Extrude & Bevel, Inflate and Rotate decided whether a fill, stroke or
member casts part of the shadow with `opacity > 0.0` and skipped it with
`opacity <= 0.0`. A NaN opacity failed both tests, so the part was
neither cast nor skipped: the shadow cast from the solids drawn added its
faces with a NaN ink, giving the shadow fill a NaN opacity, and the
shadow built for itself recorded an empty part.
Both paths now use one rule, `shadow::Ink::casts`: art casts only when
its opacity is a finite number above zero. NaN and infinite opacity cast
nothing, as 0 % does, and the art beside it still casts. Finite opacity
casts as before.
The measurement that runs before the style cascade and usvg now covers
the inputs that still got past it:
- usvg matches a selector without keeping what it found, so one whose
leftmost part matches nothing tests every set of ancestors in turn.
The measurement counts those tests for each copy and stops at
50,000,000; past that the SVG is refused with "the SVG's style sheets
take too long to match". An SVG image's matching stops at what the
SVG showing it left.
- A test counts one more for every 16 attributes a lookup scans and
every 16 bytes of the value it compares, so many classes against a
long class list count the work they do, in usvg's matching and in
the import's own cascade.
- The declarations the import's own cascade keeps, once per element,
count toward the 64 MB, including those of `:not()` rules, which usvg
drops. Each `:not()` alternative counts as a test.
- A linked SVG file is measured before its cascade runs. The linked SVG
files share one measurement: once they pass the limits together, the
files after that are shown without their links. The tests a file past
the limits made still count.
- SVG images share the element limit and the selector tests with the
SVG showing them, as they already shared the 64 MB.
The svg_filters reader measures its SVG before the cascade too.
#7569 changed the Extrude & Bevel and Inflate check so that a filled
shape's solid is built once, about the bounds the effect draws it
about, and stated that a filled shape whose solid builds about those
bounds but not about its profile's own bounds now passes. No test
covered that case.
The new test builds a group of a filled shape of 900 small curves and
a line with a 2000 pt stroke. About the shapes' own bounds, the filled
shape's solid is too complex with a repeated rounded bevel; about the
bounds that take in the stroke, its curves are sampled more coarsely
and it fits. The test asserts the first result, then that the check
passes and the solid draws. It fails when the check builds every
shape's profile about the profiles' bounds again.
With Noto Sans Arabic the fatha's ink reaches past the first letter's
left edge, so that letter's Tsume only narrows its advance and moves no
glyph. The test compared glyph positions alone and found no change.
It now compares each glyph's position and advance.
perspective.attach, perspective.release, perspective.move and
perspective.transform took their `ids` as given, so an id listed twice
was processed twice. Attach warped an object that was already projected
a second time, and a move or transform of an object not yet in
perspective attached it and then moved it twice. The Shift and Snap to
Grid reads of a move also used a document attached twice. The ids are
now taken once each, in the order first given.