Keep all solid geometry GPU-resident during navigation. Use vertex-rate instancing, compact untextured vertices, vertex-cache ordering, and cached silhouette sources without culling geometry.
The shaded surface of a curved solid was visibly coarser than its own
wireframe: the wire ran along the true rim while the facets cut inside it.
Two things were wrong, and only one was the tolerance.
`facet_res` is a resolution multiplier — the same one `scale_lod` divides the
fallback sampler's chord fraction by — and the kernel path was using it as a
sag in world units. At the default that asks for a whole unit of departure, so
nothing smaller than a metre subdivided at all and a face was only ever as
round as its own boundary. It is a fraction of each surface's own radius now,
which carries no assumption about the drawing's units.
And it is the *edges'* fraction, shared from `EDGE_CHORD_FRAC` rather than
copied. Those edges are drawn over these faces; sampling the two differently
is what put the wire across a facet instead of along its corners. `facet_res`
is deliberately not applied either, since the edges it must line up with are
built once at highest detail and never scaled.
A torus is measured by its tube rather than its ring — the tube is the tighter
bend, and sampling to the ring leaves the section a hexagon.
The primitive-volume test claimed a mesh can never read over its true size.
That holds for a convex solid, where every chord lies inside; a torus is not
one, and across the inside of its tube the chords fall outside the material.
It passed before only because the sampling was too coarse to show it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Faces were being dropped for sitting a nanometre off their own plane.
An edge is shared by two faces and in a real file cannot lie exactly on both:
the two surfaces were fitted separately and written to finite precision. Asked
for exactness, the kernel decides the edge is not on its plane, declines to
project it, and the face goes undrawn — silently, since a face that cannot be
expressed is reported as absent rather than as an error.
A micrometre in a drawing measured in metres recovers twenty-six walls of one
building. Loosening further buys almost nothing — a hundredth of this finds
one more face in sixty thousand — and past that it would start accepting
geometry that really is wrong, so it stops here.
With the tube fix that came with the kernel bump, VILLA_NEU_IFC3.dwg now
meshes 59015 of its 59033 faces, from 58922.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Two faults kept faces out of a lifted solid, and both looked like a document
problem rather than a reader one — the lift reported no loss at all.
An edge's stored parameter pair can span the right length from the wrong
place, putting a straight curve beside the face it bounds instead of along
it. And a loop's coedges are not always listed in the order the ring runs,
which the old chaining could not fix: it reversed the next piece in the list,
where what was needed was a different piece.
Either one leaves a ring that crosses itself, which ear clipping declines, so
the face is dropped without a word.
Handle 176 of Annotation-2D-Mesh-Solid-BIM.dwg now meshes all fourteen of its
faces, where it managed ten. Across that file: 105 of 107 solids complete, 865
of 867 faces. Across VILLA_NEU_IFC3.dwg: 672 of 763 solids, 58922 of 59033
faces.
The fallback sampler stays until that reaches all of them — a complete mesh is
what stops it running, and 111 faces still need it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Faces with holes were drawing the hole filled and the material around it
missing. Handle 176 in Annotation-2D-Mesh-Solid-BIM.dwg is one.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Two things the truck removal broke without saying so, both found by auditing
what is left in the manifest rather than by anything failing.
The cfg gate on `acis_export` ended up on `acad_to_render` instead, when that
module was put back after being deleted by mistake. `acad_to_render` is what
every entity converts through, so gating it left the wasm build — which is
what `--no-default-features` stands in for — unable to resolve it from thirty
places. It built at the commit before the migration and builds again now.
And deleting TRUCK_CHORD_FRAC took the constant but left its doc comment and
its `#[cfg]` behind, so the attribute landed on the next constant down and
gated a boundary-sampling tolerance that has nothing to do with solids.
lzma-sys was the last dependency in the manifest with nothing using it: no
code references it and nothing else in the tree pulls it in. The `solid3d`
feature it hung off still gates four places, so the feature stays.
Everything else declared is used. `getrandom` and `iced_widget` look unused
by grep and are not — the first is declared to turn on the feature the wasm
target needs from a crate half the tree pulls in, and the second is a
`[patch]` entry rather than a dependency at all.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The last four truck crates are gone from the manifest. What they were doing
falls into three parts, and each now has one answer instead of two.
