The kernel switch broke these tests only through their counts, which record how a kernel splits
faces rather than whether the result is right. Each now checks its intent:
- a fitted circle is the exact circle (centre, radius, a full turn), not "one segment";
- the tight bend sweeps a valid solid of Pappus' volume, πr²(40 + 5π/2), not a count of faces,
edges and vertices;
- the Pi case keeps each outer side one face, stays valid after every feature and rebuilds the
same topology at 32 whether built there or edited there, without a hard-coded table of totals;
- a certified paint stays on the wall's radius and covers as many faces after an edit, a copy
and reopening as when it was painted;
- the offset sleeve and the extended strip are checked by their exact areas only.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
box_and_fillet_have_stable_wire_records compared the encoding of a box and a filleted box
with records truck built (2f4d6acac). Since f11c0c571 an operation's result is the record
SMLib gives, which numbers the box's faces in its own order and the fillet's edges by its
history, so the golden differed in face order, face sources and edge history, not in the
encoding. Rewritten with NATIVE_BREP_BLESS=1 and checked against the geometry: the box 6
planes, 12 edges, 8 vertices, each face source's plane its face's; the filleted box 7 faces,
15 edges, 10 vertices (one edge replaced by a face: +1, +3, +2), the round a cylinder of radius
2 along x on the axis (y, z) = (2, 8), 2 from y = 0 and z = 10.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Since 09c979a4b a full circle in a profile is one arc of a whole turn, so what is built from
it is one closed edge and one wall closed along its seam. Five kernel tests still expected
truck's halves:
- a_small_hole_in_a_large_spline_face_cuts_cleanly: the Ø1 hole adds one face (its wall),
not two; the removed volume (π · 0.25 · 40) is unchanged.
- fit_loop (two mesh_convert tests): a fitted circle is Loop2::circle, one segment.
- tight_bend_counts: a Ø10 tube swept up 20, round a quarter turn of radius 5, along 20. The
circle starts at (5, 0), which runs up to (5, 0, 20) on the bend's axis, where the turn
leaves it: the bend face has no seam and both straights' seams end there. By the ladder's
rule (each whole curve once, each seam): 5 faces (2 caps, one per path segment), 6 edges
(4 circles, 2 seams), 3 vertices (the seams' ends: (5,0,0), (5,0,20), (25,0,20)), the
record's own counts.
- extruded_circles_go_as_cylinders: the record holds the wall as the cylinder it is (one
face of three), no longer a rational spline; it still goes into SMLib as its cylinder.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A folded edge's Hermite pcurve on a bend's form breaks where the edge's angle crosses a
quarter. At the 120 degree corner with flanges at 60 and 90 degrees under a tear, an edge
starts 3.25e-12 of its parameter short of a quarter: the crossing was kept as a break
(breaks within 1e-12 of the parameter merged), the pieces halved to 1.6e-12, and SMLib,
given knots 1.6e-12 apart, handed the pcurve back with 15 equal knots at degree 7 ("SMLib
export: pcurve: loop 0 use 1: given pcurve: portable B-rep: knot multiplicity"). A crossing
whose point on the edge lies within the pcurve's lift accuracy (FOLD_TOL of the edge's size)
of the previous break is that break; the end is never merged into the start.
Fixes sheet::tests::every_pair_of_angles_is_torn_or_keeps_its_notch (torn pairs 37, 37, 27).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A triangle whose corners lie on one line (or weld into one another) adds no
surface, so welding now drops it, with any corner only it used, instead of
refusing the whole mesh as having a degenerate triangle. The closed and
consistently oriented check that follows still refuses a mesh that needed it to
close, and a mesh of nothing but such triangles is refused by name. Dropping its
unused corners keeps a dropped triangle from widening the part's extent, which
the prism search reads from every position.
Regression test a_zero_area_triangle_is_dropped: the 20 x 10 x 5 box with the
case's extra triangle (0,0,0) (10,0,0) (20,0,0) converts to six faces and
V = 1000 mm^3.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
SMLib's thickening of the test sheet makes a top face of 72 x 78 control points,
which the 4096-point cap refused (TooLarge) before the solver saw it. The band
cap (2^21 entries, 16 MB) is what bounds the system's memory and its
factorisation (at most n min(n, bw)^2 work); the 72 x 78 bicubic net needs
5320 unknowns 213 apart, 1.14 M entries, well within it. The point cap now
bounds only the per-step energy sweep and the net's copies: 2^14 points, above
the band's own limit for square bicubic nets (about 90 x 90), so a face is
refused for the system it needs rather than its size. The hostile-input test
asks for 129 x 129 and checks the refusal is TooLarge.
