PathData::nearest, which the Direct Selection cursor runs to find the
segment under the pointer, projected the pointer onto every segment of
the path.
It now projects the pointer first onto a segment at the anchor nearest
to it, then skips each segment whose control-point box lies farther
away than that distance or the nearest point found so far. Each box is
measured once and the search allocates no memory. The result is the
same, including which of equal distances wins. A segment is never
skipped when the pointer or one of its control points is past 1e100 or
not finite.
Over a segment of an open path of 200,000 anchors, the cursor takes
51 ms instead of 59 ms in a release build (anchor_bench). The progress
line in docs/target-app-parity.md gives the timing.
Tests compare nearest with every segment projected over 400 random
paths, each probed at 30 points on a coarse grid (ties included), and
on open, closed, compound, one-anchor, empty, huge and non-finite
paths.
A projection now counts as invertible when its determinant, or its
adjugate's determinant, is more than 1e-15 times the largest of the six
products that determinant sums. The adjugate of the adjugate is the map
times a number, so the inverse of a projection that inverts also inverts,
is stored and reads back. The adjugate's cofactors are formed with exact
products, so the inverse of that inverse gives back the original map up
to scale, within 1e-6 of each row's largest entry, and a map whose third
row is the sum of its first two is still refused.
Homography::conjugated and Node::transform_projection judged the affine
map by an absolute determinant of 1e-15, so a scale by 1e-8 left type and
symbol instances in perspective with their old projection. Both now use
Homography::affine_is_invertible, which compares the determinant with
the two products it sums at any scale of the map. An affine map whose
determinant or inverse is not finite, such as a scale by 1e200 or
1e-200, is inverted through Homography::inverse. When the direct product
underflows to a map that is not invertible, as for a projection stored
at a subnormal scale, conjugated forms it again with its powers of two
kept apart.
- crates/geom/src/projective.rs: the_inverse_of_an_invertible_map_is_invertible,
conjugating_does_not_depend_on_the_scale_of_the_affine,
conjugating_a_map_at_a_tiny_scale_gives_an_invertible_map
- crates/doc/src/node/tests.rs: a_projection_follows_the_transform_or_is_removed
covers an invertible scale by 1e-8 and a singular map at that size
The power of two that brings a small map's entries up before a point is mapped is now reduced by the power of two of the point's larger coordinate, so no product of an entry and a coordinate reaches 1. A map that mapped a point before the scaling step does so again, and affine_at uses the same entries.
apply and affine_at multiply every entry by the power of two that brings the
largest to at least 1/2 when all entries are below 1/2, so products that would
be subnormal keep their precision. The identity times 5e-324 mapped (0.5, 0.5)
to (0, 0) before. Maps with an entry of 1/2 or more are unchanged.
adjugate_inverse no longer applies the relative determinant test to its result. That test is not symmetric between a map and its inverse: a map whose determinant is 1e-8 of its largest product can have a finite, accurate inverse whose own ratio is near 1e-16, so inverse() was None while is_invertible() was true. The result is now kept when every entry is finite and its determinant, summed from the six products in split form, is not 0.
The inverse holds every cofactor and the determinant as a mantissa and a
power of two, and forms each entry from their quotient. A map such as
diag(1e-300, 1e300, 1e-300), which counts as invertible, now gets its
inverse diag(1e300, 1e-300, 1e300) instead of none.
When a cofactor of a map with a normal but very small determinant underflows to 0, the
inverse formed from the map itself can have a w row of 0, so every point mapped through it
is refused. The inverse is now kept only when it is itself invertible, and is otherwise
formed from the map divided by a power of two that brings its determinant's products near 1.
The scale-free determinant test no longer builds a rescaled copy of the map. It
forms each of the six determinant products as a mantissa product and an
exponent sum and compares their sum with the largest of them, so a map such as
diag(1e307, 1e-157, 1e-157), whose rescaled copy overflows, is still invertible
and loads from a document. The rescaled copy is built only to invert a map whose
determinant is not finite or is below the normal range.
A homography means the same map at any nonzero uniform scale of its
entries, but `apply`, `is_invertible` and `inverse` compared w and the
determinant with absolute thresholds. A map stored with small entries
refused every point and could not be inverted, and a map with large
entries drew points past its horizon.
