The portable record's 4,000,000-item budget refused a real result: the join of a level-4
Menger sponge's level-3 sponges ended with "SMLib export: the mapped B-rep does not validate:
portable B-rep: item budget". The budget guards records from outside the kernel, and a body
the kernel built in memory from inputs already bounded where they came in isn't hostile.
Brep::validate keeps the budget for records read from outside: native design caches
(decode_native_body, behind its own 256 MiB, 1M-per-array and 20M-value caps) and STEP import.
The new Brep::validate_built runs every other check, without the cap, on what the kernel
builds: SMLib's export of an operation's result, a body's record (Body::from_record, after
the decoder has checked a file's), a part of a validated record, the record handed back to
SMLib, and the validity check. Its checks are linear in a record already in memory.
SMLib's export bounded its counts by the same budget; a count is now bounded by SMLib's own
longest array (flat::MAX_LEN, checked on entry), which keeps every offset worked out from
the counts from overflowing, and the redundant topology-length cap goes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Seg2::closest_param starts Newton from the nearest of 65 samples. On a closed segment (a full
circle) the start and end samples are one point and the first wins the tie; for a point just
before the start (in the sweep's sense) Newton then wants a parameter below 0, the clamp holds
it at 0, and the answer is the start point, off by up to a sample spacing (0.15 mm on a 5 mm
circle). The record tests' face-to-segment matching read an SMLib cylinder's side as lying
over no segment because of it. A closed segment whose answer lands on an end now also tries
Newton from the other end and keeps the nearer; every other answer is bit for bit as before.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A round face's box (form::extent) unwrapped each periodic parameter from one boundary point to
the next. Through a pole the loop goes on from anywhere round the direction the pole collapses:
a hemisphere bounded by its rim and its seam put its second seam use on the first and its box a
quarter turn off the seam, while its pcurves ran from the seam. Past the first pole the loop is
now moved the whole periods that close it. Reference case 14's re-imported dimple had a
quarter of its area missing (9601 against 9706 mm2).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The normal equations of patches about a centre were solved dense, so at most
500 free control points: the torus window, 1433 once its patches ran in their
own spaces, was refused. Each patch's interior (coordinates of points only it
holds, coupled by no row to another patch's) is a banded block now; the blocks
are eliminated exactly onto the unknowns the patches share (the spokes, the
centre, and the points rows across a spoke couple), whose Schur complement is
solved dense, and the interiors follow by back substitution. A solve takes up
to 6000 free points, as one patch does, and refuses past 4000 shared unknowns
(a 128 MB dense block) with that reason. The torus window and the pentagon now
solve (they fold, open).
The fill tests' own-edge check compares an edge's NURBS form to rounding, as
fill.rs's does.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
About a centre, every patch ran in one space both ways that held the breaks of
every half edge and spoke, so each patch carried every edge's knots: a window
cut in a torus, its edges SMLib's traced intersections with many knots each,
needed 13537 free control points. Patch k now runs along u in the space of
spoke k − 1 and along v in that of spoke k; the space along a spoke holds the
spoke and the two edge halves it meets beside its two patches, so the spoke
is one row of both, as before. The torus window needs 1433 (still over the
dense solve's 500, open). Refinement about a centre adds the middles of one
uniform grid to every spoke's space; a patch's conditions read its own
spaces' spans, and the stations along spokes the finest of them.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The fold check refused a patch whose normal, anywhere on a grid, was a right
angle or more from the normal at its middle. A fill may turn its normal that
far without folding: a dome that meets a tube's wall tangentially is vertical
at its rim (90° from its top), a window round more than half a cylinder turns
past 120°. A fold is the surface turning back over itself, which reverses the
normal between neighbouring points of the grid (24 × 24, 3.75° steps on a
half cylinder); that is what is refused now, and a normal that vanishes
inside the patch. It may vanish at a corner where two sides run on straight
(a closed loop's pieces) and on a three-sided patch's pole.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A cup's rim of seven or eight ellipse arcs meeting at corners was too wavy for
the height-field fill: its finest grid, 32 spans across, stayed 6.4e-5 off the
loop where the tolerance is 4.4e-5 (a smooth bicubic follows the height's kink
at a corner only as closely as its cells are small), and the loop went to the
general fill, seven faces where one fits. The grids now go to 64 spans (the
fit's system is banded); the first within tolerance is still the fill.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
face_tolerance (and planarity_tolerance) was 10⁻⁶ of the body's bounding-box
diagonal, which changes as a body turns: the cup turned 0.7 rad about
(1, 2, 3) grew from 54.7 to 61.6, its fill sampled to another tolerance, a
different surface (area 974.543 against 974.527). The size is now the
diameter of the smallest ball holding the vertices (solidcraft_geom::Ball,
Welzl's algorithm, expected O(n)), the same wherever the body is and however
it is turned. Twice the farthest vertex's distance from the vertices'
centroid is that too, but up to twice the body's diameter where the
vertices crowd one end: a rounded box's set-back-13 G2 corner patch then
ran 2.0e-3 past the box (corner_g2
a_long_setback_keeps_the_g2_patch_inside_the_box, limit 1e-3).
