Render: an axial shading in a transparency group paints nothing, where four engines paint a gradient

[yfedoseev]

Pre-existing — wrong in both v0.3.77 and v0.3.78, in opposite directions. Found by adjudicating all 3214 rendered pages of the release sweep against the reference panel.

What

pdfjs/test/pdfs/issue14165.pdf, page 0 (A3 landscape, 1190.5 x 841.9), mean RGB at 72 dpi:

engine	mean
MuPDF	192.90
pdfium	193.42
poppler	193.19
Ghostscript	192.38
v0.3.77	0.11 — the page is painted solid black
v0.3.78	255.00 — the page is painted nothing at all

The panel agrees within 1.05 grey levels on ~193, and MuPDF resolves 39 distinct grey levels spanning 119-255, so the correct page carries a visible gradient across most of its area.

We render it as one flat extreme in both releases: fully black before, fully blank now.

Direction of travel

v0.3.78 is 62 grey levels from the panel where v0.3.77 was 193, so this release moves substantially toward correct — but "blank" is not a gradient, and no part of the intended content is drawn. It is a different failure, not a partial fix.

What the file uses

/ShadingType 2      (axial)
/S /Transparency    (transparency group)
/BM /Normal
/CS /DeviceRGB

No image, no soft mask, no exotic colour space — an axial shading painted inside a plain transparency group with the default blend mode. That combination is the whole reproducer; the file is 19,588 bytes.

Worth noting page.get_drawings() reports 0 paths under MuPDF, so the gradient is being painted by the shading operator itself rather than as a filled path — a sh inside the group, not a pattern fill.

Why the sweep nearly filed this as a fix

Ranked by change magnitude this is the single largest mover in the corpus (254.89 grey levels), and going from solid black to blank looks like a repair if you only read the delta. It is only when the panel is consulted that both arms turn out to be wrong. This is the case the project guidance about adjudicating render deltas against a reference is describing — a coverage drop is not automatically a fix any more than it is automatically a regression.

Related colour/shading family: #1287, #1288, #1290, #1306, #1278. None of them covers this document or the transparency-group interaction.

[yfedoseev]

Re-milestoned to v0.3.78: this is a regression against v0.3.77, not only a pre-existing defect. The 3214-page render sweep measures issue14165.pdf p0 at nonwhite_frac 1.00000 in v0.3.77 and 0.00000 at v0.3.78 HEAD, against poppler 0.85629 and PyMuPDF 0.85986. v0.3.77 over-painted the gradient as a solid slab; HEAD paints nothing at all, so the content is now invisible rather than wrong. Mean-RGB distance from v0.3.77 is 254.9 — the largest single change in the sweep.

[yfedoseev]

Fixed in 8e29d375 — fix(render): paint a shading's whole ramp, not a blend of its two ends.

Root-caused by bisect to 6ceb2e32 (the function-evaluation change), but that change was right; this is its second half. The shading's /Function is a type 0 sampled function of 5120 samples that ramps white → near-black (min 10) → white. Resolving only the two ends of /Domain gives white at both, so the two-stop gradient was white-to-white and the page painted flat white. Before that commit, reading /C0//C1 off a type 0 dictionary found nothing and fell through to the black-to-white safety net — which is why the page was a near-black slab in v0.3.77 and blank after.

ISO 32000-1:2008 §8.7.4.5.3 gives the colour at parametric distance t as the function evaluated at t, so the resolver now samples the ramp at 33 points across /Domain and both the axial and radial backends build their gradient from that list.

Measured on this page at 72 dpi:

	mean RGB	coverage
v0.3.77	0.11	1.00000
before this fix	255 (blank)	0.00000
after	193.57	0.98810
poppler	193.19	0.85629
PyMuPDF	192.90	0.85986

Within 0.7 of both references on mean RGB, against a v0.3.77 value of 0.11. Coverage saturates on a full-page gradient and does not separate these — the colour does.

Corpus-wide: the fix touches 59 of 3214 pages; adjudicated against PyMuPDF, 14 better, 25 same, and the single page scored worse (04085752…) has identical coverage before and after (0.70166 → 0.70166) — it is the pre-existing #1341, untouched here. Text/markdown/HTML are byte-identical across all 2008 documents (0 files differ). 6096 library tests pass, clippy clean under --features rendering.

Fixture: a three-sample white/black/white ramp whose two ends agree and whose middle does not. It fails when the sample count is put back to two; the three fixtures from the function-evaluation change pass either way.