* feat(geom): surfaces through curve networks, and spline spaces to build them
Adds spline_space (clamped B-spline spaces, Hermite interpolation, exact
embedding by Bézier degree elevation and knot insertion, a monotone cubic
for reparameterising) and network (loft, Coons and Gordon surfaces through
curves in one or two directions, with G1/G2 cross-boundary conditions
against a neighbouring face). Part of #763.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* feat(kernel): boundary surfaces from sketch curves, 3D curves and edges
kernel::boundary_network turns sketch chains, 3D sketch polylines and edge chains
into network curves, gives G1 frames of the faces along open edges
(EdgeFaces), builds the B-spline face (sides split at chain vertices so a
join shares the body's edges) and joins it into a surface body by stitching.
G2 is refused for now (not supported yet). Part of #763.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(kernel): second derivatives of B-spline surfaces in the vendored truck-geometry
BSplineSurface::uuder and vvder multiplied the second differences by p
instead of p·(p − 1), as BSplineCurve::der2 does, so every surface of
degree p ≥ 3 (NURBS too, through its homogeneous surface) reported
1/(p − 1) of its second derivative: a cubic surface's curvature read as
half. This is why a G2 boundary surface on a cylinder measured half the
curvature it carries (#763). New tests hold both to finite differences
for degrees 2 to 5.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LGeCa5jWWdroMWvXxN3LLM
* feat(kernel): G2 boundary surfaces against a body's face
The network always carried the face's curvature across a G2 edge; the
refusal came from measuring it with truck's halved surface second
derivatives (fixed in the previous commit). EdgeFaces now reads the
face's second fundamental form from the exact second derivatives
instead of differencing the normal. A new test holds the curvature
across a cylinder's edge to the cylinder's within 0.1%, also with a
sheared far curve; G3 is refused (a network holds up to G2).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LGeCa5jWWdroMWvXxN3LLM
* test(kernel): a boundary patch with straight corners measures its own area
A stadium's arcs meet its lines at 180°, so the Coons patch is singular
at its corners. From such a corner the reference mesher's boundary walk
ran onto the B-spline's extrapolation past its domain (a planar spline
passes through every point of its plane there), and the face measured
289 mm² for its 158.5 (#763). The vendored mesher's bounded projection
(nearest_projection, #2043) keeps the walk on the domain; this holds it
there, by the faces, the measuring mesh and the display mesh. The
display path (#1653) was never affected.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LGeCa5jWWdroMWvXxN3LLM
* feat(geom,kernel): three-sided boundary surfaces (one side collapsed to a point)
A network curve may be a PointCurve: the apex where the two sides beside
it meet. The curves through it cross it at the point itself, its row of
the net collapses there, and the kernel's face leaves that side out of
its loop (three edges, one vertex at the apex). One point per patch, as
a side; refused elsewhere, when it moves, or with a continuity
condition. Tests: a triangle and a quarter disc filled flat with their
analytic areas (40 mm², 25π), by geom and by the kernel's faces and
meshes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LGeCa5jWWdroMWvXxN3LLM
* feat(doc): boundary surfaces through curve networks and loops of three or four
FeatureKind::SurfaceNetwork saves a network as `u` and `v` curves, each
a sketch chain (planar or 3D), a chain of a body's edges with an
optional G1/G2 continuity, or a point (the apex of a three-sided
patch), and an `operation` that can join the surface into the body of
its edges. A SurfaceBoundary of three or four sketch curves, refused
until now, builds as the flat Coons patch of its loop (three collapse
the side opposite the first curve). The format change is one pending
file; feature-walk goldens record the new variant's sketches, bodies
and points.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LGeCa5jWWdroMWvXxN3LLM
* feat(engine): surface.boundary through loops of three or four and curve networks
The command keeps its loop form (two to four curves of one sketch; three
and four now build the flat patch of the loop, crown refused there) and
gains a network form: `u` and `v`, each curve a sketch curve id, a
{sketch?, curves} chain (planar or 3D), a {body, edges, continuity?}
edge chain (g1/g2 with the face beside it) or a {point} for a
three-sided patch, with `operation` new or join. Edits may move between
the two forms (timeline.redefine accepts the command now that it has a
dialog).