The own-edge check of the fill tests compares an edge's NURBS form to
rounding (10⁻¹² of each coordinate): SMLib makes a conic's form again on a
round trip, the last bit of a coordinate apart.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The patch fill meets its neighbours to G4, so Construction::Patch now offers g0 to g4: the
surface.patch command takes g3 and g4 and the Patch dialog's continuity row offers
"Curvature change" and "Smooth curvature change" by itself, from the one table.
The three-round corner test patches G2, G3 and G4 and reads the seams with the Continuity
Analysis too: the middle of each seam (30-70 %) at the order asked; the whole seam G1, since at
the corners the rounds' curvatures differ and no patch can meet both.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DtL5Lzv2n8fyKqD61NCHSF
Two command scripts in examples/clean-up: a round along a plate's edge running into a boss
wall, and a corner where three rounds meet. Each takes the blend's messy face off (face.delete
heal false) and patches the hole with Curvature (g2) and, beside it, Curvature change (g3).
Gauss-Legendre constants at f64 precision.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A fill asked G^n is faired with the thin plate plus the energy of order n + 1 (each to unit
trace), so the change of curvature its conditions ask at the edges spreads over the fill
instead of being squeezed into the first span; Gram matrices on six-point Gauss-Legendre,
exact for quintics.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Rows across a patch's C0 knot lines (a side's break where it is only C0 in its parameter, such
as a circle's quarter joint in homogeneous form): the one-sided derivatives across in the
ratio of the crossing sides' speeds there, blended between them. Without them the fill rippled
along those lines (zebra stripes in steps) and SMLib's surface check refused it.
- Space::basis_left: the basis of the piece ending at a knot.
- Orders above a corner's agreement come back smoothly over the next share of the side
instead of switching on at once.
- Orders above G1 are best effort once refinement is spent: the fill is returned with its
jumps reported; the tangent plane, the points and an unfolded fill stay hard.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- One patch has a space along u and one along v, each holding its two sides that run that way;
a side's knot keeps its continuity when raised (multiplicity m + q - p), and each side is
carried in exactly by interpolation at the space's Greville abscissae.
- Refinement halves the spans only in the direction whose sides miss, and one patch's normal
equations are banded (the shorter direction numbered fastest), so a loop with a long
intersection curve refines where it needs to.
- At a corner whose neighbours agree to a lower order than asked (their tangent planes, their
twist between the two edges, or their forms), the stations near it ask no more than that.
- A station is held to the angle tolerance beyond the edge's own lean out of its neighbour's
tangent plane: a fill that holds the edge cannot come nearer.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
One order-n implementation in the linkage-curve formulation (Pegna & Wolter 1992;
Hermann, Lukacs & Wolter 1999): geom::linkage reads the neighbour as a height over its
tangent plane to order n (its jet, from the surface's exact partials of any order) and
gives each order's condition on the linkage curve, with exact gradients (dual numbers),
so fill_iso solves orders 2..n by Newton's method. Each order's row is weighed by h^(k-1)
and its jump is held, carried into the fill by Taylor's theorem, within the angle tolerance.
