Commit Graph
3557 Commits
Author SHA1 Message Date
jyvikoandClaude Opus 5.5 7cbbe2bc22 Tessellation: two curves that touch within the chord tolerance share nodes, so the face pinches there
Where two curves of a face touch or cross by less than the chord tolerance, halving their
chords never parts them, and the face was left to the lenient fill: thin_face_tests::
a_face_touching_its_own_edge_meshes (#1648), an arch dipping 2.8e-7 mm below its face's
bottom line, came out 0.659 mm2 against its boundary's 0.567 at tol 0.3.

After the halving rounds, a face whose boundary still crosses itself has each node that lies
within the chord tolerance of a crossing chord of another edge (strictly between its ends) put
onto that edge as well, the same point (edges::snap; the node is marked as snapped). In a face
that runs both edges, the snapped node takes the other's parameters, so the boundary passes one
point of its (u, v) twice (a pinch) and the chords meet there instead of crossing; the edge's
other faces get the node too, within the tolerance of their pcurve, so the mesh stays
watertight. Where the curves run within the tolerance for a stretch, the nodes either side are
shared in turn and the boundary runs out and back along one polyline (a slit). A node is not
snapped where the pinch would turn part of its loop inside out by more than the tolerance (two
curves crossing at an angle: the face is refused, as before). Nothing is loosened: the
crossing test and the triangulation's boundary check are unchanged.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 07:32:44 -04:00
jyvikoandClaude Opus 5.5 9da160892a Tessellation: two corners of a face a hair apart, within their tolerance, are one point of it
A planar boolean leaves a cut-out's corner that touches another loop's corner as two vertices
a hair apart: in the sheet refold tests (doc sheet a_cut_out_on_one_side_of_a_bend_line... and
a_cut_out_with_a_corner_on_a_bend_line...), a triangle's corner 1e-8 to 5e-8 mm from the base's
corner on the bend line, both vertices of tolerance 1.05e-4. Their loops then cross by 1e-8 in
(u, v), which no constrained triangulation can keep, and the face was left to the lenient fill.

The boundary build already makes a vertex the face passes twice one point of its (u, v). Now a
vertex within the tolerance of another of the face's vertices (the larger of the two, at most
the chord tolerance) takes that one's parameters too, so the loops touch at one point, a pinch,
as at a shared vertex. Not for two vertices an edge of the face joins: the edge says they are
two points, however near (a_ledge_beside_an_edge_as_tall_as_the_setback_goes_with_the_strip
has a 1e-4 mm step, within the vertices' 1.18e-4 tolerance), nor at a pole.

Test changed, with the reason: thread_display_mesh_stays_on_its_surfaces checked every triangle
whose three corners lie at the core radius rc as one of the core cylinder, and wanted its centre
within 2 tol of rc. The flat end disc of the M10x1.5 thread (z = 12) has its corners a hair
apart too; joined, it is meshed strictly from its boundary (a plane, e39fbdebb), and one of its
triangles has its three corners on the rim, centre at r = 3.956, on the disc. The end-disc
exemption the test already makes for its radius check now covers this check as well.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 07:32:44 -04:00
jyvikoandClaude Opus 5.5 368b34e3bc Tessellation: a curved face's grid reads the quarter turn from its surface, not its parameter
uv::divide started every curved face from ceil(range / (pi/2)) lines, as if the parameter were
an angle. A spline's is not: SMLib holds a 90 degree bend as one rational quadratic span with
v in [0, pi/2], and the far pcurve of its boundary gives v = -6.7e-13, so the range is a
rounding over a quarter, the count rounds up to 2 and the bend gets a line at 45 degrees
whatever the tolerance (display_mesh_tests::a_small_bend_on_a_large_sheet_is_meshed_round:
3 angles across the bend at the chord tolerance instead of 2).

The division now starts from the whole range, and the limits refuse an interval that spans
more than a quarter turn of its iso-lines, read from the points as an edge segment's is
(edges::QUARTER_TURN: the half chords of an arc turning theta differ by theta/2, a closed
iso-line's by pi). On a revolved surface the parameter is the angle, so the lines are the
same as before; an iso-line within a millionth of the tolerance of one point (a pole) has no
turn to read.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 07:32:44 -04:00
jyvikoandClaude Opus 5.5 1af128a69a Curvature comb: Comb length and Teeth controls in the tool
The comb tool passed only the pick, so a comb's length and tooth count
could not be changed. While the tool runs, a card at the top left of the
view has Comb length (0.25x to 8x the automatic length, default 1x) and
Teeth (4 to 200, default 40), with plain-word tooltips; moving either
runs the comb again on the curves and edges already shown. New strings
in assets/locales/pending/curvature-comb-options.json.

Test: comb_options_draw_the_combs_shown_again drags both sliders: the
shown comb's teeth grow to 8 times their length, then number 4; the
drags pick nothing, and the card goes with the tool.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 07:12:29 -04:00
jyvikoandClaude Opus 5.5 9142c6c5b5 Curvature comb: exact teeth on model edges and faces
`inspect.curvature_comb` read the display mesh for model geometry: an
edge's comb fitted a circle through every three points of the nearest
mesh polyline, a face's did the same on its iso lines' chord points. The
comb was jagged, changed with the mesh tolerance and ignored `density`.

- geom: `CurveEval::jet`, a curve's point, unit tangent and curvature
  vector kN from its first two derivatives (lines and conics in closed
  form, B-splines and NURBS from the basis functions' derivatives, a
  traced intersection from both surfaces' normal curvatures along its
  tangent, by Meusnier: kN lies in the plane of the two normals and its
  component along each is that surface's normal curvature).
- kernel (curvature_comb.rs): the edge nearest a pick by the exact
  distance to its curve (a periodic face's seam and poles left out),
  `density` teeth at even steps of the edge's parameter; along an iso
  line, the surface's normal curvature n.S_tt / |S_t|^2 along its normal.
  `IsoLine` carries the running surface parameter's span.
- engine: edges and faces use them (and `inspect.minimum_radius` on
  edges); new optional `length` multiplies the automatic scale. The reply
  keeps its shape.

Tests: a circle edge reads 1/R to 1e-9 at Coarse and Ultra display
quality alike, with `density` teeth; an extruded ellipse's edge matches
ab/(b^2 cos^2 t + a^2 sin^2 t)^(3/2) at every tooth, a/b^2 and b/a^2 at
the vertices; a cylinder face's rings read 1/R and its lines 0; `scale`
and `length` set the teeth's length; curve jets on every record kind.
The existing cylinder-edge check tightens from 0.5 mm to 1e-9.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 07:12:29 -04:00
jyvikoandClaude Opus 5.5 172c64cef8 Tests: the naming corpus checks that references resolve and builds agree
corpus_designs_rebuild_with_the_same_names_and_references compared every
corpus design with a text record of each body's face and edge counts and its
full name list, last written from truck's output (a55196e7a); 110 designs
differed only because SMLib keeps whole faces truck split in pieces. Those
counts and names are the kernel's, not an outcome.

