Flattening counted every colorized image as transparency, in the art and
in pattern tiles. A new flattener option, keeps_colorize, stops counting
an image colorized with a solid paint whose pixels are grayscale or 1-bit,
for a format whose writer draws such images in their ink. Opacity,
blending modes, opacity masks and effects still flatten them, and pattern
tiles are still expanded for those. The option is off, so exports flatten
as before; the EPS writer that draws colorized images turns it on.
The EPS pattern tile tests now check the exported colors, and a tile
group with opacity for the warning.
Found while testing crates/doc/src/units.rs:163-165 (readout)
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Found while testing crates/doc/src/setup.rs:111-113 (Quotes::opens / is_open), used by Quotes::apply at :122-131
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
The canvas keeps a hairline visible by raising a stroke's weight to half a device pixel. A width profile multiplies that weight, so a 1 pt stroke with a 6x width point swelled to twice its size at 25% zoom and shrank back on zooming in. The minimum now goes to the profile's widest point instead; plain strokes are unchanged.
Each patch of a gradient mesh was cut into 8 x 8 flat quads for SVG export,
Object > Expand and the other outputs that cannot shade a mesh, whatever its
colours, so a patch of one colour became 64 identical pieces. Neighbouring
quads are now joined into square blocks wherever the colour changes by no
more than about one 8-bit step across the block; the block's outline runs
through its quads' corners, and a patch that folds over itself is not joined.
A mesh read from a file can have any number of patches, and the view any zoom: a 150 x 150 mesh from a 3 MB file exported at 20x asked for about 23 million quads (2 GB), and the command's own cap of 50 rows and columns does not apply to a loaded one. One tessellation now makes at most 1,048,576 quads, the subdivision per patch giving way before that, and a mesh with more patches than that is malformed, like a point count that does not fit.
When moving type or a symbol instance in perspective makes the
conjugated projection overflow, Homography::conjugated now forms the
product from factors scaled to a largest entry of 1 and returns it at
that scale, which maps every point to the same place. When even the
scaled result is not invertible, the transform removes the projection
and the art is drawn flat, instead of keeping the old projection while
the flat art moves.
paint.invert and paint.complement on the Fill proxy, the Edit Colors
commands, the Blend commands and recolor.apply now recolor the color
that colorizes a grayscale or 1-bit image, as one undo step with the
other targets. Before, they left it unchanged. Expand Appearance and
Object > Expand bake color adjustment effects into that color too.
- A colorized image's pixels are not recolored; they stay the gray
coverage of its color. A stroke-only pass reaches the image's stroke
items and leaves its colorize color unchanged.
- A command that would leave a process color on a grayscale image with
an alpha channel fails with the existing message and changes nothing.
Object > Expand undoes its earlier steps when that happens.
- A colorized image counts as a filled object in a blend.
A colorized grayscale or 1-bit image draws its white pixels transparent and is written to
PDF with the image as a soft mask. Node::shows_transparency did not count it, so a PDF 1.3,
PDF/X-1a or PDF/X-3 export whose only transparency was a colorized image was not flattened
and kept the soft mask. A legacy .ai save also wrote such an image as opaque.
- shows_transparency counts a visible image with a colorize paint.
- The legacy .ai writer flattens or leaves out colorized images like other see-through images.
Found while testing crates/doc/src/blend.rs:197-201 (lerp_dash, solid_like)
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Create Gradient Mesh's To Center / To Edge highlight lightened every point
towards RGB white, so a CMYK (or grey) point became display RGB. A patch with
any non-CMYK corner then mixes in RGB, and the PDF writer separated those
colours back to CMYK with K = 0, moving ink off the K plate. Points now
lighten towards white in their own colour model (no ink for CMYK), so the
mesh stays CMYK and K shrinks with the other inks. Flat meshes (highlight 0)
already kept their inks.
The check of a selected object's masked art looked at its own place
only. A click goes down from the layer (or the isolated container) and
never reaches a Shaper group's original shapes outside Construction
mode, however deep the object lies in them; and below a layer whose
Preview is off it hits art in outline. A press inside such a path's
fill, or on a path in a Shaper group's hidden shapes, started a move.
The way down to the selected object is now checked as a click goes:
a Shaper group's part a click can't reach on it rules the object out,
and a Preview-off layer on it (below the isolated container) makes the
object count by its outline.
Tests also cover Shift with Direct and Group Selection (a marquee) and
an object a click hits over the masked art (it is selected).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EjfPR4a1jPek7r3nyFFNYv
The selected object itself was checked for being hidden, locked or a
template, but not as a click checks it: a selected guide path was
grabbed where a mask hid it while guides were locked or hidden, though
a click on it there hits nothing.