The iso fill takes rational edges (a circle's arc on its NURBS form): homogeneous rows,
each edge scaled so the corner weights agree, the inside weights the Coons blend and
fixed, so the fill stays a linear least-squares problem. Its degree is at least n + 1,
refinement adds simple knots, breaks of different sides within rounding are one, and the
weak Hermite rows act in the tangent plane at the Coons start's own speed, so they no
longer pull against the G1 rows.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LUXeKMoxpM7Vuv8Rb6E3K2
A new undoable feature (`repair`, one pending format file) and command with input and
output schemas. In order:
- A cylinder, cone, sphere or torus face SMLib refuses on its own is rebuilt on a NURBS
form of its surface laid on arcs from where the face starts
(`form::surface_on_own_arcs`), with SMLib's own trims: the import's quarter-grid form of
a face round most of a turn closes on itself, which SmTorus::IsNurbSurfaceTorus does not
recognise. Geometry unchanged.
- Free edges that pair up within `gap` are sewn by SMLib's stitch; GlueEdgesGeneral gives
each glued edge its own tolerance (at least twice its gap), the body's and the other
edges' stay. The result is checked against the input's own SMLib check, the
tolerance-bound tests (#5, #6, #59, #82) exempt.
- Loops of free edges left are holes: offered as surface.patch, and with patch_holes made
(flat only when the loop is planar and every face around it is a plane, else G1) and
sewn in.
The reply reports what changed face by face; nothing to do is refused with why.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Mesh::ray_crossings and InsideIndex counted a triangle crossed by a
Möller-Trumbore test with half-open barycentric bounds (u in [0, 1), u + v < 1).
Those bounds depend on each triangle's own vertex order, so on an edge two
triangles share the ray was in both or in neither: the bracket example's hole
centre lies on a diagonal of the base's top triangulation and was counted 0 or 2
times instead of once (bracket_example_cuts_two_through_holes_and_reopens).
Both now use ray_crosses, a watertight test: each edge's side of the ray is the
signed volume dir · ((p - o) × (q - o)), which is exactly antisymmetric in
floating point, so the two triangles of an edge see it with opposite signs, and a
ray exactly on an edge's line is put on one side by the edge's endpoints in
lexicographic order. New geom test ray_crossings_count_a_shared_edge_once crosses
a grid at every inner edge, straight down and obliquely.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A G4 blend curve is one Bezier of degree 9, and the face extruded from it a spline surface of
degree (9, 1). The record accepts splines up to degree 16 (Brep::validate), but the evaluators
behind pcurves, tessellation and Continuity Analysis capped the degree at 7
(MAX_SPLINE_DEGREE), so the SMLib export refused the face with "pcurve: not supported yet: a
spline surface of degree (9, 1)". The cap was not an algorithmic limit: every routine under it
is the general NURBS Book algorithm, sized by fixed arrays of MAX_SPLINE_DEGREE + 1 slots.
MAX_SPLINE_DEGREE is now the record's own bound, 16, and Brep::validate reads the same
constant. The display evaluator's first-derivative basis (tess::basis_ders) works on two
one-dimensional tables (A2.2 at degree p and p - 1, eq. 2.7) instead of a square one, as its
GPU twin does, so the larger bound costs a low-degree surface nothing.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A torus face covering most of its tube off the tube's quarter lines (an
exhaust collector's bend: 326° of the tube from −113.8°) needed a fifth
quarter along the tube, which STEP counts from the outer equator: its form
ran past a full turn of the tube and SMLib refused the solid. #8496 turns
frames about the axis only, and the torus label had no tube reference.
The torus label gets a tube start (`tube_start`, radians; 0 and left out
of the record by default, so older records mean what they did): v counts
from there, in every evaluator, form, pcurve, iso curve, tessellation,
placement, SSI and SMLib transfer. A spilling torus face's tube start moves
to where its tube begins, or a turn before where it ends, whichever puts
its tube circles on the grid of a round face they are about the axis of
(a pipe's cylinder meeting the bend); circles round the tube are labelled
from the tube start (brep::form::on_tube), so their forms and pcurves lie
on the face's quarter grid exactly. The saved-format change is one file,
crates/doc/src/format/pending/torus-tube-start.rs.