a_body_face_is_faired_with_its_edges_untouched_and_its_history_kept passes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Two tests counted the old 2D boolean's edges, which kept a vertex wherever an operand's corner
or edge end lay on another's edge. SMLib's planar boolean leaves no vertex that is no corner:
edges on one line within the tolerance are one edge. The results are the same regions with
fewer, whole edges, and each test's point stands (no edge shorter than the tolerance, no sliver):
- a quadrilateral cut away from below a square whose bottom edge it lies on within 1e-8 (tilted
1e-9 rad) leaves the square's four 10 mm edges (its corner at x = 4 is within 1e-8 of that
edge's line), and its union runs along y = 0 as one 14 mm edge (the quadrilateral's bottom
edge is within 1e-8 of y = 0 all along): [4, 5, 5, 10, 10, 14];
- the bent plate and the strip under it: their ends on the line x = 150 are one edge from
y = 47 to 80, six edges for every turn and offset tried, the shortest 3.
Fixes prism::planar_tests::overlapping_lines_a_hair_from_parallel_do_not_cross and
a_bent_edge_crossing_its_neighbour_by_their_corner_leaves_no_sliver.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Reading a planar sheet's faces back as regions, each hole went to the smallest outline whose
Loop2::contains held its middle: even-odd against the outline's polyline within 1e-3 mm. A
5e-5 mm square cut 1.7e-4 mm inside a radius-10 circle lies outside the polyline's chords there,
so no outline held it and the cut failed ("face 0 has a hole outside its material"). An
outline now holds a point by its polyline within a quarter of the point's distance from it (its
nearest point on the exact segments), which no chord can pass the point by.
Fixes kernel prism::planar_tests::tiny_tool_between_circle_and_display_chord_is_not_lost.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
94bd8719a (#3730) turned a BREP_WITH_VOIDS void inside out when it enclosed volume of the
outer shell's sign: SOLIDWORKS writes .F. over a void shell already facing into its cavity,
SolidCraft .T. That lived in the truck reader; the schema-4 reader (crates/step) that replaced
it took the flag alone, and the block with a 20 x 6 x 5 cavity read as 15600 (the cavity added)
instead of 15000 - 600 = 14400.
The reader now takes each void's sign from the faces themselves before it orders the record:
the divergence theorem on the record's surfaces and pcurves (sew's face_flux of p / 3, moved
to geom's brep::eval so both use it). A void enclosing volume of the outer shell's sign has its
faces turned over. Flipping on the finished record is not an option: its edges are ordered by
the faces' sides.
Fixes kernel step_in::tests::a_void_reads_as_a_cavity_whichever_way_its_shell_flag_points.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The expected 6703.5222 was a mesh measure (truck). The lid exactly: the 76.8 x 89.3 x 2.4
plate less its four 3 mm corner rounds ((76.8 * 89.3 - 36 (1 - pi/4)) * 2.4 = 16441.2344), the
skirt ring ((71.5 * 84 - 68.5 * 81) * 3 = 1372.5), less the window (68 * 61 * 2.4 = 9955.2) and
the slot (44 * 10 * 2.4 = 1056) through the plate, and the notch through the skirt's 1.5 mm wall
(22 * 1.5 * 3 = 99): 6703.534401. The window and slot lie inside the ring's opening and the
notch above the plate, so nothing else overlaps. SMLib measures 6703.534401575, 2.6e-7 off.
The test now computes that closed form and holds the body to 1e-6.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A whole circle is now one closed rational curve (Loop2::circle is one segment), whose first
control point is repeated as its last: the mean of its nine control points sat r/9 towards the
seam, and a rail scaled each station about that point. A rail bulging from radius 10 to 15
(r = 10 + 5 sin(pi z/40)) then grew the circles as if r = 10 + 5.625 sin, volume
4632.8 pi + 9000 = 23553.9 against the solid of revolution's pi * integral r^2 dz =
4500 pi + 8000 = 22137.2 (+6.4%; the engine's arc rail +6.5%).
The middle is now the curves' centroid by length, integral x ds / integral ds (8-point
Gauss-Legendre per knot span on the rational point's speed): the shape's alone, however its
curves are written; a circle's is its centre. geom's GAUSS_8 is public for it.
Fixes kernel loft_rails_tests::a_bulging_rail_makes_the_solid_of_revolution_of_the_rail and
engine loft_rails_tests::a_rail_turns_a_loft_into_its_solid_of_revolution_and_stretching_is_an_edit.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Under the blend-overflow decision (237f0ae74) a convex round running over a face's far
edge (a cliff) is built and cut from the part, so the truck-era fixtures of the reduce
and taper fits (a 2 mm fin, a plate narrowing to 2 mm, a wedge, a 10 mm cube's edge at
R20) no longer refuse their radius and the fits have nothing to react to. Each test now
uses a case SMLib still refuses, and checks the cliff case builds:
- engine fillet_fit_reduce_keeps_the_radius_where_it_fits: a 14 x 6 x 10 post on a
40 x 30 x 20 block, its four verticals at 3.5 (two rounds would cross on its 6 mm ends)
reduced to 3 on 4 of 5 edges, the block's corner kept at 3.5: 24840 - (4 - pi) 90 -
(1 - pi/4) 3.5^2 20. New fillet_fit_reduce_keeps_a_round_over_a_cliff: the fin's R4
builds, nothing reduced, at the kernel's fillet_max_each answer.