The selected object now passes the same test a click applies to it
before its art is tried: guides while guides are locked or hidden, and
a Shaper group's hidden part, don't count.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EjfPR4a1jPek7r3nyFFNYv
A selected group was grabbed by any of its contents under the pointer
where an outer mask hides them, including a hidden or locked child and
art the group's own clipping mask leaves out: none of these is drawn or
hit by a click.
A container's art now counts as a click would hit it: its contents as
the hit test tries them, through its own mask and without its hidden,
locked or template contents. The check for locked objects shares it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EjfPR4a1jPek7r3nyFFNYv
An image in a clipping mask, picked in the Layers panel, shows its
whole frame, but a press on the part the mask hides hit nothing: the
Selection, Direct Selection and Group Selection tools started a marquee
there, which dropped the selection, so the image could be dragged only
by the part that shows.
Such a press now grabs the selection by that art, as a press on its
visible part does: a drag moves the selected object (the image, not its
clip group) and a click keeps it selected. With nothing of it selected
there, or with Shift, a press there starts a marquee as before.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EjfPR4a1jPek7r3nyFFNYv
The two nested_instances_..._culled_in_bounded_time tests in the
renderer's tests checked that a frame with 2^30 differently placed
symbol instances finishes within 30 s of wall-clock time. On a loaded
machine the frame took longer than that, and both tests failed. The
time bound also passed when the frame did far more work than it should:
with the 4,096 limit on measured placements raised to 65,536, the frame
measured symbol art 131,118 times and finished in under 2 s.
SymbolBoxes now counts how many times measure_art runs a measure on
symbol art (art_measured), and FrameStats::symbol_art holds the count
for the frame's symbol boxes. The tests assert that the frame measures
symbol art at most 20,000 times. It measures it 8,234 times, and 4,160
times when each instance has a Transform effect of its own. The wait
for the frame is 300 s, as a guard against a frame that never ends.
A new test in vectorcraft-doc checks that each symbol's art box is
measured once however many instances of it are measured.
Envelope Options stored Anti-Alias and Preserve Shape Using, but nothing that
draws an image read them: a distorted image was always smoothed and always
cut into clip groups.
- With Anti-Alias off, each piece of a distorted image is marked pixelated
(saved as `pixelated`). The renderer draws it with nearest-neighbor
sampling and without mips, SVG export writes `image-rendering: pixelated`,
and PDF export writes it without /Interpolate.
- With Preserve Shape Using Transparency, each piece is the image with a
white opacity mask in the piece's shape instead of a clip group. The
renderer paints an image whose mask is one plain white path straight into
that path, as it does for an image in a clip group, when the path lies
inside the image's frame.
Docs: docs/mcp.md, docs/target-app-parity.md.
drop_beyond_canvas removed objects but left what names them: saved
selections, Asset Export entries and variable bindings kept their ids,
and layer guides of a removed sublayer named a layer that was gone. It now
runs tidy_saved_selections, prune_assets and prune_variable_bindings when
it removes objects, as edits do, and removes the layer guides of the
layers it removed.
A layer's bounds leave out its hidden objects and a clipping container's
are its clipping path's, so objects beyond the canvas inside them were
kept. Layers and groups are now looked into object by object.
A clipping path beyond the canvas goes with the objects it clips, so no
other object becomes the clip; a top-level clipping layer is emptied and
stops clipping. A removed artboard's ruler origin goes and its id is
retired, as deleting it does.
Edits refuse to leave anything beyond MAX_COORD (doc_sane checks the
artboards, the guides, the ruler origin and the objects an edit touches),
but the native loader didn't check them. A file with one object, artboard,
guide or ruler origin out there opened without a warning, and then every
move failed with "result would exceed the canvas", in-range objects
included; saving kept it that way.
Document::drop_beyond_canvas removes objects whose geometric bounds aren't
finite or reach past MAX_COORD (a layer or non-clipping group keeps its
children within it), artboards with their artboard guides, and guides out
there, and puts the ruler origin back at the origin. load_file runs it
after the depth limit and reports what it did with the other load notes; a
file whose artboards all lie beyond the canvas is refused.
NodeKind's reader buffers the fields that come before `type` as
serde_json::Values. VectorCraft writes `type` first, but a crafted file
can put a long array there: a group whose children held 50M numbers before
its type (a 96 KB gzip) peaked at 1.6 GB, and the cost grows with the
array up to MAX_JSON, before any object in it is counted.