Engine assembly: COLLECTOR_EXHAUST_VER_1 now comes in as a closed solid
(refused 3 -> 2); Poppy Humanoid unchanged (4).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
uv::closed tells a spline closed on itself (periodic, one knot interval per period) by
its first and last control rows and knots clamped at both ends, checked as k[0..=p] and
k[n..=n+p] all equal. Over its interval [k[p], k[n]] a spline of degree p never uses
k[0] or k[n+p]: it starts at its first control point and ends at its last when
k[1..=p] and k[n..n+p-1] repeat. A periodic quadratic circle written with knot
multiplicities 1,2,...,2,1 (as OpenCascade writes one) was taken as open, so its seam
edge's two uses got one pcurve and the face was refused as a loop across a closed
spline surface's seam without a seam edge (periodic_knot_tube_faces_import_and_mesh:
"pcurves failed on 2 of 2 faces"). The check now skips the outermost knots; the knots
the existing test calls unclamped (uniform) stay open.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A cylinder, cone, sphere or torus face of one whole turn has a NURBS form of the four
quarters from the one its seam starts in (brep::form::face_quarters). When the seam is
not on a quarter line of the file's frame, the seam's second use lies a period on, past
the form's end: the face's domain stopped at the form's end and SMLib measured a ball
whose seam meridian is 0.7 rad off its frame's x axis 2.2% short
(tilted_seam_only_sphere_imports_as_a_whole_ball: 32.03 vs 32.74 mm3), and a 5 mm ball
with its seam 1 rad off, holding a 2 mm void, 481.0 vs 490.1
(seam_only_sphere_void_is_turned_inside_out).
turn_quarters, which says whether turn_frames (ee4d4b5a5) turns a face's frame, counted
such a whole turn as four quarters, never a spill. It now counts the quarters the face's
angle interval needs on the grid (brep::form::quarters, as for a face of less than a
turn): five for a whole turn off the quarter lines, so the frame is turned about its
axis to put the seam on one, and the form's seam is the face's.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A cylinder, cone, sphere or torus face of less than one turn that does not
start on a quarter line of its frame, with more than three quarters left to
it, spilled into a fifth quarter of the quarter grid: its NURBS form ran past
one turn and SMLib refused the solid ("an edge lies 12-16 mm off the face's
surface"), which came in as surfaces.
Before its edges are labelled on it, such a face's frame is turned about its
own axis (brep::form::turned: the same surface, its angle parameter measured
from further on) so it starts on a quarter line and its form spans four
quarters. Round faces a circle edge about both their axes joins share one
quarter grid, so they turn together by one angle, chosen so every one of them
fits in four quarters; a group with a face that cannot be turned keeps the
file's frames. Each face's box on its label is measured once and reused.
Engine assembly: PULLEY_WATER_PUMP and MOUNT_SCAVENGE_PUMP_01 now come in as
solids (refused 5 -> 3); Poppy Humanoid unchanged (4).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DtL5Lzv2n8fyKqD61NCHSF
The STEP import's mesh check made a coarse mesh of every part (planned,
evaluated, its seams and smooth edges flagged) only to see which faces had
triangles, and threw it away. A face has triangles in the evaluated mesh
exactly where its plan has a triangle on its own nodes, so the count is now
read off the plan (TessPlan::faces_with_triangles, the condition evaluate
keeps a triangle by).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A face's form covers the quarters of its box, and the box is the extent of its
boundary's points, which run past a quarter's end by as much as the edges lie
off the surface (nanometres in a file of 1e-11 mm digits). A half tube's form
then took the quarter past each end and went round the whole tube, which SMLib
does not take for a torus: 16 parts of an alternator came in as meshes. An end
within the tolerance of a quarter's end (the angle on the face's widest circle
there) is now on it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DtL5Lzv2n8fyKqD61NCHSF
A face's boundary node on a patch side collapsed to within the tolerance of
the vertex there (capped at the chord tolerance), not to a point, is a pole:
its parameter along the collapsed direction is free, as the record check and
the STEP loop check take it (`uv::collapsed`, shared with
`SurfaceEval::pole_within`). A loop through such a corner no longer winds
round the surface as projected, and plans strictly (Poppy Humanoid's
connector, face 43: a quarter of a sphere cap on a closed rational spline).
A face the strict plan and the lenient fill both refuse no longer fails the
whole body's mesh: it is left out with no triangles, named in
PlanStats::unmeshed (logged by the display plan, counted by
Body::unmeshed_faces), and the rest of the body is meshed; a body none of
whose faces plan is still the first face's error. Body::volume_at gives no
volume for a mesh with a face left out.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
An edge raised to cover its pcurves' gap could get past SMLib's limit
(Congé19: 0.98 mm off, tolerance 1.47 mm) and its whole solid then failed
as "SMLib rejected the input". The tolerance an edge or vertex would get,
not only its gap, is now held to that 1 mm: farther off is refused with
how far. The record mapping refuses a tolerance past it with a message
rather than handing SMLib what it rejects.