- `apply` refuses a point when w is at most 1e-9 times the largest
magnitude in the w row.
- A map counts as invertible when its determinant is more than 1e-15
times the largest of the six products the determinant sums, taken on
a copy rescaled by a power of two. An inverse whose determinant
overflows or underflows is taken on that copy.
The Spiral tool places its anchors on a logarithmic spiral, with the
radius shrinking by the decay each quarter turn, but gave every handle
the direction of a circle's tangent and a length from an averaged
radius. The curve bulged off the spiral between anchors and turned at
each anchor, so the spiral looked squared: up to 2.8% off its radius at
the default 80% decay and 23% at 30%.
The handles now follow the spiral's own tangent, with lengths that
differ at each segment's outer and inner end. The curve stays within
0.03% of the spiral at 80% decay and 0.8% at 20%. Anchor positions and
counts are unchanged.
A stored projection whose determinant is not finite is read back as the
same map divided by the largest magnitude in its w row, the scale the
overflow branch of conjugated already uses, so points that overflowed at
the stored scale draw: the identity with every entry 1e308 maps (2, 2)
to (2, 2). The overflow branch of inverse scales the map so that the
largest of the products its determinant sums is near 1, so a map with
large entries and a small w row, such as a scale by 1e305 with a move
by -3.9e307, is still inverted and read back.
The two overflow tests in persp_text.rs build their stored projections
with a w row of 1, so they still overflow after the read-back.
Homography::from_array accepted nine finite numbers only when det().abs()
was above 1e-15. With large entries the determinant can be inf - inf, so a
valid stored projection was rejected on reading and its type drew flat, or
it can overflow to inf, so a projection that cannot be inverted was
accepted.
When the determinant is not finite, inverse() now inverts a copy divided
by a power of two (largest entry between 0.5 and 1), and treats that copy
as not invertible when its determinant is at most 1e-15 times the largest
of the six products it sums. from_array and the overflow path of
conjugated accept such a map only when that inverse exists. Maps with a
finite determinant are judged as before.
Follows #4912.
Homography::then_after returned products whose entries could overflow to
infinity or become NaN. A move, transform or grid reproject of type or a
symbol instance whose stored projection has large finite entries stored such
a product, and the type was then drawn flat, since a projection that is not
finite reads as none.
- Homography::checked_then_after returns None when an entry of the product
is not finite.
- project_node composes each type and symbol instance projection with it and
returns false, leaving the object as it is, when one would not be finite;
moves and transforms in perspective and Define Perspective Grid with
reproject then return the horizon error and leave the document unchanged.
- transform_map, attach_homography and size_homography return None for such
a product.
- Edit Text returns the horizon error, and does not enter isolation mode,
when the type cannot be shown flat where it is drawn.
When the conjugate a ∘ self ∘ a⁻¹ overflows, conjugated now forms every
entry of the product as a sum of terms whose factors are split into a
mantissa and a power of two, and scales the whole product by one power of
two from its w row at the end. Scaling each partial product to a largest
entry of 1 could round an entry to 0 that the next factor would restore,
so a projection with entries far apart in size, such as a scale by
(1e160, 1e-160) seen through the same scale, was removed and the art drew
flat.
When the conjugate of a projection overflows, it is now formed one factor
at a time, and each partial product is divided by its largest entry
magnitude. Scaling the affine factors up front let the w row underflow to
0, so a scale by 2 seen through a move by 1e308 returned None and the
projection was removed, though its conjugate is finite.
When conjugating a projection overflows, the scaled-down result is now divided by the largest entry in its w row. A valid conjugate is no longer dropped by the determinant cutoff at the arbitrary scale, and every point of a kept projection passes the horizon test in apply instead of none.
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.
Homography::apply checked that w is in front of the horizon but not the quotients, so a point with w just above 1e-9 and a large x or y, or a NaN w, mapped to an infinite or NaN point. apply now returns None unless the mapped point is finite. inverse, conjugated and affine_at return None when their result has an entry that is not finite.
A closed arc from Arc Segment Tool Options, an Arc tool drag or
shape.arc with lengthX and lengthY closed through the box's bottom-left
corner with Base Along X, which makes a convex arc a quarter-ellipse
sector. With Base Along Y it closed through the corner a convex arc
bulges toward, so it drew the thin sliver between the arc and that
corner instead.