The test, renamed corpus_designs_keep_their_references_and_name_alike_every_build,
now asserts invariants only, with no record:
- every design loads; one that does not is a failure naming the design and
  the step (oracle/83-boundary-loop-thicken: surface.thicken, SMLib status
  1001), not a record line;
- every name a design stores (fillet and chamfer edges, faces a feature
  picks, a sketch's face, sheet metal edges) finds a face or edge: none
  `gone`, no sketch face `lost`;
- two builds give the same name -> face map: each named face's area and
  centre and each named edge's length and middle (to 1e-6 of their size), and
  the same face or edge found for every stored name.
The design already exempt for unstable piece order
(regressions/closed_sweep_torus_cut) is still only printed for the second.

On develop it fails on two designs: oracle/83 above, and
regressions/cut_up_to_face_leaves_skin, whose Extrude13 face
`F47:top#[...]` finds nothing (the truck-era record had it `gone` too). All
159 designs name alike in both builds. The records in tests/naming/ are no
longer read; CONTRIBUTING no longer asks to re-record them.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 07:12:24 -04:00
jyvikoandClaude Opus 5.5 c1328212d3 Tests: curved bodies' contact stop is checked, not the sweep's internal step
ordinary_curved_bodies_keep_a_practical_sampling_step asserted
ContactGeometry.step > 1.0, an internal sampling step derived from the
bodies' display meshes. What a user gets is the drive: it either stops at the
contact or is refused for running out of the sampling budget. The test,
renamed ordinary_curved_bodies_stop_at_their_contact_within_the_sampling_budget,
keeps that: the slide drives to -30 without a budget error and stops in
(-5.02, -4.9) (the sphere r 5 about x = 20 and the box from x = 15 meet the
block's x = 10 side after 5).

It still fails on develop with "motion exceeds contact sweep collision
budget": the planar faces meshed as grids, which the planar-tessellation fix
(#9307) addresses.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 07:12:24 -04:00
jyvikoandClaude Opus 5.5 51855589d8 Tests: Open in EasyPrint's STL is read back, not measured in bytes
open_in_easyprint_hands_the_host_a_3mf_or_an_stl_that_reads_back asserted the
STL was 84 + 50 * 12 bytes: twelve triangles, the tessellator's count for the
fixture's box (the fixture is a mesh import of box_solid's display mesh, 124
triangles today).

The STL is now parsed (solidcraft_io::read_stl): it has the imported mesh's
own triangle count, every side is shared by two triangles running opposite
ways (closed and consistently wound), it encloses the 10 x 20 x 30 box's
6000 mm^3 (divergence theorem, to 1e-6; the corners are exact in f32) and its
bounds are (0, 0, 0)..(10, 20, 30).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 07:12:24 -04:00
jyvikoandClaude Opus 5.5 2be7b1b6fa Tests: the extruded ellipse is checked by its area, not its face count
free_form_profiles_extrude_exactly asked for an ellipse extrusion of exactly
3 faces and a spline-and-line extrusion of at most 5: how many pieces the
kernel cuts a closed side into. It now checks that the sides are exact.

Ellipse a = 10, b = 4, extruded 5: area = 2 pi a b + 5 P, P = 4a E(e),
e^2 = 1 - b^2/a^2. By the Gauss-Kummer series, P = pi (a + b) sum_k
(binom(1/2, k))^2 h^k, h = ((a - b)/(a + b))^2 = 9/49: P = 46.026225191330
(a 200000-point midpoint integral of sqrt(a^2 sin^2 + b^2 cos^2) agrees to
1e-13). A = 80 pi + 230.131126 = 481.458538 mm^2, asserted to 1e-4; a faceted
side would be short of it. Only the top and bottom are planes
(face_types.plane = 2, merged pieces counted once).

The fit spline closed by a line: the planes are its top, bottom and the line's
side (face_types.plane = 3); its volume stays checked against the profile
area.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 07:12:24 -04:00
jyvikoandClaude Opus 5.5 978ae1d512 Tests: a mirrored sheet's flat pattern is its source's mirrored, as the code makes it
a_mirrored_or_patterned_sheet_body_is_sheet_metal_with_its_flat_pattern
(1171c4c2a) expected the mirrored copy's model turned over: the same outline,
its bend down and its pocket on top. That describes a SheetBody::moved that
never landed. The code on develop is SheetBody::copied (d9d6af51c, #3354): a
mirror moves the frame and mirrors the flat x -> -x with the same thickness
side, the opposite hand's blank with each bend turning the same way, which
a_mirrored_sheet_copy_is_the_opposite_hand_with_the_mirror_image_flat_pattern
asserts on an asymmetric bracket. The YZ mirror keeps z, so a pocket cut from
below stays below and the flange on y = 0 still turns up: the flat pattern
reads blank x 0..100 -> -100..0, pocket x 20..30 -> -30..-20, side bottom,
bend up.

The test now asserts that: the same flat size, the outline and the pocket
outline equal to the source's with x negated (to 1e-6), pocket side bottom and
bend up on both. The volume doubling, the bend table and the flange rebuilt
on the copy (point containment) stay.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 07:12:24 -04:00
jyvikoandClaude Opus 5.5 5ea43b5c29 Tests: a tray's sheet names are checked by where their points lie
a_trays_faces_have_readable_sheet_names_that_follow_its_size asked for
exactly two unnamed faces (the hole's wall in two halves) and for the back
rim's named point at y = 199.5, the strip's centre line. The first is how many
faces the kernel makes a cylinder of; the second is where the face's roomiest
display triangle lies (naming::face_spot), any point on the face.

Now: there is an unnamed face and every one is the hole's wall (r 3.3), and
the back flange's far rim point lies on that rim: z = 30 and 199 <= y <= 200
(the flange 1 thick, its outside at y = 200), 0 <= x <= 400, to the 1e-3 the
test had. The right flange's top (the plane x = w - 1) keeps x = 319 and 399.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 07:12:24 -04:00
jyvikoandClaude Opus 5.5 5e7e1485c0 Tests: the knife-edge R2 fillet is checked once, in closed form
fillet_the_knife_edge_at_the_tip (features_more_tests.rs) and
blend_a_knife_edge_where_a_curved_wall_runs_out (tests/robustness.rs) build
the same #851 part with the same script, round the same knife edge at
(62 1/6, 0, 5) with R2 and measure the removed volume through the same display
plan. The first integrated the removed area on a grid (5e-4); the second has
it in closed form. Drop the first and keep its tolerance on the second.