- engine fillet_fit_taper_follows_the_room_along_the_edge, spline_fillet_takes_over_a_taper,
kernel taper_follows_the_room_along_the_edge, ui-egui fillet_spline_handle_tests: the
narrowing plate stands against a wall behind its slanted side (z -15 to 30), so its top
and bottom end in concave edges: R3 is refused and tapers 3 -> 1.8 as before. Without
the wall, R3 is cut over the 2 mm top's far edge and kept constant: at x the section
loses (1 - pi/4) 9, or 3 w - (F(3) - F(3 - w)) where the top is w < 3 wide, 6522.7745566.
- The wedge closing to a point: on its own R16 is cut over its cliffs (the section
integral, 4195.578821, SMLib within 4e-8) and kept; its room's taper is checked as a
spline fillet. Against a wall (new fillet_fit_taper_closes_to_a_point) the taper must
close to 0, which SMLib fails today (status 1001).
- kernel taper_mesh_tests: the collapsed taper is built from taper() as a spline fillet.
- ui-egui the_feature_menu_copies_a_warning_message: the 10 mm cube's four verticals at
6 are reduced to 5 (a full round), which the warning says.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The owner's blend-overflow decision (2026-10-25, 237f0ae74): a convex round or chamfer
whose rail runs off its face across a cliff (an opening's rim, a hole, the part's
outline) is built on the faces continued and what it removes is subtracted, so it never
bridges an opening; a round may use up a face; a chamfer of a face's full width stays
refused. These tests still expected truck's refusals. Each now expects the decision's
body, with its answer from the geometry:
- issue_1080 chamfer_into_a_pocket_is_cut_by_it (was ..._is_refused_with_the_size_that_fits):
chamfer 6 into the pocket 4 mm in removes 360 - 20 (its triangle less the part over
the pocket): 14660; 6 along the top and 2 down, 120 - 20/3; 2 and 6, 120.
- issue_1080 fillet_into_a_pocket_is_cut_by_it_and_one_that_fits_builds: R5 removes
(1 - pi/4) 25 along 20 less 10 (5 - (sqrt(24)/2 + 12.5 asin(1/5))): 14893.0344442.
- issue_1080 chamfers_crossing_each_other_are_refused: still refused (they cross);
chamfer_max and fillet_max find 2 mm.
- issue_1080 rounds_on_both_sides_of_a_wall_build_as_one_fillet: R7 alone opens a slit
where its arc passes the inside corner: V0 - (104 (1 - pi/4) 49 - 100 a), a = the square
[25, 30] x [35, 40] outside the circle about (25, 35), 5 (5 - sqrt 24) - (F(5) - F(sqrt 24)):
38458.8958265. Both at R7: the arcs about (25, 35) and (23, 33) cross only outside both
quarter arcs and the inner round's material holds the slit, so V0 - 4 (1 - pi/4) 49 =
39509.9380404 (was expected refused).
- issue_324 a_round_into_a_hole_is_cut_by_it (was the_radius_offered_beside_a_hole_fits):
R4 over the r3 hole: 8000 - 90 pi - (640 (1 - pi/4) - the round's height over the
hole's width, integrated with y = 17.5 - s^2): 7580.2873154.
- issue_851 arc_top_blend_uses_up_the_wall_and_refuses_a_collapsed_one: a round of the
wall's full height uses the wall up (one face for the wall, the removal integrated as for
the other sizes); the full-height chamfer stays refused.
- kernel tests: fillet_max_finds_the_largest_radius_that_works keeps the post (3.5 -> 3)
and the no-edge case; its disc and notch cases move to
rounds_past_the_bottom_of_a_wall_are_cut_from_the_part: the 2 mm disc's R2.5 turned
section, pi (338 - integral of (169 - rho^2) dz), rho = 10.5 + 2.5 cos(theta) (the same
formula gives SMLib's R2 994.15188 exactly), 957.6922389; the notched block's outline
R7, 11925 - (1 - pi/4) r^2 (140 - 8 c), c = r (10 - 3 pi) / (3 (4 - pi)) (SMLib's R7 within
5e-8), and R8 less its part over the notch, integral of 2 (5 g - g^2/2), g = 8 - sqrt(64 - t^2).