NURBS evaluation. truck was carrying B-spline curves and surfaces for the
SPLINE entity, hatch boundaries, BLEND's endpoint frames, the spline preview
and ACIS spline-surface faces. The kernel now has both, over one de Boor
written across however many coordinates a control point holds — so a plane
curve and a space curve cannot drift apart, and a surface is the same
algorithm applied twice. The rational and the polynomial cases stop being
separate types: weights absent means polynomial.
The entity conversion. Every LINE, ARC, ELLIPSE, SPLINE and polyline was
built into truck topology purely so it could be sampled back into points.
They are sampled through `entities::curve` now, which is where each one's
geometry is already defined once and what EXTRUDE and REVOLVE read — so a
circle drawn on screen and a circle handed to the Model tab come from the
same definition. That retires four of TruckObject's variants and the module
is renamed for what it does.
SWEEP and LOFT. Both only ever produced a mesh, so both are built from point
lists: a band of quads per span, and a lid where a profile closes. Lofting
profiles of different densities resamples them by distance rather than by
index, so a circle lofted to a square no longer twists.
Two things worth noting for anyone reading the old comments. The tolerance
globals that lived in the tessellation module were never about truck at all —
they are the per-frame chord height, and they move to `curve_tol`. And the
rule that a profile handed over as `Lines` silently broke EXTRUDE and REVOLVE
no longer holds: those read the curve directly and never look at this channel.
The two `automation` test failures are unchanged from before this and are not
caused by it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Every 3-D solid OCS makes or opens now goes through cadkernel, which holds
analytic surfaces — a cylinder is a cylinder record, not a spline that looks
like one. That is what lets a solid built here be written back out as exact
ACIS instead of a triangulation, and a facetted export is a one-way door: the
next application to open the file would have no way back to the circle.
The Model tab's seven primitives, the Design-group booleans, 3DROTATE,
3DMIRROR, 3DALIGN, SLICE, SECTION, POLYSOLID and PYRAMID all come from the
kernel now, and `scene.solid_models` holds a `Body`. A mirrored solid needed
its own care before; the kernel reverses the loops on the way through, so it
comes back the right way out rather than lighting black.
EXTRUDE and REVOLVE sweep the profile `entities::curve` already defines, so a
circle, an arc-bulged polyline and a closed spline arrive as the same thing.
A circle extrudes into four cylinder patches rather than a run of chords, and
a profile turned about an axis becomes a cone, a sphere or a torus. A spline
profile has no analytic side wall and is refused rather than approximated.
ACIS goes both ways through the codec's own bridge: `lift` to read and
`append` to write. That drops the 679-line truck conversion and the planar-
only exporter, and the tessellator keeps its bespoke sampler as a fallback for
whatever the kernel cannot yet express — reported rather than passed off as
whole, since a mesh missing a wall looks finished.
truck-shapeops is no longer used and is gone from the manifest. truck-modeling
stays for now: fourteen files use it as a 3-D NURBS evaluator, which the
kernel has no answer for yet, and SWEEP and LOFT have no kernel operation
behind them either.
The two `automation` test failures are unchanged from before this and are not
caused by it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The kernel gained a B-rep layer — topology, analytic surfaces, surface
intersection, face splitting, point classification, booleans — and
acadifc gained the bridge that lifts an ACIS document into it and lowers
it back, provenance-driven so an untouched body saves byte for byte.
The reachability test is a compile-time one on purpose. A dependency
chain that stops resolving is caught where it breaks rather than the
next time somebody reaches for the far end of it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The pattern's lines were clipped by sorting their crossings with the
boundary and taking every other span. That is right where a line cuts a
boundary cleanly and wrong wherever it grazes one: a line through a
corner crosses twice in effectively one place, and from there every span
is inverted — the fill lands in the holes and the solid comes out empty.
A 1e-5 dedup on the crossings hid some of it and could not fix the rest,
because which crossings to merge is not the question.
The kernel divides the line by its crossings and then asks each span
whether its middle is inside, which costs a containment test apiece and
cannot get out of step. It also accepts a curved boundary rather than
only a polygon, which this call site does not need yet and a section
fill will.
What stays is the PAT frame — the family angle, spacing, origin offset,
the k range and the dash walk — which is a format convention rather than
geometry.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
SPLINEDIT's REVERSE reversed the control points and then regenerated a
clamped *uniform* knot vector. That is only harmless for a spline whose
knots were evenly spaced to begin with; for any other — which is most of
them once a modeller has been near one — the curve came back a different
shape, bulging where it had been taut. It now mirrors the knots within
the domain, which is exact, and swaps the end tangents to match.