The ball 2 from the back face y = 0 and from the r 21 wall about (62 1/6, -21)
is centred at (62 1/6 - sqrt(168), -2). The removed section is the trapezoid
wall centre, tip, (kx, 0), ball centre (sqrt(168) wide, 21 and 2 high) less the
wall's sector theta = acos(19/23) and the ball's sector pi - theta:
  10 (11.5 sqrt(168) - 220.5 theta - 2 (pi - theta)) = 119.648812 mm^3
the constant the kept test has; the dropped test's grid gives 119.6487 at a
0.004 x 0.001 step. The kept test's fillet bound goes from 0.5 to 5e-4 of
that (0.0598), the dropped test's. Its chamfer row (159.236451, the chord
triangle less the wall's segment) already says the chamfer takes more.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 07:12:24 -04:00
jyvikoandClaude Opus 5.5 03054211b3 Tests: the arc-touching cut is checked by its volume and area, not its face count
cut_whose_end_touches_an_arc_side_leaves_no_sliver compared the touching cut's
face count with a clear cut's, a count of how the kernel splits its faces. It
now compares what the two parts are.

The plate's arc runs from (0, 20) through (-10, 10) to (0, 0): centre (0, 10),
r 10, so it reaches x = -10 only at y = 10. The r 5 cut along x about
(y, z) = (12.5, 12.5) spans x -10..7.5 (touching) or -10.1..7.4 (clear). Both
open the same patch of the arc; the slice x 7.4..7.5 lies wholly inside the
plate (y 7.5..17.5, z 7.5..17.5 in 0..20 x 0..27.5). So
  V(clear) - V(touching) = pi 5^2 0.1 = 7.853982 mm^3
  A(touching) - A(clear) = 2 pi 5 0.1 = 3.141593 mm^2 (the hole wall; the
  floor disc is the same)
both to 1e-4 (the measure integrates the exact faces to 1e-8). A sliver of the
arc left where the end touches adds about a square millimetre of faces.

Seed 14's half (the fuzz generator's seed, then a shell) is pinned as its own
test, shell_after_a_cut_whose_end_touches_an_arc_side, with the commands the
generator gives today: the touching cut, then a 2 mm shell opening the top
face y = 20. It fails at the shell (SMLib status 1001), a kernel gap.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 07:12:24 -04:00
jyvikoandClaude Opus 5.5 21c1ed541e Kernel tests: rim rounds meeting at a concave corner are mitred, not swept
The owner's final ruling of 2026-10-30 reverses the swept corner 6539c070a
and 8dc344d22 used: rim rounds meeting at a concave corner beside an
unpicked sharp edge are mitred by default (a user who wants it swept rounds
the corner edge too; a swept corner with a sharp inner edge is a future
option). SMLib's swept result is the defect, so both tests now expect the
mitre and fail today on it.

rounds_that_fit_are_closed. At depth u the section is the outline moved in
by d = r - sqrt(r^2 - (r - u)^2) with every corner mitred: the area less
L d, plus d^2 at each of the six convex corners (the two strips overlap in
a d x d square) and less d^2 at each of the six concave ones (the mitre
also takes the d x d square beyond both strips), the overlap of the two
spandrel prisms being r^3 (5/3 - pi/2) per corner over the round. Six and
six cancel, so both ends remove 2 L r^2 (1 - pi/4) with L = 107.6:
11.545566 (the swept corners gave 11.514718).

fillet_tabbed_plate_outer_loop_and_partial_selection_after_hole. Where both
sides of a concave corner (+-4, sqrt(609)) move, the moved R25 arc and the
moved side x = 4 - d now meet where they cross,
y = sqrt((25 - d)^2 - (4 - d)^2), as where only one side moves; the swept
arc branch goes. Check: A(0,0) - A(d,d) agrees with the Steiner form for a
mitred outline, L d - d^2 (pi + theta - 2 tan(theta/2)), to second order
(17.190976 against 17.190968 at d = 0.1; the arc's curvature moves the
mitre point at third order, which the Green's-theorem area has exactly).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 07:12:24 -04:00
jyvikoandClaude Opus 5.5 4d7553dd8b Kernel tests: a ruled strip's moved handles are checked on its far curve
a_strip_off_a_curved_face_is_ruled_to_its_far_curve lifted the free
strip's handles 1 mm and wanted the strip's bounding box to rise, which it
need not (the far curve's corners are fixed and the box's top may be
elsewhere). It now checks what the handles do: the plan's far curve takes
the lifted handles as its interior control points and keeps its corners,
so it rises by 1 - R0 - Rn inside; the built strip has that curve for an
edge (33 points of it within 1e-3 of the strip's edge polylines at 1e-4),
and its middle, risen more than 0.1, is more than half that from every
edge of the default strip.

It fails today on its premise: the strip's edge (the cup's straight rim at
x = 14) is a line, so its far curve, on the edge's degree and knots, has
no interior control points, and ruled_far_handles gives an empty list
(the old witness passed nothing to move). A real gap: the far curve of a
strip whose rulings turn along a straight edge is the chord between its
corners.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 07:12:24 -04:00
jyvikoandClaude Opus 5.5 a45d56aa94 Kernel tests: the slot plates' STEP files are graded by their bounds and volume, not surface kinds
rounded_slot_exports_cylinders required no B_SPLINE_SURFACE, four
CYLINDRICAL and two TOROIDAL surfaces; ramp_cut_slot_exports_sheets
_bounded_by_their_faces five CYLINDRICAL and two B_SPLINE (SMLib keeps one
more round an exact cylinder). Which kind a face is written as is the
writer's choice; what #612 and #1198 were about is that no surface reaches
past the part. Both keep that (every CARTESIAN_POINT's x within [0, 60],
control points and placements included) and their round-trip volume: the
first against Pappus, the plate (800 + 100 pi) 4 less (1 - pi/4) along
the straight edges and the arcs at 10 - r (10 - 3 pi) / (3 (4 - pi)), at
1e-3; the second against the exported body's at 1e-4. Both pass today.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 07:12:24 -04:00
jyvikoandClaude Opus 5.5 a3848d1d0c Kernel tests: modifier histories are graded by what each face is, not how many changed
Three history tests pinned truck's sewing in tallies:
- shell_of_a_painted_body wanted changes [0, 2, 8, 0, 5], the side faces
  split across the rim's level; SMLib's shell leaves them whole. It now
  checks what paint follows: the painted face 0 comes through as exactly
  one face (its geometry checked against face 0 by verified()), and each of
  the five inner walls is one Generated Offset of the face it lines, the
  open top none, as shell_lines_the_walls_with_offsets has it.
- fillet_after_box_cuts wanted the painted cuts to leave faces in pieces
  and the fillet's edges to meet merged faces; it keeps the outcome for
  every cut on both blocks: the round stands in place of the picked edge,
  and composed with the cut the history holds over the block and the tool.
  merged_meetings has no use left and goes.
- surfaces_keep_their_faces wanted every stitched face Kept; whether
  sewing changes a face's class is not the contract. It now checks each
  stitched face has one origin, the face it was sewn from (geometry checked
  by checked()), the five walls the open box's faces 0-4 and the lid the
  patch, and the four seams naming two edges each.

split_a_face is left as it is (the owner's ruling: a face whose boundary
only gained a vertex is Kept), failing until SolidCraft's split history
says so ([3, 2, 2, 0, 0] against [5, 0, 2, 0, 0]). The three tests above
pass today.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 07:12:24 -04:00
jyvikoandClaude Opus 5.5 4ff38619c1 Kernel tests: the rounded shell's planar roles are graded by name and exact volume
rounded_shell_planar_roles_are_exact_distinct_and_transformable required
the four outer r6 rounds to be coalesced extrude sides (face_was_coalesced:
SMLib's round is not one any more), no floor-height seam between corner
faces, 19 faces in the cup and in its STEP copy, and the volume of a 1e-4
mesh within 0.2. It keeps what the roles are for: the right wall's role
resolves to the face at x = w/2 and the floor's to the face at z = t, both
still after the transform; each straight outer wall is one full-height
face, in the cup and in its STEP copy; and the cup's volume is exact.