- kernel tests: chamfer_wider_than_sloped_basin_rim_is_cut_and_a_full_width_one_refused:
past the 1 mm and 0.5 mm rims the chamfer runs over the slab's front edge and is cut:
the clipped triangle's area (30 + z centroid), its length 60 - h between the sloped end
walls (the same formula gives chamfer_sloped_basin_rim's fitting chamfer to 1e-12);
a chamfer as wide as the rim is refused.
- blend_corner sloped_rim_fillets_stay_inside_the_material, equal_slope_rim_loops_stay_
inside_the_material: R4 and chamfer 2 over the thin rim build, every vertex inside the slab.
- engine blend_fit chamfer_through_the_openings_is_cut_by_them: the report's 9 mm chamfer
is cut by the openings, 3350 (the rig's fillets/1080 case).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Since bcbb99005 a blend larger than its faces is SMLib's refusal (KernelError::TooLarge,
"the fillet is larger than the faces beside its edges allow"); truck's room check, whose
words these tests still matched, went with blend_fit. The refusals themselves stand:
- fillet_consuming_a_side_face_makes_a_full_round: R3.001 and R3.5 on the 14 x 6 x 10
box's four vertical edges, past the full round, are refused as too large (no longer
"larger than the neighbouring faces"); the body unchanged is its own volume before
(SMLib's measure of the box is 839.9999999999998, not 840 bit for bit).
- a_chamfer_that_would_grow_the_part_on_a_convex_edge_is_refused: refused as too large;
the word "convex" was the old check's.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
When the kernel's room pre-check went (bcbb99005), a fillet or chamfer refused as
too large stopped saying what would fit: only SMLib's NO_ROOM words were left. The
rig expects the refusal to name it (fillets/4782-fillet-two-end-verticals-past-a-full-round:
"the largest radius that fits is 7.5 mm", fillet_max for the exact fit).
A constant fillet with the exact fit, and an equal-distance chamfer, refused as too
large now add "; the largest radius (distance) that fits is X mm", found by the same
bisection fillet_max runs (up to 16 blends, three significant digits), now shared with
a new kernel::chamfer_max. Where the search itself fails (a candidate failing for
another reason), the refusal is left as it was. The reduce and taper fits and the
kernel's own searches call the kernel blends directly and do not pay for it.
Tests: the 14 x 6 x 10 post's R3.5 names 3 mm; a 4 mm rib's crossing chamfers and
rounds at 3 name 2 mm.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
sweeps_along_paths expected one wall for the circle swept through 41 points of
a smooth curve: a point path is the polyline its points make, swept line by
line and mitred at each point since 43e8b0349 (sweep_path's own test expects
40 + 2 faces since d9274aaf1), so the circle's one segment has a wall along
each of the 40 lines.
The holed square along the L-shaped path expected two faces per segment. The
corner turns the profile about Y, so each line at a constant y (the outer
loop's segments 0 and 2, the hole's 1 and 3) sweeps the plane y = c it lies in
along both legs, and the two walls meet edge to edge on the mitre in that one
plane: one L-shaped planar face. The lines at a constant x sweep two planes,
two faces. check_sweep takes the count per segment.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
sweep_origins read a sweep_path tool's faces by the truck face order (walls in
profile order, then the start and end caps), and fell back to swept_origins,
which names a face without an edge in the profile's plane an end cap. SMLib's
sweep along lines and arcs is made piece by piece and mitred at each corner
(tangency-labs/smlib#43), so most walls start further along the path: 39 of a
44-face sweep came out as end caps (105 of 110 before the previous commit). The
sweep records each face's origin as it builds it (Body::history), and
sweep_origins now returns those, finding by geometry only a face the tool names
no origin for. swept_origins says what it holds for.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A sweep along a path with corners sweeps each run past its corners and cuts it
back to the plane halving the corner. The lead past the corner was the
profile's reach times tan(turn / 2), enough only for a profile square to the
path. A point standing s ahead of the path and w from it needs
w tan(turn / 2) - s past a corner (w tan(turn / 2) + s before one): a profile
at a slant to its path reaches across the mitre by its slant as well, and a run
carried short of it leaves a notch with a piece of the run's section in it.
The section keeps both distances along a run and through each corner (it
turns with the path; a twist keeps them, an end scale grows them by at most
its largest scale), so the lead is taken from the profile's points at the
start, for the plain and the controlled sweep. A profile square to the path
(s = 0) leads exactly as before.
The walls' history seeds sat at the middle section of each run, which for a
slanted profile on a run shorter than its slant lies past the mitres, on the
next run's wall: half the walls came out generated. Each piece's seed now sits
half way along the stretch its own point keeps between the mitres, carried
there exactly along lines and arcs (as SMLib's sweep turning with the path
carries the profile), at the station there in the controlled sweep.