Tests: analytic areas of loop patches and of a hyperbolic paraboloid
bounded by four 3D lines (every mesh vertex on z = xy/10), a loft
across two sketches, G1 and G2 measured across a cylinder's edge with
inspect.continuity (angle < 0.1°; curvature jump 1e-16 with G2 against
0.027 with G1), 26 hostile networks refused without panicking, and u/v
added to hostile_params_never_panic.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LGeCa5jWWdroMWvXxN3LLM
* feat(ui): Boundary Surface dialog
Curves and Cross curves take sketch curves and edges (picks that meet
end to end in one sketch or body are one chain); Continuity and
Operation (New Body or Join). Two to four sketch curves of one sketch
with nothing across are the loop patch; anything else is a network.
Edit Feature brings either back (per-side orders, crown and stations
the dialog does not show are kept). Strings in en, ja and zh.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LGeCa5jWWdroMWvXxN3LLM
* docs: ROADMAP M8 and the bridge patch's notes for boundary networks (#763)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LGeCa5jWWdroMWvXxN3LLM
* test(oracle): a boundary surface through a loop, thickened (#763)
Case 83: four sketch curves (a square's three sides and a half circle
for its top) make one boundary surface, thickened 2 mm; analytic values
from tools/oracle-analytic/cases_surfaces.py (V = 2A, area 2A + 2P,
centroid by moments). Recipes gain boundary_surface, thicken and remove.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LGeCa5jWWdroMWvXxN3LLM
* refactor: boundary surfaces, review cleanups
Project onto the face once per derivative evaluation (the frame is
passed to natural() and reused by scale()), keep each track's
finite-difference step instead of rescanning its table, export the
kernel's edge-pick and chain caps and the default station count for
the engine, doc and dialog to share, map TooLarge spline errors to
KernelError::TooLarge, reuse the engine's operation parser, and skip
the dialog's sketch lookup when the picks are no loop.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LGeCa5jWWdroMWvXxN3LLM
---------
Co-authored-by: jyviko <2501648+jyviko@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: bherbruck <brennen.herbruck@gmail.com>
A scale across a revolution's axis keeps the face a surface of revolution and is written as one; any other refusal from the writer is an export error, not a panic.
A point within the point tolerance of a boundary curve's end is that end in the boolean, whatever its parameter, so text cuts no longer leave unmatched sliver vertices (#391). Robustness goldens re-recorded for seeds 0, 34, 39, 41 and 57 with reasons.
The mesher samples edges through a fallible evaluation, so an intersection curve whose projection fails to converge no longer panics; a boolean result with an edge that does not evaluate is refused with an error. Fixes#1252.
A crash inside a graphics driver left only "an unclean exit between
commands": no backtrace, no module, no hint that it was at startup.
- The crash journal now keeps the startup step the app is in (window,
graphics instance, surface, adapter request, adapter chosen, device
creation, renderer and viewport shader compiles, surface configure,
first frame and its present, running) and the renderer as soon as the
adapter is picked. egui-wgpu tells the app about its steps through a
small hook. An unclean exit before the first frame names the step.
- Windows: an unhandled-exception filter (src/native_crash.rs, the one
Windows-only module allowed unsafe) writes the exception code, the
faulting module and offset, and a minidump next to the journal; the
next start puts them into the report.
- The Windows CI job runs a child that faults on purpose and checks the
report.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
When projecting a face's boundary loop onto its surface fails from an unhinted first point, the walk is retried once from the first point an unhinted search does find; a walk that succeeds is unchanged. A fillet strip on an extended surface now meshes, so the rim cut in the reported case builds in one boolean.
#970's lenient retry of a face's parameter boundary searches the whole surface as a last resort; on robustness seed 20 it found an edge's points across a cylinder's seam, half a turn apart, and the Cut failed. A lenient boundary whose points jump half the parameter range or more is refused, and the face is left undivided as before (#970's elbows step at most a fifth of the range). Seed 81's volume moved by 0.1 mm3 with #952's hole direction; recorded.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Cutting one lofted round tube from another failed in truck's divide_face ("parameter boundary") for some section counts: the intersection points lie on the loft's seam, the previous point's parameter can sit a hair outside the range, and the Newton search from there fails.