Proof. The box w x 60 x 30 with its four vertical edges rounded r6 is
(60 w - (4 - pi) 36) 30; shelled t inward, open at the top, its inside is
the box moved in by t with its corner rounds at 6 - t:
((w - 2t)(60 - 2t) - (4 - pi)(6 - t)^2)(30 - t). Graded on SMLib's
volume (integrated on the surfaces, to 1e-8) at 1e-6. It passes today.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 07:12:24 -04:00
jyvikoandClaude Opus 5.5 5229bbedd4 Kernel tests: a small bend's display mesh is graded on the display mesh alone
a_small_bend_on_a_large_sheet_is_meshed_round also pinned the chord-only
mesh (exactly two stations across each bend, its sag over 0.29 r) and
required every flat side face to keep the chord-only mesh's triangle count:
facts about another mesh, not what the display shows (the chord-only mesh
now puts three stations across the r1 bend). It keeps what the display
mesh must do: each bend a quarter turn in more than five facets, each
turning at most DISPLAY_ANGLE, sagging at most r (1 - cos(angle / 2)) and
less than the chord tolerance. It passes today.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 07:12:24 -04:00
jyvikoandClaude Opus 5.5 65718cd418 Kernel tests: the five- and six-sided fills are graded by what they make, not their face counts
a_five_sided_window_fills_g1_and_through_a_point required five faces and at
least 15 edges (five patches about a centre); a_hexagonal_rim_fills_through
_a_point... required (6, 18). How many patches a fill is made of is the
fill's own business: both keep their outcome checks (check_face: the rim's
own edges, within the face tolerance, a clean mesh; G1 to the wall within
0.1 degree; the point on the surface; the sealed cup holding more than the
plain fill's).

Both still fail on outcomes: the pentagon's fill through a point has a
folded triangle in its mesh, and SMLib's stitch refuses the hexagon's fill
through a point (its control net doubles back on itself, a cusp in a
curve).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 07:12:24 -04:00
jyvikoandClaude Opus 5.5 87e9123d3a Kernel tests: a crowned patch's hostile crowns are refused for their range
a_crowned_patch_fills_its_middle_and_keeps_its_edges checked only that NaN,
infinity, 3.01, -0.91 and 1e300 are refused; it now checks why: "a patch's
crown must be between -0.9 and 3" (patch_edges_crowned's range test).

The bump check is left as it is. The audit proposed measuring it in model
space along the loop plane's normal, but it already is: both patches are
height fields over the same plane on the same parameters, so the distance
between their points at the domain's centre is the bump along the normal,
and the record's domain is the fit's box in mm (diagonal 52.5). The crowned
patch's middle rises 7.656 where the documented crown x size is 13.125:
the crown's bump reaches 7/12 of what patch_edges_crowned promises on this
pentagonal cup (geom's own test of fill_loop, on a fine hexagon net, gets
it to 1e-9), a real gap, still failing.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 07:12:24 -04:00
jyvikoandClaude Opus 5.5 c69dbc5d39 Kernel tests: the collapsed taper's mesh is held to its B-rep, the taper to validity
collapsed_taper_mesh_stays_accurate_as_tolerance_decreases compared each
display mesh's volume with the taper law's integral,
6000 - (1 - pi/4) 40 integral r^2, at 1e-4: the built fillet removes about
10 mm3 more than that (5212.9 against 5223.1) at every tolerance, so it
measured the fillet, not the mesher. It now holds each mesh to the B-rep it
meshes: a mesh within its chord tolerance of the faces has its volume
within area x tolerance of the exact one (SMLib's, integrated on the
surfaces to 1e-8), the bound fillet_mixed_arc_and_line_removed_volume
_converges uses. It passes today.

taper_follows_the_room_along_the_edge keeps its law and closed-form
volumes (the 10 mm3 against the law is the failure it still shows) and
drops the two internal checks at its end (no unmeshed face, a face with
three vertices) for the outcome: the wedge's tapered fillet is valid.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 07:12:24 -04:00
jyvikoandClaude Opus 5.5 a0c24b87c0 Kernel tests: the back edge then the arc is one more pair of the rounds-in-sequence test
the_arc_rounded_after_the_back_edge_matches_both_together ran
one_then_the_other(BACK, arc_top()) as a test of its own. It is the same
check as every pair of a_round_ending_in_an_earlier_round_matches_both
_together (the round in sequence against both together, within the
display meshes' 5e-3), which had only the other order (arc_top(), BACK):
the pair joins that list and the separate test goes. Nothing is loosened.

Today the list fails before it reaches the pair: SMLib fails the arc's
round after the long front edge's (status 1001); the back-then-arc pair
failed the same way on its own.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 07:12:24 -04:00
jyvikoandClaude Opus 5.5 659bd866b3 Kernel tests: the #871 part's fitting rounds take their exact volume
rounds_that_fit_are_closed counted faces (the input's plus one per picked
edge: 38, where SMLib makes 50 with its corner faces). It now keeps
validity and no crossed loops and holds the R0.5 rounds on all 24 end-face
edges to the exact volume.

Proof. The part (plate 14 x 21.5 x 10, block 14 x 9.948 x 10 on it, the
5.2 x 7.1 window through both along x) is
14 (215 + 10 (b - a) - 36.92). At depth u into an end face's rounds the
section is that face's outline moved in by d = r - sqrt(r^2 - (r - u)^2),
each edge parallel to itself: the area less L d, L = 2 (21.5 + 20) +
2 (5.2 + 7.1) = 107.6 (the stepped outline's perimeter is its box's), plus
d^2 at each of the six convex corners (two rounds mitred), less pi d^2 / 4
at each of the six concave ones (the two shoulders and the window's four
corners: rounds meeting at a concave corner beside an unpicked sharp edge
are swept round it, the owner's ruling of 2026-10-30). The wall between
window and block side is 1.256 > 2 r, so the offsets stay apart. Over the
round the integrals of d and d^2 are r^2 (1 - pi/4) and r^3 (5/3 - pi/2),
so both ends remove
2 (L r^2 (1 - pi/4) - (6 - 6 pi/4) r^3 (5/3 - pi/2)) = 11.514718.
Graded at 1e-7 of the part's volume. It passes today (aarch64, SMLib).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 07:12:24 -04:00
jyvikoandClaude Opus 5.5 5af467aef0 Kernel tests: the box rounded after separate bottom rounds takes its exact volume
issue2422_vertical_after_separate_bottom_rounds compared the face count and
volume (within 0.05) of the vertical round, picked with and without its
bottom mitre, against the same round picked alone, and required every face
meshed. It now holds each of the three picks, on the 40 x 80 x 27 box and
on the same shelled 1 mm open at the top, to the exact volume; the corner
probes stay.