A flat 4 mm square swept along 21 points of a rising sine (provenance test):
110 faces, 66 of them notch faces, and walls left unnamed; now 44 (two caps, the two
walls in the path's plane, a wall of each slanted side along each line), volume
16 cos(slant) L to 1e-6 (new test a_slanted_profile_meets_itself_at_every_mitre).
The square round a closed 122-line polyline (issue_2005
closed_polyline_sweep_shares_its_starting_section): the first section leaned
by half a chord's turn left a 6.2 mm^2 piece of it at the closing mitre, named
the start cap; now 246 faces, every one a wall.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The test cut three regions from a plate that keeps its meshes and
expected the memo to hold one mesh afterwards: the old kernel's booleans
meshed their operands. On SMLib (ops::boolean -> smlib_boolean, since
42c41464a) a boolean reads the exact record and a tool apart from the
body is found from the exact boxes, so no boolean meshes anything and the
memo is empty after them. The test now asserts that, then that meshing
the kept body twice at one tolerance makes one memo entry, bit-equal to a
fresh mesh, which is the reuse the memo is for.
Fixes kernel body::mesh_memo_tests::regions_cut_from_a_body_keeping_its_meshes_are_unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Split Face records its lines on the body (split_lines): heal and the
certified-planar regularizer keep the faces apart along them, and they
make the body regularization-protected. On SMLib the boolean, the n-way
union and Shell carried the protection over but dropped the lines, so a
body with an authored split lost its record after a Union with partial
contact (issue #674/#782's case), and a shelled one after Shell. Each
result now keeps its operands' lines, each line once
(Body::inherit_split_lines), as the old kernel's coplanar join did.
regularization_protection_tests: SMLib's shell makes no section strips,
so the painted cup no longer has more faces than the ordinary one. The
ordinary cup is the box's five kept faces, their five linings and the
rim: 11. A paint over a whole face authors no boundary inside it, so the
painted cup has the same 11 faces; the split side's line stays an edge on
the wall and its lining: 13. The test now asserts those counts exactly
(it asserted more than 11 for both, the old strip route's witness) and
that the split lines survive.
Fixes kernel issue_674_tests::partial_contact_retains_an_authored_split_away_from_the_contact
and body::regularization_protection_tests::stamped_shell_keeps_authored_painted_and_split_boundaries.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The test helper side_by_side copied the second body's regions by hand,
dropping its voids and keeping its infinite region (region 0) as if it
were a bounded one. Its closed shell was then the outer shell of two
regions, which the portable record refuses ("a closed shell bounds two
regions and encloses one"). The helper now joins the two records with
geom's brep::part::compound, which puts each body's outer shells in the
one infinite region and keeps its bounded regions; the face sources are
still found by plane afterwards, so nothing the test checks changes.
Fixes kernel body::planar_sources::tests::coincident_competing_source_planes_and_tool_planes_stay_unknown.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Body::parts (and so Body::lumps) mapped each piece's faces through the
body's face-source table with a silent `get`, so a table shorter or longer
than the body's faces gave its pieces sources of other faces, or none. The
table is by face position: one whose length is not the body's face count
belongs to another body. The pieces are now refused with "lumps: face
source table does not match the body", as before the SMLib port (the check
went with the old lumps_with_clearance in 66e31bede).
Fixes kernel body::lump_tests::grouping_lumps_does_not_infer_absent_or_misaligned_sources.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Which input plane each result face lies on was found by scanning every input plane whose
normal x was within the resolution's angle: quadratic on axis-aligned parts (two thirds of the
faces have normal x = 0), so it was capped at 16 384 faces a body and 2 M comparisons, and the
join of the 20 level-2 sponges of a level-3 Menger sponge (7 326 faces) was refused past the
comparison budget.
The planes are now indexed by their unit normal, in cells twice the bound on two matching
normals' difference (asin(angular) < 2 angular), and in each cell by their offset from a common
centre, searched in the window linear + |a.o - c| |n_b - n_a| that
d_b - d_a = (b.o - a.o) . n_b + (a.o - c) . (n_b - n_a) gives, plus rounding slack. The index
offers a superset; same_support still decides each match, and candidates are taken in the same
order as before, so every answer is unchanged. A plane with a normal not of unit length is
offered to every query. Both caps are removed: results are bounded by inputs capped where they
entered. The descriptor-size budgets on kept roles are unchanged.
Rig, --known-failing, results compared case by case with develop's: booleans (129) and shell
(28) identical except booleans/5126-join-menger-sponge-level-3, which now passes in 38.9 s on
aarch64 and leaves KNOWN_FAILING.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The curves were held on a 1-2-5 ladder from the world origin, so count 3 on a 40 x 30 face gave 3
one way and 2 the other. They now follow the command's contract: the held parameter at
lo + k (hi - lo) / (count + 1), k = 1…count, over the face's extent. Faces of the body on the same
surface record are spaced together over their combined extent, so a face a boss splits still
shows the same lines straight across the split.