Such a closed arc now closes through the corner across from the one a
convex arc bulges toward with either base, so the Y-axis arc, open or
closed, convex or concave, is the X-axis arc mirrored top to bottom.
Open arcs and arcs with Base Along X are unchanged. The two-point
shape.arc form, whose params list no baseAlong, now closes through
(x1, y2) whatever baseAlong says.
Docs: docs/target-app-parity.md.
CodeRabbit findings #4, #34, #46, #51, #77, #101, #170, #171, #188:
- `vectorcraft_geom::flatten`: kurbo's flatten with each curve's tolerance raised so it makes
at most ~4096 segments (non-finite curves become a line); the canvas outline and
variable-width strokes use it.
- Plug-ins: a heavy filter's budget is the host's `Limits::heavy_fuel`/`heavy_wall_time_ms`,
never more; a start function runs on one fuel slice; the effect cache holds at most ~1M
anchors of inputs and results, and doesn't keep a result past a quarter of that.
- Loaded textures are cached by their PNG (base64 spacing doesn't matter), and live textures
hold at most 512 MB together.
- Revolve draws the source path when the faces' paint copies would pass 256 MB.
- Color books: names with prefix and suffix take at most 64 MB.
- ICC: localized-name (`mluc`) records over 1 MB together are refused before moxcms copies them.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
CodeRabbit findings #58, #75, #82, #113, #159, #175, #187:
- Live evaluation (expansion, envelope mapping and everything inside) makes at most 4M
objects (`MAX_GENERATED`); deep expansion counts what each live object makes, and envelope
image pieces count before they are cut.
- Clipping runs copy at most 8M clip path elements together; past it runs are plain groups.
- Exports read at most 256 placed documents; the rest show their previews.
- Graph design and marker budgets count opacity-mask art (`Node::count_with_masks`).
- Native files read at most 32M anchors (4M on wasm) of paths; the web build allows 1 GiB of
entries after the JSON.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Replace length/angle with Length X/Y, reference point, Type, Base Along, Slope and Fill Arc; extend geom and shape.arc params.
Co-authored-by: nrq6993-a11y <nrq6993-a11y@users.noreply.github.com>
anchor_arc_distance measured every segment again for each anchor, so a reshape cost the square of the anchor count; path.reshape now measures the positions once, and the two helpers only it used are gone. The default radius stays half the subpath's bounding-box diagonal, now measured along the path: half its length reached the far corners of a square, moving art the grab is nowhere near (reshape_moves_the_grabbed_point_and_its_neighbourhood).
PathData::bounds measured each segment by its start, end and extrema,
as kurbo's bounding_box does. For a curve with a handle at about 1e308
in size, or an infinite one, kurbo's extrema computation overflows and
drops the extreme. The box then held only the anchors. The canvas check
after an edit reads that box for the selected objects, so path.create
kept a curve with handles at 1e308.
The new segment_bounds measures a segment by its control points when
one of them has an x or y that is infinite or larger in size than
f64::MAX / 16, and by its extrema otherwise, bit for bit as before.
curve_bounds joins those boxes as kurbo does, and PathData::bounds uses
it, so such a curve's box holds the whole curve and the canvas check
refuses an edit that leaves one selected.
Docs: docs/mcp.md, docs/target-app-parity.md.
Shift pressed while dragging a segment had no effect on the curve:
Direct Selection's curved-segment and spine-segment drags, the Anchor
Point tool and the Pen with Alt sent the pointer offset alone, and the
segment bent freely. The reference app keeps the handles in the
perpendicular direction and of equal length while Shift is held.
path.reshapeSegment and object.blend.spine.reshapeSegment take
perpendicular (default false, which keeps today's behavior). With it,
both of the segment's handles stand at right angles to the line between
its anchors, on the same side and at the same length, and the point at
t lies as far from that line as the moved point does. A smooth anchor
at either end stays smooth, as it does in a drag without perpendicular,
and a segment whose anchors coincide bends as it does without
perpendicular. With Direct Selection, Shift pressed once a curved
segment's drag has started sends perpendicular; Shift held at the press
still toggles the path's selection, as before. Direct Selection's
spine-segment drag and the Anchor Point tool's segment drag send Shift
as perpendicular too, so the Pen with Alt does as well. A straight
segment dragged with Direct Selection already moves at 45° steps with
Shift.