The margin over a surface gap moves to solidcraft_geom::brep::pcurve
(SURFACE_GAP_MARGIN) for SMLib's pcurves to use too.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
EvalSurface and SurfaceEval give every partial derivative up to total order four: in closed
form on cylinders, cones, spheres and tori, by basis-function derivatives and Leibniz's rule on
rational splines, and by the rotation's own derivatives on a revolution (which had no second
derivatives before). normal_section turns them into the curvature of a normal section and its
first two arc-length derivatives by solving the section's Taylor series order by order.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LUXeKMoxpM7Vuv8Rb6E3K2
A corner ball's edge may run from its equator over the sphere's pole to a vertex just past
it (Poppy Humanoid's creative_commons_license: 1.8° past, 0.016 mm off the pole). On the
sphere's parameters the pole is a line, and the edge's pcurve would have to jump half a turn
along it: "pcurve starts 2 from the end of the use before it, on another period". Each
edge that passes through a pole of a face it bounds (a sphere's, a revolved face's on its
axis) inside its interval is now split at the nearest point to the pole (golden-section
search from 16 samples), the new vertex placed on the pole; the loop's half turn is then at
a vertex on the pole, where a loop may run along it. Each solid's notes say how many edges
were split. The geometry check takes a vertex at a pole to within the vertex's tolerance, as
the reader already does (SurfaceEval::pole_within).
Without the split the edge's jump read as a 0.024 mm gap and was covered by a tolerance:
the measured gap now only covers what the file's geometry really has.
Test: a spherical triangle (radius 0.5) with an edge over the pole to 1.8° past it reads
with its spherical-excess area and its edge split.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LUXeKMoxpM7Vuv8Rb6E3K2
A range sampled in even steps ends at a + (b − a)·1, which may round one ulp past b, where
find_span gave no span and every B-spline evaluation there failed. The import's loop area
(the reader's sampling of each pcurve, used to turn over a face whose outer loop runs
clockwise) then had no value, the face was left as written, and the record check refused
it: "outer loop runs clockwise in (u, v)" on 7 of Poppy Humanoid's solids. The same step
had refused closed-form pcurves (426924937). find_span now takes a parameter within 8 ulps of
the domain's magnitude past an end as that end; farther is outside, as before. It fixes the
pattern for every caller rather than each sampling loop.
Poppy: 232 → 242 bodies, 15 → 7 solids not imported; the kernel tests as before.
Test: Poppy's interval's last step rounds past its end and finds the end's span; 1e-9
past is still outside.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LUXeKMoxpM7Vuv8Rb6E3K2
The closed-form pcurve's lift check sampled t0 + (t1 − t0)·k/n, which at k = n can round
one step past t1 (1.2985843833752289 for an end of 1.2985843833752286), where the pcurve
on [t0, t1] has no point: the closed form was refused, and on a plane, which has no
parameter box for a kernel's pcurves, the face with it ("a face on an unbounded surface
without closed-form pcurves", Poppy Humanoid's Combine1 and others). The end is now t1.
Test: Poppy's interval on a spline on a plane; without the fix, "no exact pcurve of a
B-spline on a plane".
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LUXeKMoxpM7Vuv8Rb6E3K2
Brep::validate lets a pcurve's interval differ from its edge's by PCURVE_INTERVAL (1e-9
of the magnitude: rounding room), but check_geometry then evaluated the pcurve at the
edge's own ends and samples, which may lie just past it: "the pcurve has no point at
2.0000000000082903" on four of Poppy Humanoid's solids, circle edges whose end vertex
sits on a quarter of their form, where SMLib's pcurve ends on the knot. Both of the
check's pcurve readings (the use's points and the loop's area) now take the edge
parameter on the pcurve's interval, through one reader.
Poppy: 195 → 199 bodies, 41 → 37 solids not imported; the corpus survey and the kernel
tests are as before.
Test: a pcurve whose interval ends 1e-11 short of its edge's validates and checks;
without the clamp, "the pcurve has no point at 1".