New case cov-isocurve-40-by-30-face-count-3: x = 10, 20, 30 and y = 7.5, 15, 22.5, to 1e-9 mm.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Both walked the picked curve's display polyline: a point a third along an R10 sketch arc was
0.87 µm off (8.6603, 5). They now evaluate the path's exact curves by arc length: lines and arcs
in closed form, a free-form span at the parameter its arc-length integral reaches (body edges by
their record curves, 3D sketch curves by their arc, spline, line or linked exact curves; only a
curve that exists as points alone, an intersection edge or a projected curve, runs along them).
The arc's point is now within 1e-12 mm of (10 cos 30°, 5, 0).
A t outside 0…1 was clamped silently, which the contract (`t: expr, 0…1 along it`) does not
say: it is refused ("t is 1.5: a point along a path is 0 (its start) to 1 (its end) of the way
along it"). The construction cases grade the arc and line to 1e-9 mm, past-the-end is refused,
and a new case takes the middle of a symmetric cubic spline, (20, 7.5).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A face of degree 1 in one direction (an extruded spline) was refused with "a fairness term does
not fit the net's degrees": the thin-plate energy's S_vv asks for second derivatives a linear
direction does not have. On every span that term is exactly zero, so it is dropped; the energy is
the bending along the curve and the twist. A face whose held layers are its whole net (a
single-span cubic with its edges kept, an extruded curve's two rows) has no interior control
point: it is left exactly as it is, its energy reported unchanged, 0 steps.
cov-fair-single-span-face-has-nothing-free-to-move passes and leaves KNOWN_FAILING.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
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>
Drafting a cube's side about its bottom edge (on the neutral plane) tilts the side and moves the
edges it shares with the top and the two faces beside it: those four are trimmed differently
(Modified); the bottom and the opposite side are untouched (Kept). Truck called all six
Modified; SMLib's history called the moved neighbours Kept. Needs the SMLib fix "history: a face
whose edges an operation moved is Modified, not Kept" (SMLib branch devgreen-kernel-records,
cdaf205) and a pin bump.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With SMLib integration-overnight + b78bd4a (the fixes of 2026-10-29), three
expectations that held a refusal now build, each checked against its exact answer:
- rig fillets/1207-fillet-box-end-edge-into-unblended-round (was ...-refused): the end
face's top edge of a cube rounded r3 along its top back edge, rounded r1 alone
(tangent chain off), builds. Its round runs on along y through the r3 round's end
until that surface drops below it at y = 17 + sqrt(5): V = 8000 - 20 (9 - 9 pi/4)
- 17 (1 - pi/4) - 0.218236 = 7957.505202 (the last term the section under the round,
integrated exactly); SMLib gives 7957.505245.
- blend_smlib smooth_edges_and_unfinished_tangent_chains_are_refused_saying_why: the
same cube builds, at that volume.
- modify_history_tests::smlib::draft_tilts_a_side_and_keeps_its_name: the rod's side,
a cylinder along the pull, drafts 5 deg to a frustum, radius 5 at the neutral plane
and 5 - 10 tan 5 deg at the top, at its exact volume.
tests/kernel-cases/fillets/KNOWN_FAILING drops 1557-fillet-vertical-corners-after-
bottom-rounds, which passes now; the other eight still fail, so the gate holds
(81 pass, 8 known failing, no regression).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Since an operation's result is its record (f11c0c571), a mirrored body is
SMLib's transform read back from SMLib's export, and two SMLib defects there
made it invalid: a mirrored cylinder came back inside out the wrong way
(SmCylinder::IsNurbSurfaceCylinder; its edges 10 to 30 mm off the surface),
and every edge's own tolerance was reset to the body's (SmBrep::Transform).
Both are fixed in SMLib (fix/mirrored-cylinder-from-nurbs e02e63e, c603293,
with e008e82 for a scaled body without vertices);
this test needs the SMLib pin moved past them.