PEDIT's Spline option evaluated a uniform cubic B-spline from four
blending polynomials written out by hand, clamped the open case by
repeating end control points, and then pinned the two endpoints
afterwards to undo the drift that leaves. A clamped knot vector says all
of that directly, and the closed case gets a genuinely periodic curve
rather than one whose seam is a special case.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
BLEND summed Bernstein terms using a binomial-coefficient table with
entries for degrees three and five — the two it builds — and a fallback
of one for everything else. Correct today, and quietly wrong the moment
a third continuity mode is added. De Casteljau needs no coefficients at
all, works at any degree, and is steadier: the Bernstein terms are large
coefficients times small powers and they cancel, while repeated
interpolation only ever averages neighbours.
Curvature through three points moves with it, since both are the 3D
arithmetic a blend needs and neither has a planar counterpart to
delegate to.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A hatch's area was not the area of anything. Its boundary was flattened
into a ring by pushing an arc's *centre* into it, and a spline's control
points — neither of which is on the boundary — and the shoelace formula
was run over the result. Boundary edges now go through the kernel, which
measures each for what it encloses: an arc contributes its sector, a
spline is integrated.
AREA had three measurements between them. The bulged-polyline one
carried its own arc corrections; the planar entities it did not cover
fell through to mass properties. Both are now one call on the entity's
curve, which gets the bulges, the ellipse and the spline right in the
same place. What is left is the genuinely spatial — a 3D polyline, a
spline whose points wander off any plane — measured as rings of points
by Newell's method rather than by projecting them flat.
Regions were doing the same Newell sum by hand. They share it now.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
FILLET enumerated its own candidates twice, once for a line against an
arc and once for two arcs, each with its own quadratic, its own four
sign combinations and its own conventions about which root meant what.
Both are the same construction: a circle of the given radius tangent to
a curve has its centre on one of that curve's offsets, so the fillet
centre is where an offset of one crosses an offset of the other.
That now comes from the kernel, which does it for any pair of straight
and circular curves and agrees with the corner formula where both apply.
What stays here is the part that genuinely belongs to the command —
which candidate the two picks meant, whether the touch lands on the
drawn part of an arc rather than elsewhere on its circle, and which side
of each entity to keep.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
624 lines become 70. What went is the planar arrangement itself —
splitting segments at their crossings, welding coincident endpoints and
tracing the bounded faces of the resulting graph — which is not a
drawing-specific problem. It is the same one a B-rep boolean solves in a
face's parameter space once the intersection curves are projected into
it, so solving it in the kernel means the boolean does not arrive with a
second copy.
What stays is reading the wires, which is where the drawing's own
conventions are: NaNs separating the runs inside one WireModel, and the
double-single coordinate pair that has to be recombined before anything
is measured.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
DIVIDE and MEASURE did nothing at all on an ELLIPSE or a SPLINE. Their
length function answered `0.0` for both — the `_ => 0.0` arm — and the
command reported the entity type as unsupported. On a polyline they
measured the chords and ignored the bulges, so points landed short of
where they belonged; on an arc or a circle they read the centre as world
coordinates when the entity stores it in its OCS.
All three now walk the entity's own curve, which the kernel measures
properly: closed forms where they exist, and the integral of the speed
where they do not.
A path ARRAY had the same confusion in a different place. It sampled the
path at even *parameters*, which is the same as even distances on a line
and a circle and nothing like it on an ellipse or a spline — copies
bunched up wherever the parameter ran slow. It also now closes properly:
on a closed path the last copy no longer lands on the first.
LENGTHEN measured an ellipse with a hundred and twenty-eight chords and
then bisected on that, and a spline with forty rounds of thirty-two
chords apiece. A chord sum reads short, so "current length" was already
below the true one before any delta was applied. Both now ask the kernel,
which also makes the spline case one evaluation instead of about
thirteen hundred.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The point lists were built three different ways and none of them scaled
with the geometry. A full ellipse was cut into sixteen segments per half
and an elliptical arc into thirty-two, whatever its size, so anything
large read as a polygon; a circle worked out its own segment count from
the chord tolerance; an arc was handed to truck's arc-through-three-
points builder, which cancels at survey coordinates.