Proof. Horizontal sections, s up from the bottom. Below s = r the four
separate R2 bottom rounds (mitred at the corners) leave the rectangle
moved in by d = r - rho, rho = sqrt(r^2 - (r - s)^2), and where the
vertical round meets them (a sphere octant) its corner is a quarter disc of
radius rho: (L - 2d)(W - 2d) - (1 - pi/4) rho^2. Above, LW - (1 - pi/4) r^2.
With the integrals over the bottom r of rho (pi r^2 / 4), rho^2 (2 r^3 / 3),
d (r^2 (1 - pi/4)) and d^2 (r^3 (5/3 - pi/2)):
V = LW r - 2 (L + W) r^2 (1 - pi/4) + 4 r^3 (5/3 - pi/2)
  - (1 - pi/4) 2 r^3 / 3 + (H - r)(LW - (1 - pi/4) r^2) = 86174.445361;
with four such corners it is issue1557_all_vertical_corners_match_analytic
_volume's formula exactly. The shell's cavity, 38 x 78 x 26, is clear of
both rounds (along the corner's diagonal the R2 arc stays 2 - sqrt(2) in,
the cavity's corner is 1 in), so the cup is V - 77064. Graded at 1e-6.

Today SMLib's check refuses the bottom rounds' result (edge and vertex
tolerance 0.97 against 0.01), as before.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 07:12:24 -04:00
jyvikoandClaude Opus 5.5 a8f393ca36 Kernel tests: the tabbed plate's rim rounds take their exact volume
fillet_tabbed_plate_outer_loop_and_partial_selection_after_hole only
checked that every subset of the R25 tabbed plate's seven top-outline edges
builds and stays inside the plate. It now first holds the hole's rim round
and the three selections that end only at the plate's sharp concave
corners (the R25 arc's two halves, the tab's tangent run of sides, R2
corners and top, and both) to their exact removed volume; the sweep over
all 127 subsets stays as it was.

Proof. The outline's area A(da, dt), the arc moved in by da and the tab's
run by dt, each element parallel to itself, by Green's theorem: an arc
about c of radius rho from a to b gives
1/2 (rho^2 (b - a) + rho (cx (sin b - sin a) - cy (cos b - cos a))), a
segment p -> q gives 1/2 p x q. At a concave corner (+-4, sqrt(609)) where
only one side moves, the elements meet where they cross (the round runs
into the other wall: the section is the union of the two sides' half
regions); where both move the corner is swept, an arc of radius d about the
corner, traversed clockwise from the arc's normal to the side's (the
owner's ruling of 2026-10-30 for rounds meeting at a concave corner beside
an unpicked sharp edge: the profile swept round the corner edge). Checks:
A(0, 0) = 312.5 (pi + 2 theta) + 4 y + 8 (cy - y) + 8 + 2 pi (disc sector,
triangle, tab, R2 corners), and A(0, 0) - A(d, d) = L d - pi d^2 by the
Steiner formula for the swept offset (total smooth turning 2 pi + 2 theta,
the two swept corners -theta each), agreeing to 1e-15 at d = 0.1, 0.3,
0.5; A(0, 0) - A(0, d) agrees with a 0.004 mm grid to 2e-6. The round's
section at height r sin(phi) into it is the outline moved in by
d = r (1 - cos phi), so it removes the integral of
(A(0, 0) - A(d)) r cos(phi) dphi over [0, pi/2] (midpoint rule, 8192
steps, converged to 1e-7 against 4096). The hole's rim round is Pappus:
(1 - pi/4) r^2 about the axis at 1.5 + r (10 - 3 pi) / (3 (4 - pi)).
Graded at 1e-5 of the removal, the bound the D prism's rounds are held to
(the blend surfaces are SMLib's fits).

Today the three selections build with these volumes (aarch64, SMLib); the
sweep still fails where a subset ends a round at a tangent joint (the x = 4
side without its R2 corner): SMLib cannot end a blend inside a tangent
chain, "the blend cannot end there".

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 07:12:24 -04:00
jyvikoandClaude Opus 5.5 3569298a40 Kernel tests: a chamfer whose leg runs off the thin rim is refused, naming the largest
The owner's blend-overflow ruling of 2026-10-30: a chamfer whose leg runs
off the face beside its edge over another face is refused as too large,
naming the largest size that fits (a 2 mm chamfer on the basin's 1 mm rim
would lower the slab's front face from 30 to 29.382); a full-width chamfer
stays refused. Two tests wanted those chamfers built and cut from the slab:

- chamfer_wider_than_sloped_basin_rim_is_cut_and_a_full_width_one_refused
  becomes chamfer_wider_than_sloped_basin_rim_is_refused_naming_the_largest:
  (margin, d) = (1, 2), (1, 3), (0.5, 4) and the full widths (0.5, 0.5),
  (1, 1) are TooLarge, the body unchanged, and the message names a largest
  size just below the rim's width: the full width is refused, so the fit
  search (three significant digits of the size asked, step
  10^(floor(log10 d) - 3)) ends one step below it, in
  [margin - step, margin). The removal formula for the cut chamfer
  (basin_rim_chamfer_removed) has no use left and goes.
- equal_slope_rim_loops_stay_inside_the_material: the 2 mm chamfer of the
  1 mm rim's loop is refused the same way (largest in [0.999, 1)); the R4
  round (tangent side, builds) and the 0.5 blends stay.

The half-width chamfer on each rim, which only had to take something off,
now takes its exact volume. As in chamfer_sloped_basin_rim: the triangle
with legs d on the top and d down the wall towards (20, 10) (run
20 - margin over the 20 drop) has area d * drop / 2 with
drop = 20 d / sqrt((20 - margin)^2 + 400), and the chamfer's length at
depth h is 60 - h between the end walls x = 20 + (30 - z) / 2 (planar:
x depends on z only), so it removes d * drop / 2 * (60 - drop / 3);
graded at 1e-8 as that test is.

Today SMLib refuses the chamfers as too large, but the message does not
name the largest that fits yet ("the chamfer is larger than the faces
beside its edges allow"), the SolidCraft change issue_1080's R30 test
waits for too.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 07:12:24 -04:00
jyvikoandClaude Opus 5.5 65e46a9628 Kernel tests: a round as tall as its wall and more builds on the disc
The owner's blend-overflow ruling of 2026-10-30: a round that runs past
its face on its tangent side is built on the face continued and cut from
the part, keeping its radius. loop_fillet_as_tall_as_its_wall_consumes_it
still wanted R2.001 and R2.5 on the R13 x 2 disc refused as taller than the
wall; they now must build, valid, with the exact volume.