The buffer is read through a counting visitor: each JSON value counts
against MAX_BUFFERED_VALUES (32M, 4M on wasm32) as it is read, and values
replayed from the buffer aren't counted twice. load_file reads within it;
the same file is now refused at a 1.1 GB peak.
Opening a native document now reads at most MAX_READ_NODES objects (4M,
512K on wasm32) and refuses the file past that with a message naming the
budget. Paths already read within MAX_READ_ANCHORS and the web build's JSON
cap is 256 MB, but the number of objects had no bound: a gzip'd document of
empty groups costs about 270 bytes per object once read against some 50 in
its JSON, so 10M groups in a 24 MB file took 2.7 GB to open.
NodeWire's id is read through counted_id, so each object counts once as it
is read. load_file reads through a private load_file_within so a test can
exercise the budget without a 4M-object document.
A blend whose keys are blends draws each of its steps as a copy of a key,
and that copy makes its own steps when it is drawn, so step counts capped
per pair multiplied with nesting: five levels of 15 steps were 1.3 million
steps from a 23 KB file, 3.4 GB to export.
A blend now makes at most 100,000 steps in all (MAX_BLEND_STEPS), counting
the steps of the blends inside its keys and steps. Past the budget each pair
of keys keeps the same share of its steps; a copy of a blend spaced by
distance counts the most steps its spacing can make, since scaling it up
adds steps. Blends within the budget draw every step as before.
Review follow-ups:
- items_per_target maps each requested row on its own. Rows deleted or
duplicated together, one shared and one not, fell back to the literal
indices in every object, so the other objects could lose the wrong
items. The first object takes its own rows; each other object takes
its copies of the shared ones only.
- effect.dialog takes `aligned` (the panel passes it on a mixed
selection's shared effect row), and its OK and preview run
effect.setParams with it, so the dialog edits each object's own copy
of the effect instead of the first object's index in every object.
Tests: removing a shared stroke with an unshared fill leaves the other
object's fills; an aligned dialog's command carries `aligned`.
With several objects selected whose appearances differ in anything (two
circles whose fills differ but whose stroke and effect are the same),
the Appearance panel showed only "Mixed Appearances": no stroke, effect
or opacity row, so attributes the objects share couldn't be edited
together. It now lists what they share, as the issue describes:
- `vectorcraft_doc::appearance::SharedStack` (with `align`, a stacking-
order longest-common-subsequence match) gives the fills, strokes and
object-level effects every object has an equal copy of, and where each
is in every object's own stack.
- The panel's object row names the objects' common kind ("Path",
"Ellipse"; "Mixed Objects" only when kinds differ, as the Control bar
does), with an italic "Mixed Appearances" row under it when they
differ, then only the shared rows with their usual controls. The
Opacity row reads "Mixed" when opacity or blend mode differ (an edit
sets them all). The object row shows no fill chip when the fills
differ. Rows aren't dragged in a mixed selection: where a row would go
in each of several different stacks isn't one place.
- The engine maps a shared row to each object's own item: the active
item (a row of the first selected object's stack) always does, so the
Color, Stroke and Transparency panels edit every object's matching
item, and the item and effect commands do with `aligned: true`
(appearance.setItem / removeItem / duplicateItem, effect.remove /
duplicate / setParams / lightingPreset, transparency.set), each as
one undo step. moveItem and effect.move refuse aligned rows of
differing stacks. Without `aligned`, explicit indices keep their
meaning for single objects and scripts.
- "Mixed" is added to every catalog.
Tests: the shared stack of the issue's case, of a stroke at another
index, of nothing alike and of equal stacks, and that matches never
cross; editing the shared stroke edits both strokes (at index 1 and 2)
in one undo step that restores both, also through stroke.set on the
active row; removing a shared effect and adding a fill reach both, and
reordering differing stacks is refused; the panel lists kind, note,
Stroke, Feather and Opacity, not Fill, the full stack for alike objects
and "Mixed Objects" for a path and type. The existing object-row test
now expects the kind for two rectangles and the shared rows when their
opacity differs.
proxy_png decoded with image's default limits and no header check, while
every other decode path in the project uses MAX_RASTER_ALLOC counted as
RGBA. The header now decides first, and the decoder gets the same budget,
so an image too large to draw gets no preview by header alone, and a
large-but-in-cap image the app itself can open now gets its preview.
Co-authored-by: rahul05ranjan <11793455+rahul05ranjan@users.noreply.github.com>
dotCooCoo