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LUXeKMoxpM7Vuv8Rb6E3K2
The pcurves close a face's loop through a pole to within the face's tolerance (a
point where crossing half a period moves less than that): there the periodic parameter
is free and the loop jumps across. The reader's check that no loop crosses a seam
without a seam edge took a pole only exactly (1e-9 relative), so it refused the loops
the pcurves had closed near one. Poppy Humanoid's nearly round pin ends (a spindle
torus of major radius 0.005, minor 0.345) have their apex vertex at the tube's top,
3.6e-5 mm above the true pole: their meridians end 1.7e-6 rad short of it, and 13
faces were refused. Both now use one test, SurfaceEval::pole_within.
Test: a point 2e-7 mm from a sphere's axis is a pole within 1e-6 mm, not within 1e-7.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LUXeKMoxpM7Vuv8Rb6E3K2
fill_loop's bicubic height field over the loop's plane, fitted to the
neighbours' derivatives as well as the loop: every condition of a graph is
linear in its heights (G1: the gradient from the face's normal; G2: the second
derivative along the face's leave direction from its normal curvature), so
one banded least-squares solve with the thin-plate energy decides the surface.
Its parameters stay plane coordinates, so an edge's pcurve is still the edge
projected onto the plane, exactly.
Where a round's section meets a plane at a corner of the loop, the section has
the round's curvature and the plane is flat across its own curve: no surface
is G2 to both there. The G2 rows fade in over 15% of the run from each corner;
asked over a few cells instead, the plane's flatness to the second order along
its curve pulls the surface off the whole side. Grids go to 128 spans, which
G2 to a plane along a curved line needs. The thin-plate weight is a thousand
times fill_loop's: with the slope and curvature rows acting a little inside
the loop, the weaker one let the surface rise past a plane it is G2 to (0.06
mm at a box corner of r5 rounds set back 8).
loop_frame is fit_once's plane, shared.
Tests: the hole a setback leaves at a box corner of r5 rounds set back 8,
within 2.2e-5 of the loop, 0.027 degrees and 4.7e-4 per mm of its neighbours
clear of the corners; set back 7.5 to 15, under its planes but by under 7.2e-4;
a hole in a cylinder filled with the cylinder's own second order; refusals.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
An edge is left out of the shaded view only where its two faces meet
with one tangent plane (normals within 0.5 degrees) and one curvature
across the edge: at each station the faces' normal curvatures square to
the edge agree within 0.5 degrees' worth, |k1 - k2| <= 0.0087 x
max(|k|, 1/L), L the edge's length, so the continued faces turn apart
by no more than that over a radius of curvature or the edge's length.
Fillet boundaries (0 to 1/r), corner-patch joins and a plane tangent to
a cylinder are drawn again, straight ones included. A modelled thread's
flank pieces (one sweep) differ by at most 0.0046 and stay hidden.
Second derivatives of a spline surface now exist at a boundary point a
pcurve carries a rounding step past the knot range, as its point and
first derivatives already did.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The branch rebased onto main keeps each of its commits; where both sides
changed the same code, these are the pieces neither side's commit had.
- Cargo: one [features] table in the CLI's manifest (two met in the merge).
- geom: a NURBS surface's `closed` set once in a placed copy (both sides
added it).
- kernel: the whole-face extrude tools (#2040) work around truck's boolean
on truck types, which left the workspace with the SMLib kernel; the
module, its tests and the doc crate's tool and tests go. The display
planner's flat fill (tess/fallback.rs) has no caller with the schema 4
mesher, and the tessellator's seam and uv modules are the branch's. A
body's boundary tolerance from a STEP file belongs to the truck body the
schema 4 record replaced: `with_boundary_tol` goes. `unmeshed_faces`
counts the coarse mesh as the branch does. A loft's section re-split
passes the lead's perimeter shares to main's `resample` (a circle breaks
there when the matched corners do not go round in order). Main's tests of
the mesh-based measure (the SMLib kernel measures exactly) and of the
IGES preprocessor name (no such field on the branch's import record) go.
- doc: a feature result carries no boolean diagnostics on the branch, so
the model cache neither writes nor reads them, and a pattern copy's miss
test runs the intersect as it is. A placed body's worker result is a
Result: a body taken from the previous placement is Ok. A thread's
clipboard remap takes main's arguments, the tapped catalogue names a
metric size by ThreadForm::Metric, and main's tests unwrap the placing
calls that return their error.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
At a pole or apex `EvalSurface::eval` takes the normal from a step beside the
point, trying (+h, +h) first. At the far end of an interval (a sphere's north
pole at u = 1) that step lands past the end, where the extrapolated surface
folds back over the pole, and the normal came out turned over: the pole vertex
shaded dark with a light ring around it (ladder 01/005, every sphere and dome).