Proof. The round's section in (rho, z) is the r square at the top edge
outside the circle about (13 - r, 2 - r), clipped to z >= 0. With
t = z - (2 - r) = r sin(theta), the round's inner radius is
rho(theta) = 13 - r + r cos(theta), theta from asin((r - 2) / r) to pi/2,
and the disc keeps pi (169 * 2 - integral (169 - rho^2) dz), dz =
r cos(theta) dtheta (midpoint rule, 20000 steps). At r = 2 it is Pappus,
2 pi 169 - (1 - pi/4) 4 * 2 pi (13 - c), c = 2 (10 - 3 pi) / (3 (4 - pi))
(checked to 1e-9 in the test); at r = 2.5 it is 957.692239 mm3, the
ruling's number, the same integral as
rounds_past_the_bottom_of_a_wall_are_cut_from_the_part's. Graded at 1e-6
as that test is.

Today SMLib fails the R2.001 round (status 1001): a convex round past its
face is not yet built by subtraction, the cause fillets/KNOWN_FAILING gives
4782 and 5158.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 07:12:24 -04:00
jyvikoandClaude Opus 5.5 e9963439b1 Kernel tests: the round into a hole moves to the rig, with the hole cut the test's way
a_round_into_a_hole_is_cut_by_it (issue_324_tests.rs) built rig case
fillets/324-fillet-over-hole-overlap's body, but cut its hole with a cylinder
longer than the block (z -1 to 11) where the rig extrudes a sketch circle
flush. The new case fillets/324-fillet-over-hole-cut-by-overlong-cylinder
builds it that way (solid.cylinder, base z -1, height 12, cut) and holds it to
the same hand-derived answer: block 40x20x10 less the r3 hole at (20, 14.5)
(282.74) less the R4 round on the top edge at y = 20, 40 (16 - 4 pi) = 137.35,
plus what both remove, the integral over y in [16, 17.5] of
(4 - sqrt(16 - (y - 16)^2)) * 2 sqrt(9 - (y - 14.5)^2) = 0.375830 (re-derived
with the substitution y = 17.5 - s^2, 200000 steps):
V = 8000 - 90 pi - 40 (16 - 4 pi) + 0.375830 = 7580.287315, and G1 on the
blend. It passes today (aarch64, SMLib), as the flush case does, so it needs no
KNOWN_FAILING line; the unit test, a duplicate of the two cases, is dropped.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 07:12:24 -04:00
jyvikoandClaude Opus 5.5 51641123cf Kernel tests: drop the mixed arc-and-line mitre test, the rig's 485 grades it
fillet_mixed_arc_and_line_joint_preserves_both_section_radii put R0.5 on the
D prism's top outline and checked 258 points of each mitre edge against the
cylinder and torus to 1e-7 in r squared, through the record's edges. The rig's
fillets/485-fillet-d-prism-top-outline-arc-and-flat builds the same prism with
the same radius and grades the removed volume to 5e-5 (a 0.12 mm3 loss shows)
and G1 on every blend seam, which is what the mitre's accuracy is for.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 07:12:24 -04:00
jyvikoandClaude Opus 5.5 2f88ca2f62 Kernel rig: everyday chamfers of box edges and pocket rims
A box edge chamfered 2 and by two distances (3 and 1, and flipped, told apart by the
centre of mass); a pocket's rim chamfered 1, by two distances, 4.9 (fits), and 5 and
5.5 (refused by the owner's rule on full-width chamfers: built at 5 today).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 07:07:31 -04:00
jyvikoandClaude Opus 5.5 fd571d076a Kernel rig: everyday fillets after fillets
A box edge rounded, then the edge meeting it at the corner, in both orders (both the
three-edge rolling-ball corner); r3 then r2 (the r2 round's torus about the r3 round's
axis, by Pappus); an L-bracket's front inner outline after its bend round; a block's
top edge after the arc round at the cylinder (fails today: G0 at the round's end on the
crease); a curvature round after a curvature round with Tangent chain off.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 07:07:31 -04:00
jyvikoandClaude Opus 5.5 06e802ffee Kernel rig: everyday rounds of an L-bracket's bend
Inside bend r3, outside r5, both concentric (r3 and r8), the inside r35 using up both
inside faces, r36 refused naming 35, and the inside bend Type: Curvature.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 07:07:31 -04:00
jyvikoandClaude Opus 5.5 cd7a099f42 Kernel rig: everyday rounds and chamfers of a hole through a plate
Top and both rims rounded r1 and chamfered 1; the top rim rounded r10 (as deep as the
plate), r12 (past it on its tangent side: built by subtraction, fails today with
status 1001) and Type: Curvature; chamfered 9.9 (fits) and 10 (the full depth:
refused by the owner's rule, built today). Volumes by Pappus.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 07:07:31 -04:00
jyvikoandClaude Opus 5.5 d027399e0c Kernel rig: everyday rounds of a pocket and a boss
A rectangular pocket's floor edges, rim edges and both rounded r2, the rim r5 as deep
as the pocket; a boss's root r2 and r10 (as tall as the boss) and its top rim r2.
The rim cases fail today: SMLib puts a corner patch where two rim rounds meet at a
concave pocket corner, where the rounds running into each other make a mitre (four
faces more; the volumes agree to 5e-4).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 07:07:31 -04:00
jyvikoandClaude Opus 5.5 f78e7fd66d Kernel rig: everyday rounds on a block joined to a cylinder
A block whose sides meet the cylinder at creases: the concave arc at the wall above
the block's top r2, the cylinder's top rim r2, the block's long top edge r2 with
Tangent chain off and Type: Curvature. A block as wide as the cylinder, its sides
running tangent into the wall: the arc and the long top edge each rounded alone
(built or refused as continuing tangentially). Volumes by Pappus or one-dimensional
quadratures of the exact sections, the round ended by the planes it runs into.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 07:07:31 -04:00
jyvikoandClaude Opus 5.5 57090ba2e3 Kernel rig: everyday box corners and box edges, rounded and chamfered
Three edges at one box corner rounded r1 and r3 with the rolling-ball corner, r3 with
Corners: Curvature at its default setback and at 3.5, and refused at 3 (the radius);
the corner chamfered 2; all twelve edges rounded r2; one edge rounded Type: Curvature.
Volumes in closed form (strips plus sphere octants, half-space intersection, the
rounded box as a Minkowski sum, the quintic section's area 113/756 s^2); G1 at every
round's seam, G2 for the curvature round and the corner patches.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 07:07:31 -04:00
jyvikoandClaude Opus 5.5 b66821c829 Kernel tests: the third round at a mitred corner takes its tangent chain, a sphere octant
Stale since 84bd2daad (the Fillet feature rounds each pick's tangent chain). After a and b are
rounded at 2 mm and meet in a mitre, the mitre's seam leaves c's end along c, so c's chain is c
and the seam; rounded at 2 mm, as the issue1557 tests pick, the corner is a sphere octant.