All three now sample the entity's curve at the render pass's chord
tolerance, so how round something looks follows how big it is and how
far in the view is zoomed. The ellipse also stops casting its local
coordinates to f32 on the way in — the vertex path splits f64 into a
double-single pair precisely so that precision survives to the shader,
and throwing it away first defeated that.
What does not change is the kind of object each entity hands over.
EXTRUDE, REVOLVE and SWEEP read their profile and path out of Contour or
Curve and have no arm for a bare point list, so an arc or an ellipse
given as Lines would quietly stop being usable as either. The arc keeps
its circle_arc edge for that reason; only the ellipse's control points
change, and they keep the same two-half-edge wire.
The ellipse now decides it is closed from its own endpoints rather than
from its stored parameters, so a file that writes an end a hair short of
a full turn still produces a wire a plane can attach to.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The same shape was being read out of an entity's fields in three places
and the three had drifted. An arc's tessellation cut it at 20 segments
per radian, TRIM's sampling at 16, and its snap points came from a
fourth expression again; a spline was sampled at a fixed 96 points
however long it was; TRIM read an arc's centre as world coordinates
when the entity stores it in its OCS, so an arc carrying the (0, 0, −1)
normal a MIRROR leaves behind acted as a boundary on the wrong side of
the drawing.
entity_curve returns a PlanarCurve: the shape, plus the plane its
coordinates are read in. The plane comes from the entity's OCS — its
extrusion normal through the DXF arbitrary-axis algorithm — not from the
UCS. A UCS is the working plane a user drew in and an entity does not
remember it: draw a circle in a UCS turned 30° about Z and the stored
normal is still +Z, which is why the header keeps the UCS separately.
The algorithm itself stays in transform.rs, where the format is
understood and where #142's degenerate case was fixed once.
Entities that store world coordinates and may not be planar at all —
LINE, SPLINE — are checked rather than assumed. A spline whose points
wander in space returns None instead of being flattened onto XY, which
would move it.
entity_curve_xy folds the plane away for the editing commands, which
work in plan view and need every boundary in one frame before two of
them can be intersected.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The codec declared nalgebra and never used it — its Vector2 and Vector3
are plain structs of its own. It was pulling eleven crates into every
build here for nothing.
Also picks up cadkernel's Vec2, which replaces the vector arithmetic the
kernel had been writing out by hand.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
`line_seg_ts`, `arc_seg_ts`, `ellipse_seg_ts` and `spline_seg_ts` each
walked every boundary kind and solved that pair by hand — four nested
matches over the same seven cases. They are now four descriptions of the
target as a `Curve` handed to one `cut_params`, which asks the kernel.
Their signatures are unchanged, so the twenty call sites are untouched.
Three things follow from it.
Boundary splines and target splines stop being sampled. A cut on a spline
now lands where the boundary actually crosses rather than where a fixed
sampling said it did, because the kernel subdivides and converges to a
tolerance instead.
Rays and infinite lines as boundaries are handled by the kernel's extent
rather than by a case here — a ray rejecting what lies behind its origin
is the kernel's job now.
`extend_spline` shoots a real ray. It used to build a line a million units
long and cut against that, which both invents a limit and, at survey
coordinates, is comparable to the coordinates themselves. The ray's
direction is a unit vector, so what comes back is the distance to the
boundary. That also fixes the hit position, which multiplied by the
million-unit extent a second time and landed the extended end far past
where it was supposed to.
`cut_params` clamps to 0..=1 only for a bounded target; a ray has to keep
the parameter it comes back with, since the distance is the answer.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Two faults, reported together.
Fit-point splines could not be edited at all. A SPLINE is stored either as
a control polygon the curve approaches or as fit points it must pass
through with optional end tangents, and `spline_to_nurbs` only read the
first. A fit-point spline has no control points, so every command that
touched one — TRIM, BREAK, LENGTHEN, OFFSET — got nothing back and did
nothing. Two fit points with start and end tangents, which is what a
drawing stores for the simplest curve, was the clearest case. Fit points
now interpolate into an exact curve through cadkernel, so they are as
editable as any other spline.
Trimming deleted the whole spline when the surviving piece reached the
curve's own start. `trim_spline` cut away everything before the interval
and everything after it, and a cut at either end is a no-op that reports
nothing to do — which the end case handled and the start case propagated
as failure, dropping the piece. That is why one side of the cutting line
worked and the other emptied the drawing.
Elevation is read from whichever point list the spline actually has. A
fit-point spline has no control points, so it was falling to z = 0 on the
first edit.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