Steps now stay inside the active interval of a spline or a revolution's
profile. The pole test checked only that the normal was radial, either way;
it now checks the side too.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
form::extent took the box of 32 points per edge. A face's NURBS form is built
over that box, so where the boundary peaks between two samples the form
stopped short of its edge: the re-imported rod's bore ended at v = 9.99977
while its intersection with the rod reaches 10, and SMLib found the edge
1.9e-4 off the bore (SmBrep #4, SmEdge #13). Each sample higher or lower
than its neighbours now brackets the edge's extreme there, found by golden
section.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0113gDVmStT5tCGKW9Fxgc5K
form::extent unwrapped each loop of a face from its own first point, so a
hole beside the seam (a cross bore through a rod at the rod's seam angle)
came out a period away from the outer loop. The box over both was more than a
turn, cut to one turn from the hole: the face's NURBS form ran over quarters
3..7 while its seam uses lay at 4 and 8, and SMLib found the re-imported
rod's edge ends outside its face domain (SmBrep #70-#72). Each loop after the
first now goes the whole periods along that keep the face's box smallest.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0113gDVmStT5tCGKW9Fxgc5K
With touching: separate (the default) a join groups the joined body's
pieces by face contact. The groups that hold the material of the body
before the feature stay in it (each unless it touches one already kept;
the first group where the lineage doesn't say), and so do the groups that
touch nothing: a virtual grouping later operations take as one body. Every
other group that touches another along an edge or at a point becomes a
body of its own, and the feature's notice says so. The joined body keeps
its name and never holds an edge or point contact. A join that only
grouped pieces apart gives no notice, so KeptSeparate::Apart is gone.
one_body still keeps every piece in one body.
SMLib's export copies an edge or vertex that touching solids share once
per solid, so the record never listed those contacts:
Brep::region_contacts now takes such copies (the same curve and interval
between vertices at the same points, a vertex at the same point) as the
edge or vertex they share. A contact SMLib leaves out of the topology (the
3612 spoke tangent to its hub) is found on the exact geometry, the least
distance within the record tolerance, and reported as KeptSeparate::
Touching ("along a line or at a point").
The pending format change join-touching-bodies stays: older designs kept
pieces touching along an edge or at a point in one body too. Its notes say
so, and the reopened legacy emboss still expects touching: one_body.
Tests restored to their form before cc49081df and 733d3bd19: kernel-cases
2896-join-arm-end-near-ring-wall-gap (one body of two shells), 3379, 3455
and 3526 (no touching: one_body); the engine tests of posts joined while
apart (twice), two cubes for interference, a ring with a cylinder, the
buried tab and a_join_adds_only_its_sketch's 0.1 mm gap.
join_meeting_nothing_keeps_the_first_body expects one body and no notice.
The #3378 spoke stays 1 mm inside the hub.
New tests: blocks touching along an edge are split with the along-an-edge
notice; blocks apart stay one body without one; a ring with an arm on its
corner edge and a hub apart gives ring and hub in one body and the arm on
its own (doc); the engine's notice for an edge contact; the record's
contacts of cubes sharing an edge or a vertex (kernel).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Brep::region_contacts lists each pair of material regions that shares a
face, an edge (used by faces of both) or a vertex, with the most it
shares; pairs not listed are apart. Body::region_contacts reports it for
a boolean's result (lump positions as in Body::lumps), and Body::parts
cuts a body into groups of lumps (lumps is parts of one each). Exact
facts from the record only: no geometry, no decision.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017itGh5fpMEousnZ91cYcVt
(cherry picked from commit 64a1d688e68e5f0c5b1f9e4fe0d2facfaa4744be)
A degree-1 patch of four control points on a parallelogram (a plane as an IGES
or STEP B-spline writes it) is that plane at every parameter, so its pcurves
take the plane's closed forms instead of going to SMLib (which drew the rim
circle of 3806's base as a 385-point spline).
SMLib may run a one-face sheet's loop against ours. A use's pcurve is the same
either way, so the loop's way round is read from an edge the face uses once and
every use is matched with it (3806 failed with 'a use SMLib did not keep').