The boundary is built as before; only when that returns None is it built again leniently, starting each search from the hint clamped into the parameter range and, as a last resort, searching the whole surface. Nothing that succeeds today changes. The boolean-retry guards are untouched.
Fixes 5 of the 6 failing hollow-elbow cases in #968 (new test crates/engine/src/tests/loft_cut.rs, volumes within 0.3% of exact). 90 degrees / 14 sections still fails in a later step and stays open in #968.
#921's grid lines through the vertices that split a ruling edge also went into the meshes Boolean operations intersect, which changed seven robustness seeds (seed 2's sphere cut failed). Deflection::rulings now asks for them only in the kernel's own tessellation (display, export, measurement). The seeds #936 (rejected instead of unsupported), #946 (26 and 85 build, others fail differently) and #921's display side (about 1e-6) moved are recorded with their reasons.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Moving a window between monitors of different scale can briefly ask for a
surface larger than the device's largest texture (11512 x 6808 against 8192 in
the report), and Surface::configure failed validation, which wgpu turns into a
panic. The surface is now clamped to the device limit with its aspect ratio
kept, and egui draws the whole window at the matching smaller scale, so the
frame is only softer until the size settles. The device also asks for the
adapter's own texture limit (16384 on most desktop GPUs) instead of 8192, and
the resize path and an unsupported surface log instead of unwrapping.
Fixes#869
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
The exact surface search could return an arbitrary parameter at a collapsed NURBS boundary, which corrupted the trimmed triangulation domain and lost volume on fine meshes of a fillet tapering to zero. Search results in the boundary projection are now validated, preferring a bounded parameter at a pole and snapping to a finite endpoint only when its transverse tangent has collapsed and the residual improves. Adds four regressions for decreasing chord tolerances, pole parameters, unbounded planar support and provisional preview.
Fixes#796.
A boolean that cuts a pocket into the top of a cylinder (Shell does this) splits the
outer wall's straight seam edges at the pocket's floor. The wall's parameter grid has no
division along its rulings, so the triangulation fanned out from the two split vertices
to the far rims and its triangles cut through the part: the shelled Bolted Flange sample
drew and exported its rim cut off (mesh volume 29201 mm3 for a 40796 mm3 body).
Give the grid a line through each vertex that splits a boundary segment running along a
ruling, in a direction that needs no division of its own and within the grid budget.
Fixes#819.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
truck's face and edge IDs are their addresses, so walking maps keyed by them followed heap layout, and the ribbed lid's side-wall hole passed or failed by chance. The planar boolean's rebuild and glue now take truck's connected components in the shell's order, and the planar boolean's own maps are ordered. Healing's union-find keeps the lesser root and takes a run's ends in order of use. The face classification after an intersection orders each component's faces by the shell first. The lid now builds the same body on every run, with a test that builds it six times in one process. Its 0.5 mm side-wall holes, and one sensor enclosure hole of the same kind, now fail every time and are marked failing in the baseline.
Fixes#707.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Refs #303. file.save / file.save_as can store a warm native geometry cache (cache: true, off by default) keyed to the document hash, app version and adapter revision. Opening shows a bounded preview from it until the rebuilt session replaces it. Any cache failure counts as a miss.
Display meshes also bound the angle one facet turns, so bends and small holes on a large body are drawn round instead of as chamfers or squares. Bounded per face, and only the display mesh changes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Choosing 8x anti-aliasing in Preferences › Graphics on a device that
only allows 1x and 4x for the window's format made every later start
panic in wgpu's pipeline validation (`Sample count 8 is not supported
by format Bgra8Unorm on this device`), so the app could not be opened
again without editing graphics.json by hand.
- egui-wgpu (vendored): request TEXTURE_ADAPTER_SPECIFIC_FORMAT_FEATURES
when the adapter has it (WebGPU only guarantees 1x and 4x), and clamp
the sample count to the largest one the device supports for both the
colour target and the depth/stencil format. The supported counts are
exposed as `RenderState::msaa_supported`.
- At start, a saved count the device can't do goes back to the default
(4x, or the largest supported count below it), is written back, and a
"Settings were changed" notice says what changed.
- The Anti-aliasing list greys out unsupported counts, `ui.prefs`
rejects them, and an invalid saved `msaa` loads as the default.