Docs: docs/mcp.md, docs/target-app-parity.md.
A corner's radius was capped at half its shorter side whatever its neighbours did, so one corner of a 100 pt square rounded alone stopped at 50 pt. Each corner now rounds until its cut reaches the cuts its neighbours ask for: the whole side when the neighbour is square, halfway when both ask for more. The corner widget drag, the red limit outline and CSS border-radius export follow the same limit.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Both readers skipped a clip of at most six straight segments whose
bounding box contained the page. A diamond touching the page edges has
the page as its bounding box, so its clip was dropped and everything it
cuts away showed. geom::hit::polygon_covers_rect answers it properly
(one convex polygon containing the page's corners); any other path is
kept as a clip.
Fixes#1542
Live Corners cut a corner between a curve and another side with the handles of
the arc between straight sides. The curve has turned by the time the cut meets
it, so those handles reached past where the tangents at the cut's ends cross,
and from a middling radius on the cut bulged out of the shape or looped. Each
handle is now the share an arc takes of its tangent up to that crossing.
Corners between straight sides are cut as before.
The Anchor Point tool drags any segment of a selected blend's spine into
a curve, straight ones too, and Direct Selection reshapes a curved spine
segment as it does a path's (a straight one it leaves, as on paths). Both
run the new `object.blend.spine.reshapeSegment`, which bends the segment
with the same math as `path.reshapeSegment`, now shared as
`SubPath::bend_segment`.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A marquee drawn wholly inside a filled shape counted as touching it, so
a marquee over a background (the usual way to pick small objects off it
in Outline view) selected the background too. As in Illustrator, the
marquee now needs the path itself: an anchor inside it or a segment
crossing its edges.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A corner is now any anchor that isn't smooth where two sides meet at an
angle, straight or curved, as in Illustrator. The cut meets each side at
its setback measured along the curve, leaves both sides along their
tangents there, and each curved side is trimmed to what the cuts at its
ends leave of it. The largest radius is still the one taking the cut
halfway along the shorter side, measured along the curve. Corners
between straight sides draw exactly as before.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
fill_contains promised to close open subpaths (as their fill does) but
counted kurbo's winding, which leaves the closing edge out: a point beside
it read as outside. An open path (an anchor deleted from a closed one) lost
its fill clicks on that side, and a reflection moved the edge across, so it
looked like Reflect broke selection. Open subpaths are now closed by a
straight line back to their start before the winding is counted; closed
paths are borrowed as they are.
Tests: the open path from the report, on both sides and even-odd; reflected
shapes, imported SVG hearts, a selected shape dragged by its fill.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
shapes::polar_grid drew at least one radial divider (`radial.max(1)`), so radial 0, which
the command and the Polar Grid tool (Left arrow while dragging) accept, still drew a spoke.
Zero now draws none; the loop never divides by it.
The test checks 0, 1 and 2 radial dividers with no concentric ones give 1, 2 and 3 paths.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
# Conflicts:
# ROADMAP.md
An ellipse kept pie angles in its model, but nothing drew or set them. Now
the Transform and Properties panels show Ellipse Properties with Pie Start
Angle, Pie End Angle and Invert Pie, as in Illustrator, and
object.setLiveShape takes {pieStart, pieEnd, invertPie}. The pie runs
counterclockwise on screen from the start angle (0 at 3 o'clock) to the
end angle and closes through the centre; 0 to 360 is the whole ellipse,
an end of 0 means 360, and Invert Pie swaps the angles to show the other
part. One undo step each; a bad angle is an error. The labels are
translated in every complete language. The on-canvas pie widget is still
to come.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Shaper fits a hexagon's rotation by scoring 0, 90, 180 and 270
degrees, but a hexagon at 180 is the one at 0 (and 270 the one at 90):
those pairs score the same up to rounding, so the platform's atan2 picked
between them. FreeBSD's libm gave 270 where Linux gives 90, failing
tilted_rectangles_ellipses_and_hexagons_snap_to_their_angle_steps on the
FreeBSD build+test job. Candidates a whole number of periods from an
earlier one are now dropped. A new test reproduces the flip on Linux
with the old code (180 instead of 0 at tilt -9).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GQPh8UDgMyARR5Hat9UchC
dotCooCoo