A plate with a through hole and a box with two rounds meeting at a corner,
mirrored in x = 30 and read back from their records: SMLib finds both valid
and their volumes are the originals'.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
SMLib makes a torus one face, so the face at (25, 0, 0) was the whole torus,
not a patch within half a turn each way. The patch is now the torus's piece
with x, y > 0 and z > 1, bounded by iso lines both ways (planes through the
axis, and z = 1 round the tube), and its thickened volume is the formula's
over the angles its mesh spans. (The sphere cap above it builds once SMLib's
offset takes an open edge's second trim point away from its first: SMLib
devgreen-kernel-geometry; waits for the pin.)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
SMLib thickens a sheet across a convex sharp edge as other CAD does: each
face's offset is extended until the offsets meet (and trimmed where they
cross inward), where truck bridged the gap with a wedge. Owner decision:
keep SMLib's result. The exact values:
- three faces of a 10 box's corner out by 1: three slabs, a 1 × 1 × 10 block
along each edge and a unit cube at the corner, 300 + 30 + 1 = 11³ − 10³;
- into the inside corner by 1: 10³ − 9³ (it was refused); half each way:
10.5³ − 9.5³ (refused too);
- two faces across one edge: 2 · 100 + 1 · 1 · 10 = 210.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A sweep along points follows the lines between them (no wall is fitted through
the points, sweep_sheet_points' contract, as sweeps_along_3d_paths already
says): along a quarter circle of 41 points the sheet is a wall per line, 40,
its area 10 × the polyline's length (within 10⁻⁴ of 10 × 15π); the bent chain
is 2 × 40 walls sharing every seam, 2·41 + 3·40 edges.
SMLib's sweep takes guide rails along a cornered path when each rail turns
with the path at its corners; a straight rail along a bent path is refused
for that reason, not as "guide rails along a path with sharp corners".
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
SMLib revolves the cone frustum's sheet over angles 0…1 as one face, where
truck made two of half a radian each; the face's own span is angles 0…1 by
profile parameters 0…0.8: √125 · ∫₀^0.8 (5 + 5t) dt = √125 · 5.6.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
SMLib makes a cylinder's side one face closed round its seam, and a sphere one
face; the tests picked "one half" of the side or "one sixth" of the sphere, which
were truck's pieces. The offset test moves the whole side (one face, area
2π·12·20); the thicken test takes half the side from a cylinder split at x = 0
(π·(11² − 10²)/2·20) and the whole side inward (π·(10² − 9²)·20); the sphere
test takes the cap above z = 5 (solid angle π: π/3·(11³ − 10³)).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
prism_measure's uncertain-material check built a 20 x 20 x 1e-5 box and asked that it get no
cap certificate. SMLib refuses that box (its loops fail SMLib's orientation check): at 1e-5 it
is half a body's tolerance (2e-5) thick, no solid. Refused or built without a certificate, the
property holds.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
edge_at (matchsurf) and the mixed fillet mitre check evaluated `e.curve` at parameters from
`e.interval`, which is the interval on the edge's NURBS form when it has one (SMLib's edges
do): they picked and sampled the wrong points. They now evaluate `e.parameters()`, as the fill
helpers do since aa534cc6d's fix on the kernel-geometry branch. (Both tests still fail, for
their own causes: the G2/G3 match leaves a curvature jump of 2 at the seam, and SMLib's mixed
blend lies 4e-7 off its section circle.)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Truck split what it swept: a big arc into two faces, a full circle into halves, a turn into
two or three parts, a sphere into six patches, a torus into six; it numbered boxes' faces in its
own order and sewed stepped cuts slab by slab. SMLib sweeps each profile segment into one face,
closes a full cylinder, sphere or torus along one seam edge, and cuts without slabs. The record,
face source, helix, display mesh, measure and boolean history tests still expected truck's
pieces. Each expectation now follows the geometry, checked by the tests' own independent
measures: every face's origin against the profile segment its mesh lies over (unchanged),
faces picked by where they lie instead of by truck's face numbers, and the merged counts by the
kernel-cases ladder's rule (each whole curve once, and each cylinder's seam): a cylinder 3
faces, 3 edges, 2 vertices; a ball cut from a slab 7, 14, 10. The truck-only stepped-prism
check (walls coalesced across slabs) becomes the property it stood for: the stock's walls stay
whole, one face each, Kept.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
SMLib's draft extrusion takes one loop, so a tapered extrude of a region with holes was refused
("SMLib draft extrusion of a profile with holes"), and designs that use one no longer opened
(modern-plastic's Rest: design_corpus, action_matrix a_small_hole_through_the_wall_of_a_plastic_rest).
As other CAD does, every wall now leans by the taper relative to the pull on the material's
side: the outer loop is drafted, and each hole, taken counter-clockwise, is drafted the other
way and cut from it, so a taper that narrows the outer boundary widens the holes. Exact: each
loop is SMLib's draft extrusion of it, the walls' roles are found as the foot's are. Refused,
saying why, where the taper would make a hole meet the outer wall or another hole at either
end of the extrusion.