- New messages translated in en, de, ja and ko.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
ops::boolean's result carries its history over [a, b] whichever path made
it, recorded by the code that builds the faces:
- truck path (incl. jitter retries): shapeops history mapped through the
order-preserving copies and the cut's turned tool, then through healing
- heal records which input faces each healed face merges and which input
edges each joined or exact edge replaces (HealMap); opt-in, so heal_keep
callers pay nothing
- trivial shortcuts, the void, glue, faces pushed apart (push_faces records
its own history; the slab cut back on a union maps to the operands'
faces in that plane), the nudged retry, by_lumps and join composition
- prism booleans: section segments carry tags (operand cap edge, the wall
beside it, its vertices) through the 2D boolean; walls, caps and edges of
the extruded or sewn result follow from them
- planar and faceted polygon booleans: polygons carry their operand face;
faces join the pieces across edges, edges come from the pieces either side
History gains renumber/over (and consuming forms) for composition;
record gains regions/traced/build helpers. truck-shapeops
*_with_history now return the solid with an Option history. Paths that
cannot report (bore narrowing via offset, inexpressible lineage) leave the
result without history. Boolean results, tolerances and fallback order are
unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01J53sR6gEaVt56soFDpTck8
and_with_history / or_with_history return, with the solid, the operand face
each result face is a region of (and whether it is that face untouched) and,
for each result edge, the operand edge it is (a piece of) or the pair of
operand faces whose intersection made it. Faces and edges are recorded where
they are created (face division, intersection edges, split edges, edges
rebuilt on a moved vertex), resolved to operand terms at creation, so the
passes folding in extra shells compose without bookkeeping. and / or wrap
the same pass and drop the history; a missing record never fails them.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01J53sR6gEaVt56soFDpTck8
* test: reproduce #261 (Rectangular pattern of a hole cut scales super-linearly)
* Patterns of many hole cuts scale near-linearly (#261)
Two quadratic costs:
- Meshing a face tested every point against every segment of every boundary
loop, so a plate face with hundreds of holes meshed in quadratic time. The
vendored truck-meshalgo inside test now skips loops whose parameter box is
clear of the point (they wind zero times around it).
- A pattern cut each copy in turn, rebuilding and re-checking an ever more
complex target per copy. Copies whose boxes are clear of each other are now
one tool of many pieces (kernel::union_apart, no intersection work), cut once.
512 holes now take ~7 s instead of not finishing. The test's volume tolerance is
widened from 1e-6 relative to 0.1 % of each hole: the measure meshes the bores
and reads each hole ~0.06 % small, on main too.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Hole-pattern test: a budget CI can meet; coarser apart-box meshes (#261)
CI runs every test binary at once and is about ten times slower than a
machine running this test alone (15 tests passed 60 s in the last run), so the
512-hole pattern (about 6 s alone) hit the 60 s timeout there. The budget now
only has to tell linear from the quadratic hang: 300 s timeout, 120 s budget.
union_apart meshes each copy coarsely (2 % of its size, box grown by twice that).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Hole-pattern test checks the scaling, not a wall-clock time (#261)
CI runs every test binary at once: the 512-hole pattern (about 6 s alone, 10 s
CPU single-threaded) took 177 s there, so no fixed time is reliable. The test
now patterns 128 and then 512 holes and requires the second to take under 8x
the first (linear is 4x, measured 3.5-4.2x here; the old quadratic cost was 16x
or more), each with the volume check and a 900 s hang guard.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: bherbruck <brennen.herbruck@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Moving a body transforms its intersection-curve edges, and truck-geometry 0.5.0 scales each curve's tolerance by the operator norm of the matrix. matext4cgmath 0.1.0 computes it from eigenvalues through a quartic whose Newton refinement loops while the residual is above a tolerance and never gives up. For a translation matrix all four eigenvalues are 1, and near that quadruple root the iteration can cycle forever (about one translation in eight of the boolean retries' shifts, by a scan of sizes 0.1-2000 mm). matext4cgmath is vendored with both Newton loops stopped after 100 steps (vendor/README.md, ATTRIBUTION.md).
truck master no longer computes the norm in transform_by, so the vendored copy can go once a truck release after 0.5.0 is adopted.