Proof: the 20 mm cube less its three rounds, each (1 − π/4) r² over its 18 mm straight part, and
the corner's 2 mm cube less the octant, r³ (1 − π/6): 8000 − 3·(1 − π/4)·4·18 − 8·(1 − π/6) =
7949.834793498635. Measured 7949.834793513291 (1.8e-12 relative), 10 faces, valid; the check is
tightened from 0.05 mm³ to 1e-11 relative.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 07:07:27 -04:00
jyvikoandClaude Opus 5.5 42d4d0ecde Engine tests: rounding the long edge then the left upright matches the three edges their chains take
Stale since 84bd2daad, which gave solid.fillet each pick's tangent chain by default. After the
long top front edge is rounded, the front-left upright meets the round's end (the quarter circle
on the left face) along its tangent, and that circle meets the top-left edge along its tangent:
the second fillet rounds the upright and the top-left edge. The "together" fillet picks the
three edges ([20, -21, 10], [0, -21, 5], [0, -10, 10]); one by one and together remove the same
volume within the display meshes' 2e-4 (it read 60.649 vs 43.185 with two picks).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 07:07:27 -04:00
jyvikoandClaude Opus 5.5 96b248d288 Kernel tests: a slot narrower than the taper closes partway up and builds; a closing hole is refused
tangency-labs/smlib#64 builds a tapered wall that closes partway up (its neighbours meet above
its apex), so the 0.2 mm slot of a 10 mm square tapered 10° out over 2 mm now builds where it
was refused. The test, renamed for what it checks, asserts the slot's solid is valid with its
exact volume and keeps the ring's refusal (its 0.3 mm hole closes entirely).

Proof of the volume, t = tan 10°: every wall moves out of the material by z·t at height z. The
square grows to 10 + 2zt, so without the slot ∫₀² (10 + 2zt)² dz = ((10 + 4t)³ − 1000)/(6t).
The slot narrows to 0.2 − 2zt and keeps its length 8 (its end wall and the square's side both
move up by zt); it closes at z = 0.1/t = 0.567 and takes ∫ 8 (0.2 − 2zt) dz over 0..0.1/t =
0.08/t. Volume ((10 + 4t)³ − 1000)/(6t) − 0.08/t = 213.984095421783; measured to 1e-6.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 07:07:27 -04:00
jyvikoandClaude Opus 5.5 01d0d6bb93 SMLib bindings: a curve list holds up to 16384 curves, as the C API now takes
tangency-labs/smlib#63 raised SMLIB_CURVE_MAX_CURVES from 4096 to 16384; CURVE_MAX_CURVES
follows it, so a profile of 5000 segments (a modelled thread, a fine polygon) extrudes:
kernel tests::measure_of_a_body_with_many_faces passes (5002 faces, 15000 edges, 10000
vertices).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 07:07:27 -04:00
jyvikoandClaude Opus 5.5 1661fe292e SMLib: pin 985faae, the batch of 2026-10-30 evening (tangency-labs/smlib#63–#65)
- #63: a knife edge (two faces tangent along it, folding back) rounds to the ball touching both
  faces back from it; a curve list takes up to 16384 curves (SMLIB_CURVE_MAX_CURVES); a stitched
  solid is a nested solid, its cavities void; booleans pair each face only with the faces whose
  box comes near its own, the closure test of a new face searches from both its sides, and a new
  edge from a faceuse shell to a wire keeps the faceuse shell: a row of Menger sponges joins in
  near linear work.
- #64: a tapered wall that closes partway up ends at its apex, its neighbours meeting above (it
  was refused as SMLIB_ERR_WALL_CLOSES); a round of revolution offset in past its radius has no
  offset; a round along an arc turns from its start section, not from the arc's frame.
- #65: capi tests, the Menger row join counts its work (smlib_work_counters) instead of timing it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 07:07:27 -04:00
jyvikoandClaude Opus 5.5 2e123eb3dc Kernel tests: a swallowed tip's volume is the drafted wedge's, every wall at the taper
tapered_extrude_swallows_a_short_tip expected the prismatoid whose two tip
corners run straight to the top's tip (h/6 (A0 + 4 A_mid + A1)), the solid
the old offset-and-rule construction built: its tip triangle leans about 40
degrees, not the 5 degree taper. With every wall drafted 5 degrees the tip's
wall is a triangle that closes where the long sides meet across it, at
z* = 0.126237833 (offset 0.011044379), and above that the long sides meet
along a ridge. The section at height z is the profile's lines moved in by
z tan(a): a quadrilateral below z*, a triangle above, each area quadratic in
z, so Simpson's rule on each piece is exact: V = 72.883284979193. The old
value, 72.882657253174, is 8.6e-6 off, past the test's 1e-6. SMLib's
DRAFT_CORNER_ROUND path builds exactly this solid (72.883284979193, 6 faces);
the default curvature rule still refuses it (SMLIB_ERR_WALL_CLOSES at
(20, -0.01, 0)), so the test fails until the fork builds a wall that closes
partway up.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 07:07:27 -04:00
jyvikoandClaude Opus 5.5 1817cffaf3 Shell: a basin shelled past its 4 mm rim measures its exact sharp offset
shell_thicker_than_a_rim_is_refused expected the "rim or ledge" refusal from 9508ee749,
added because truck's face offset turned the top face inside out and returned a body of
about twice the volume. SMLib builds the exact sharp-edged offset, and its volumes are the
closed form, so the test now checks those (renamed
shell_of_a_basin_past_its_rim_keeps_the_rim_solid).