Test: a 40 mm square with a 10 mm square hole, 30 up at 5° inward, is the one frustum less the
other to 1e-6 with ten faces; at 30° the hole would pass the outer wall and is refused.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A face's edges are projected exactly: lines, and arcs read off SMLib's NURBS within the
planar-edge tolerance (Body::planar_edges). An SMLib shell's offset corner is a NURBS whose
arc came out with its centre 2e-6 mm off and its ends up to 6e-7 mm from the lines they meet,
beyond the link's merge tolerance (1e-7): the inner loop of the sensor enclosure's case top
projected as sixteen loose curves, the offset of one found no chain to follow, and the Lid lip
extrude found no profile l13..a16 (the five acceptance scenarios, action_matrix
an_axis_through_a_corner_round_that_a_parameter_moves). Each edge's pieces now end on the
edge's vertices where they lie within the edge's and the reading's tolerance of them (the
vertex is the point both edges share), and an arc whose ends moved takes the centre nearest
its own that is equally far from both, so the sketch keeps both ends on its circle.
Body::planar_edges_tolerance names the reading tolerance planar_edges uses.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
map_planar_sources took every plane of a cut's tool as competing with an input face on it,
facing either way, and dropped that face's exact role. A text pocketed 1 mm down from a block's
top has its tool's top on the top's plane facing up, as the top does; the top then lost its
CertifiedExtrudeCap, and the sketch on it reported "the top cap of Extrude1 no longer exists"
(a_face_sketch_saved_with_fitted_letters_reopens_and_builds_the_same). A cut's tool faces enter
the result turned round, as the walls of the hollow, so a tool face can only take the place of a
face on its plane facing against it (a pocket's floor, a face the tool sat on from outside). Only
that orientation competes now; the existing blocking cases (a tool's bottom on a pocket floor, a
tool face against a wall) are unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
same_support took two planes as one support only within 4096 ulps (normals 9e-13 apart,
origins about 1e-12). Faces SMLib rebuilds through a boolean lie within its own resolution of
the plane they were made on, not within ulps: a cap at z = 10 notched by a cut came back with
its origin 2.1e-11 off and its frame turned 1.36e-12 rad, so its CertifiedExtrudeCap was
dropped and the sketch on it reported "the top cap of Extrude1 no longer exists".
Supports are now compared at SMLib's own coincidence rule, SmPlane::CoincidenceCheck: normals
within the angle the tolerance makes across the plane (`tolerance / size`) and each origin within
the tolerance of the other plane, the tolerance being the record's `tolerance` (SMLib's per-brep
XSectTol3d, the one its booleans test surfaces for coincidence with). The body's size stands
for the plane's extent (at least as large, so no looser); rounding stays the floor. The normal
window the candidates are searched in follows the same angle.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The truck booleans answered a tool clear of the body without intersecting (the operand kept,
its record shared). The SMLib boolean always rebuilt the body, so a cut that could not reach a
body still handed back a new record: a hole far from a second body reported that body changed
(#2904's test), and a body clear of a pocket's cutter was no longer the same solid. A cut or
intersection whose operands are apart (their exact boxes, grown by twice the tolerance as
union_apart takes them, do not meet) now gives the body itself with its faces and edges kept,
or nothing, with no SMLib call.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Two readings of the SMLib record went wrong in emboss_cylinder:
- The footprint arcs were labelled on the sketch's radial frame. The assembly gives each new
circle edge the NURBS form on its label's quarter grid, and the wall face keeps its own form on
its frame's grid (a cylinder's seam lies at an odd angle, build::SEAM_ANGLE). The two grids
differ by a non-quarter turn, so the arc's form is not a straight line in the face's (u, v)
and no exact pcurve exists: "no exact pcurve of a NURBS on a spline surface". The arcs at the
wall's radius are now labelled on the wall face's frame (form::on_frame), so their forms lie on
the face's grid and the pcurve is exact.
- wall_angles took the edge's interval as the circle's angles. On an SMLib record the interval
is on the edge's NURBS form (2πr long for a whole turn), so a wall spanned some 63 rad: a
profile across the seam was accepted, and a wall of radius 0.1 was refused as off its seam.
It now reads Edge::curve_interval.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A face that meets itself along a line (the Pinch case's boss side, turned 30 degrees: the hole's
wall touches it along a vertical line from the hole's bottom to the top rim) has its loop run up
that line and back down. Its motorcycle trace runs from the slit's tip to the bottom and the
iso-line grid meshes it, but the band between the grid cells and the boundary was filled by a
parity flood that flipped at every constraint, the slit's too: the band on one side of the slit
was taken for outside, the face did not follow its boundary, and the lenient fill took over.
The band now crosses a side the loops run along both ways without counting it, as the
triangulation itself already did. Delaunay's sag refinement had the same flood and left the far
side of a slit unrefined (the lenient fill's 0.52 mm^3 shortfall); it takes the slits too.
The Pinch case now meshes strictly: the boss side is 628.3177 mm^2 against 200 pi = 628.3185, and
hole_tangent_inside_a_stepped_boss_cuts_exactly passes. New tess test: a square with a slit
meshed by the cut grid follows the slit both ways and covers the square's area exactly.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>