kept as they were recorded:

The prototype comparison produced valid, readable one-page PDFs from both backends. pdfinfo reported PDF 1.4 and Letter-size media pages; pdftotext extracted the fixture lettering; Ghostscript parsed both with -sDEVICE=nullpage without errors. Both PDFs exposed the expected TrimBox (9, 9, 603, 783), BleedBox (0, 0, 612, 792), CMYK operators and fill/stroke overprint state. Two repeated emissions from each backend were byte-identical in the prototype run. Direct writer output was 1,111 bytes; ReportLab output was 1,568 bytes. These figures describe the fixed fixture, not a production renderer.

Black-plate check (manual). To verify that lettering lands on K alone and that colour plates beneath it survive without knockout, run Ghostscript’s tiffsep device:

gs -dNOPAUSE -dBATCH -sDEVICE=tiffsep \
   -sOutputFile=/tmp/spike-%d.tif \
   /tmp/scenet-direct.pdf

In Ghostscript’s tiffsep device, %d formats the page number rather than the plate. For a one-page document, Ghostscript writes the composite as /tmp/spike-1.tif and writes one separation TIFF per colorant named after the separation: /tmp/spike-1(Cyan).tif, /tmp/spike-1(Magenta).tif, /tmp/spike-1(Yellow).tif, and /tmp/spike-1(Black).tif. (Alternatively, -sOutputFile=/tmp/spike.tif produces /tmp/spike(Cyan).tif, /tmp/spike(Black).tif, etc.).

Tested with Ghostscript 10.04.0:

  • The fixture positions 100% K lettering (0 0 0 1 k) over a tint rectangle with no K (a plain C40 tint, 0.4 0 0 0 k).

  • Cyan plate (spike-1(Cyan).tif): Shows the solid 40% cyan rectangular tint block with no lettering-shaped knockout hole beneath the lettering, proving that /OP true and /OPM 1 overprinted rather than knocked out the underlying cyan channel.

  • Black plate (spike-1(Black).tif): Shows the lettering “SCENET PRINT SPIKE” and the bounding box border in 100% black over an uninked background, confirming the text lands cleanly on K alone.

  • Magenta (spike-1(Magenta).tif) and Yellow (spike-1(Yellow).tif) plates: Completely blank (0% ink), confirming no unintended ink contamination or colour conversion occurred.

  • Composite (spike-1.tif): Displays the black lettering and border cleanly over the cyan tint.

This was a structural/parser check, not PDF/X preflight. No PDF/X-capable preflight tool is available in the execution environment, so conformance, output intent, printer-specific ink limits, and colour separations still require manual prepress validation. The fixture uses a built-in Helvetica face and live text for simplicity; it does not yet exercise Scenet Panel Core glyph outlines or its full range of drawing primitives.

How it ended#

Every step the spike planned was taken: the comparison on a real page (above), tests first, the emitter beside emit/svg.py, scenet build --pdf, the print how-to and the determinism documentation. ReportLab was not added, so no licence notice changed, and MCP is unchanged. The manual check became a test: Ghostscript separates every print example in CI, and qpdf reads every gallery PDF. No PDF/X conformance is claimed: that needs a preflight tool, and an output intent, which are their own issue.