Proof. Slab [0,100]×[0,60]×[0,30]; pocket lofted from x 20..80, y 4..50 at z 30 to
x 30..70, y 20..40 at z 10, its sides the planes x + z/2 = 35, x − z/2 = 65,
y + 0.8z = 28, y − z/2 = 35 (normal lengths √5/2, √5/2, √1.64, √5/2). Shelled open at
z = 0 by t, the cavity is the slab shrunk by t (x t..100−t, y t..60−t, z 0..30−t) less the
pocket grown by t (each plane moved out by t·|n|, floor z = 10 − t). For t ≤ 6 the grown
pocket stays inside the shrunk slab in x and at the back (50 + 0.618t ≤ 60 − t); its front
reaches y = t at z0 = (28 − (1 + √1.64)t)/0.8, below 30 − t from t ≈ 2.70 (the rim stays
solid above z0). With W = 30 + z + √5 t,
  grown(t) = ∫[10−t, z0] W (7 + 1.3z + (√5/2 + √1.64)t) dz + ∫[z0, 30−t] W (35 + z/2 + (√5/2 − 1)t) dz,
  V(t) = 180000 − grown(0) − (100 − 2t)(60 − 2t)(30 − t) + grown(t),
grown(0) = 33866.667, so V(1) = 17105.6240415, V(4) = 63560.2492284,
V(6) = 90113.9202101. SMLib measures 17105.6242, 63560.2492, 90113.9203 (≤ 1e-8
relative). The test computes the closed form and checks each shell to 1e-6, with the
centroid inside its box and on x = 50 as before.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 05:56:11 -04:00
jyvikoandClaude Opus 5.5 785b78bea5 Surface Fair test: fair the B-spline patch on a saddle rim, not an offset sphere
The test offset the sphere's top patch to get a fitted B-spline sheet. That held on truck,
whose sphere came in several patches and whose sphere offset was fitted. On SMLib the
sphere is one face: offsetting it closes up, so it is a solid by design (surface_offset.rs,
"Faces that close up make a solid", and its test a_sphere_is_fitted_and_refused_past_its_centre
asserts exactly that), and SMLib's offset is exact, a sphere of radius 22 with no control
points, which surface.fair refuses. Body::is_surface was right: the record lists the offset's
closed shell as the outer shell of a material region (Parts::canonical), so it is a solid
whether read from edge uses or from regions.

The test now fairs the patch on the non-planar rim of a box cut by a sideways cylinder (the
cover in patch_trim_thicken_and_stitch), a one-face B-spline sheet: energy 0.7185 → 0.6938 in one
step, edges unmoved bit for bit, undo/redo/reopen and the refusals as before.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 05:56:11 -04:00
jyvikoandClaude Opus 5.5 298869f4dc Engine tests: the hostile-parameter test builds its starting model once per command
hostile_params_never_panic built a box, a sketch and a rectangle afresh for each of the ~3000
calls it makes per command. SMLib serialises every call behind one process-wide lock, so those
boxes were built one after another whatever the number of threads: 20 threads made 4000 boxes
in 7.8 s, one thread 200 in 0.31 s (1.5 ms each). The fixture was nearly all of the test's time
(8814 of 8823 thread-seconds over the first 72 commands on the aarch64 box; the commands
themselves 9 s, none over 0.5 s), and the test took 2195 s in develop CI.

It now builds that model once per command and gives each call a copy of the session with its
design, model, undo, redo, open sketch and log, sharing the bodies as undo does. Every command
still gets every parameter name with every hostile value, from the same starting state. A new
check runs Cancel Sketch, two undos and a redo on a copy and on a session built afresh and
finds the same designs, with the copied session untouched.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 05:55:52 -04:00
jyvikoandClaude Opus 5.5 f37b954988 Native B-rep: a body is written as it serializes, not through a JSON value
encode_native_body turned the whole saved record into a serde_json
Value (a map with an owned key for every field of every face, edge,
curve and pcurve), checked it, counted it, then wrote it out. Building
that tree was most of the cost: on the Menger level 4 sponge (3 tools,
--save) the 101 bodies the model cache encodes, 535 MB of encodings,
took 24.1 s of encode_native_body, 20.8 s of it in to_value.

binary::write_serialized is a serde Serializer for the tagged binary
format: it writes each value as it is serialized and applies the same
checks as it goes (nesting, string and array budgets, floats within
1e12, the value budget, the size limit). An object's entries are put in
key order when it closes, as the JSON map held them; floats that are not
finite are null, as to_value made them. The bytes are the same: a test
encodes plain, blended, cut, extruded, painted, split, open, round and
far-placed bodies both ways and compares them, and all 101 Menger bodies
(6 to 28 412 faces, 15 kB to 68 MB) encoded both ways in one run were
byte for byte the same. The JSON-value writer stays for the tests that
build hostile records.

Menger L4 3-tool --save, two runs each: encode_native_body 24.1 / 25.3 s
before, 3.0 / 3.0 s after (the longest call 1.34 s -> 0.29 s); the run
41.2 / 41.4 s wall before, 39.1 / 39.2 s after (the encodings run on
16 threads).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 05:55:40 -04:00
jyvikoandClaude Opus 5.5 918429c6dc Sheet machining: a trimmed To Object tool is measured as the prism its record proves
A To Object extrude to a curved face runs its tool through the target and
cuts the target out of it. The SMLib boolean carries no extrusion face
sources (they stood for the untrimmed profile), so sheet_tool_span, which
read the tool's depth and footprint only off its extrusion caps, gave up
and the intersect was refused as unrepresentable.

The tool's current geometry says what it is: the kernel's prism
certificate (prism_measure, the one component copies are compared by)
reads off the exact record whether a body is a uniform prism along a
direction, its two cap levels and its section of lines and arcs. It is
now public as certified_prism_section, and sheet_tool_span uses it when
the caps don't certify the tool: the levels are the depth, the section the
footprint the To Object cut is bounded by. A tool trimmed to something
that is not a prism (a curved roof over it) is still refused.

partial_depth_intersect_curved_to_object_preserves_current_trimmed_cap_footprint
passes: the tool is the square 20 ... 30 less the pin's circle (radius 2),
z 0 ... 1, 100 - 4π, as the test's own witness extrude measures it. A
kernel test reads that section off the trimmed lump.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 05:55:40 -04:00
jyvikoandClaude Opus 5.5 73b0532212 Face capture test: extruding a split face in an uncaptured component is refused
The refusal "capture the source occurrence position before generating a
legacy face profile across components" lost its test when Offset Face got
an exact history name (b6c3bf3c5). A half of a box's top that Split Face
divided still has neither a construction certificate nor an exact history
name, so Extrude from that face, in a component whose occurrence has an
uncaptured move, takes the legacy path and is refused there; the document
and the pending moves stay as they were.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 05:55:40 -04:00
jyvikoandClaude Opus 5.5 a84369fe71 Sheet convert: a wall cut off at its strip keeps each corner with its own edge
end_at_wall takes the plate's end strip off a wall that stands on part of
an edge of the fixed face: the strip's two corners and its two edges go.
They were filtered at their places in the face's lists, counted from the
bend's edge, which pairs corner j with edge j only while the places do
not run across the lists' start. SMLib merges the plate's end strip and
both walls at x = 100 into one L-shaped face whose loop starts on the
strip, so the corners left were each followed by the next one's edge:
the bend's edge took the wall end's face and the wall end took the
plate's. The bend then ran the wall's full width and no relief cut it.
Both lists are now kept counted from the bend's edge.

a_wall_on_part_of_an_edge_meets_a_wall_at_a_relieved_corner_and_is_relieved_at_its_end
passes: bends [22, 92] (was [26, 96]). From the geometry: T 2, R 2, the
setback R + T = 4 on the plate from y = 0 and from x = 100; the round
corner relief 8 across is centred on the setbacks' corner (96, 4), so it
takes 4 off the end of each bend there: the long bend from x = 0 to
96 - 4 = 92, the short one from y = 4 + 4 = 8 to 30, 22 long.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-30 05:55:40 -04:00