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BnQawySUFx57iaJZq7z1Fm
exact_pcurves gives each use's pcurve where it has a closed form (or is given)
and None where it has none, without closing the loops; pcurves_within closes
them once every use has its pcurve. TrimFace, TrimEdge, TrimUse and the Trims
trait carry a face to a kernel that makes the pcurves of the uses with none.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BnQawySUFx57iaJZq7z1Fm
along took u or v the edge's own parameter at slope ±1. A rim SMLib makes (an
extruded or revolved circle, four quarters with scaled knots) runs the face's
quarter grid at another steady pace, and a closed rim's two ends are one point,
so neither its ends nor slope ±1 give its line. The pace is read from the edge's
points a quarter and three quarters along, unwrapped from the middle; the line
is taken only where it lifts onto the edge within the tolerance.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BnQawySUFx57iaJZq7z1Fm
Validation refuses a circle or ellipse edge, or a cylinder, cone, sphere,
torus or revolution face, without its NURBS form. Test fixtures are built on
forms; seam and domain defects use the form's parameter (one turn = 4).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BnQawySUFx57iaJZq7z1Fm
- A meridian of a sphere or a tube circle of a torus (a circle in a plane through the axis) has
its exact pcurve on the label: u fixed at the arc's azimuth, v its angle from the outward
radial, alpha + s t.
- A placed revolution has a NURBS form: its profile's control points turned about the axis on
the quarter grid from angle zero, weights multiplied, then its matrix; the record takes forms on
revolutions.
- `form::extent`: a face's parameter box from its boundary projected on its label (the collapsed
parameter at a pole bounds nothing; a boundary that never goes round covers the turn), what its
form is built over with no pcurve involved.
- A face that goes all the way round gets a form of exactly one turn, closed and so periodic.
- `along` projects inside the face's box (a spline extended past its knots, or a profile turned
past the face, has other preimages), and takes a closed form within the record's tolerance, as
the closed forms on analytic surfaces are.
Tests: fits are gone, so curves with no closed form on their face (a tilted circle on a sphere,
an ellipse on a cylinder, an intersection record, a curve weaving across a thread flank) are
asserted refused; the closed-spline fixtures are built as forms. geom 166/166.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BnQawySUFx57iaJZq7z1Fm
`On::fit` and everything it used (samples, least squares, breaks, filled poles) are deleted, and
the STEP reader no longer retries a face at ten times its tolerance. A pcurve is exact (closed
form, an affine image, or a straight line along a form's parameters) or the face is refused:
"not supported yet: no exact pcurve of <curve> on <surface>".
The rig and ladder are as with the fit (333/372, 101/116; none lost; 3806, already failing, now
refused on import). 16 unit tests that build sphere meridians, torus seams, closed spline
bands, intersection curves and surfaces of revolution on their labels now refuse: those paths
need meridians labelled on their face's meridian plane and revolutions on the quarter grid.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BnQawySUFx57iaJZq7z1Fm
`form::edge` is the one rule sew and the STEP reader share: a circle about a revolved face's
axis is labelled on that face's frame, and a circle or ellipse edge gets its form. The STEP
reader now gives every imported edge its label and form before any pcurve, and every
cylinder, cone, sphere and torus face its form over the extent its label's pcurves give; the
pcurves, the outer-loop choice and the turn-over test then run on the forms, and the record
carries them.
With the pcurve fit turned off, the STEP and IGES circle cases pass (8 of the 9 import cases
that needed it; 3208's ellipse on an extruded spline surface remains).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BnQawySUFx57iaJZq7z1Fm
The bend's rim arcs carried frames of their own (x toward the arc's start), a fraction of a
quarter from the bend cylinder's, so their forms were off the cylinder's quarter grid and no
straight pcurve lay along them. `form::on_frame` relabels such a circle on the face's frame (the
same circle, its interval shifted); its form is then on the face's grid, its pcurve u = ±s + c.
`along` takes u or v as the edge's parameter either way along, offset. The volume test steps to
a pcurve's ends exactly (t0(1-f) + t1 f), as an interval's end may now carry rounding.
With the fit off, 38 of the 44 sew cases pass (from 29). With it on, the rig is unchanged
(333/372, ladder 101/116, no case changed).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BnQawySUFx57iaJZq7z1Fm