Tests: moving_a_body_with_an_intersection_curve_edge_ends (kernel) cuts a cylinder with a tilted box and moves the result by one of the shifts that hung (hangs on main). examples/regressions/oblique_cut_of_cylinder_hang.json is the variant of the oblique cut that hung the app; it hangs on main, and since the previous commit it passes without reaching the shifted retries, so it is replayed by the engine test as a scenario only.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Three step_in tests and two split tests failed in debug builds only:
truck's debug_new asserts simple loops and two distinct end vertices, but
faces whose loop runs a seam edge both ways and full-circle edges on one
vertex are valid. Meshed faces and edges are now built unchecked. ci
(release) was unaffected; the kernel tests pass in debug and release.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AS5Tn3xieewUSg3uxrMVdL
truck's connected components came out in hash order of face addresses, so a
boolean's faces, and everything built on them (fillet bands, STEP entity
order), changed from run to run. That made the oracle 35 export differ run to
run. Its old step_export flakes came from the orientation check's drifting
search, removed in 2db9e66. The result's shells and faces now keep the
order of the shell they come from (vendored truck-shapeops). Oracle results
are unchanged (73 PASS before and after). Test: the tee of oracle 35 with its
junction fillet exports byte-identical three times in one process (it failed
before).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AS5Tn3xieewUSg3uxrMVdL
- Surfaces of revolution follow ISO 10303-42 strictly in reader and writer: the
normal is d/d(angle) x d/d(profile), the opposite of truck's revolution, as
Fusion reads it (fixes Fusion inverting fillet sphere corners and the dimple).
step_orientation_errors flags faces whose loop runs clockwise round their
normal; engine, oracle and kernel round trips assert it.
- Revolved faces with analytic profiles are written as native PLANE,
CYLINDRICAL/CONICAL/SPHERICAL/TOROIDAL surfaces, and the pieces truck splits
a revolution into are merged back into one face per surface (with a seam
edge for full turns), so face counts match other systems.
- A horn torus (tight sweep: tube touching its axis) is written as one rational
B-spline face whose loop passes the pole twice, with truck's zero-length pole
edges collapsed, instead of a degenerate surface of revolution; the reader
starts such figure-eight loops off the pole (Fusion's own part 50 now imports
with the right volume).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AS5Tn3xieewUSg3uxrMVdL
Meshing (vendored truck-meshalgo) works within a budget so hostile files
cannot hang: bounded edge and surface divisions (point caps, depth and
round limits, NaN-safe), and a grid budget per face shared out per shell.
Surface grids refine only the direction that bends; rulings left undivided
get cells comparable to the other direction, so trimmed cylinders (holes,
tee joints) no longer lose volume to long triangles. The helical sweep
imports in 12 s instead of 7 minutes.
Small sphere patches get poles away from them even when the file runs a
seam through the patch (a pyramid's corner blend). Faces the mesher cannot
mesh are reported as import warnings. Test: wild B-spline and far-off
boundaries import in bounded time.
STEP corpus 65/66 (50-sweep-tight-bend: a horn-torus bend whose inner side
collapses to a point; Fusion's own area for it is also off).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AS5Tn3xieewUSg3uxrMVdL
Vendored truck-stepio (writer only): an intersection curve's second surface
was written at the first one's id (duplicate ids for every boolean edge),
and revolved surfaces lost their orientation (the modeling-surface sense was
always true; inverted or mirrored revolutions came out inside out). Export
now writes intersection edges as C1 cubic B-splines through exact points,
so files carry plain 3D curves and no redundant surfaces.
Vendored truck-meshalgo: hinted parameter searches that run off a B-spline's
domain are rejected or clamped (meshing their grid never finished).
STEP import: straight revolution profiles are bounded in the meridian plane
and stop at the axis; loops on revolutions start away from the axis; face
extents include edge end points. Edges between pieces of one split surface
draw as seams.
kernel::step_validate checks a STEP file (duplicate ids, dangling
references, dense numbering); the oracle replay now exports every rebuilt
part, validates it and re-imports it with no warnings and the same volume
(all 32 rebuildable cases pass).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AS5Tn3xieewUSg3uxrMVdL
