Skip to content

port: smlib#124 — the smooth-vertex End, the mixed End's overflow refusal and the seam-vertex continuation (Part of #2358) - #2364

Merged
bherbruck merged 27 commits into
tangency-labs:mainfrom
brett19:fix/2358-smooth-end
Nov 3, 2026
Merged

bherbruck merged 27 commits into
tangency-labs:mainfrom
brett19:fix/2358-smooth-end

Conversation

@brett19

@brett19 brett19 commented Nov 2, 2026 •

Copy link
Copy Markdown
Collaborator

Scope (2026-11-02): a faithful port of tangency-labs/smlib#124, which landed in the fork's replay/1.0.1 at 619c2f0 (see jyviko's comment on tangency-labs/solidcraft#13825). It replaces the earlier "built in solidcore from smlib#81" scope: the fork commits are the reference, matched as they are, with no fallbacks or per-case branches.

Added 2026-11-03: smlib#130 (alpha-1-fixes, tag v1.1.0-alpha.1), ported under fillet_blend_corner:

  • a763ac1: an edge whose faces meet tangentially ends no tangent chain (end_at_smooth_vertex skips it via check_tangential_edgeuse_at_vertex), so the round ends on the end face past it.
  • 9822462: a round that goes second meets the rounds that went first as one blend (FilletExecutive::add_earlier_pass_face / is_earlier_pass_face; OverflowFirst::add_earlier_pass_face for the session). The caller must record the first pass's faces: SolidCraft's two-pass overflow_first wrapper must call add_earlier_pass_face with the faces the first call made, as the fork's capi does.
  • 9b5b521 (capi tests): used only as the source of the cases. The new crates/solidcore/tests/issue_2364_alpha1_ends.rs builds them from SolidCraft's dumps, including the two-call overflow_first form. Without the fix all three fail (13 / 10 / 15 faces).
Case Fork 9822462 (built from source, -ffp-contract=off) solidcore
3369 slot bar, r 2 11 faces, 9763.665662831 11 faces, 9763.665662831
5158 plate, two calls 8 faces, 1681.903764986 (closed form 1681.903758348) 8 faces, 1681.903764966 (the first pass already differs 2e-8: 1706.889016313 against 1706.889016333)
1080 notched block, two calls 13 faces, 10198.718249145 13 faces, 10198.718249145

every-block-cylinder is still capped on the fork at 9822462 (9 faces, 19085.770744810), and solidcore matches. There is no solidcore-only refusal.

Review fixes (2026-11-03):

  • bherbruck 1: the issue_2314 arcs are pinned at 1e-12 to the measured post-main values, 0.15913583961037148 and 0.15913583978226598, which are 3.5e-11 and 5.4e-11 from the fork because of the solidcore-native exact spine. Pappus is asserted at 1.2e-9 and 6e-10 (measured 1.12e-10 and 6.0e-11). In issue_2223, the rim is pinned at 7.9586729268771705 and rim plus corner at 8.117808766496637, both at 1e-12; Pappus is asserted at 1.1e-9 (measured 1.02e-10).
  • bherbruck 2: the WR-697 addendum now says what moves RIM_REMOVED. It is only the volume accuracy (1e-8 to 1e-10); neither the End nor the seam continuation moves it.
  • bherbruck 4: the addendum explains the Join's NO_ROOM across more than one edge. Like the fork's NoteFitRadius(0.0), it names fit 0, because the refusal is topological and has no radius that fits.
  • jyviko: WR-641's removal is filed as Remove WR-641 (fillet_through_extended_rail_end) once the smooth-vertex End is the construction #2500, and deleting the dead indexed FilletBrep::trim_original_surfaces as Delete the dead indexed FilletBrep::trim_original_surfaces (p6.rs) #2501. That function has no callers outside its own unit tests, and the difftest trait impls are missing() stubs.

Units: blend (fillet executive, fillet geom, blend corner End and Join)
In scope (in this order, all under the existing fixes.fillet_blend_corner, WR-697 addenda citing the fork SHAs):

  • d657afa: TrimOriginalSurfaces: an edge of more than two faces bounds the trim (marked in DoFilletingSteps), and a leaking trim is SM_FILERR_BAD_MERGE (not dropped). Regression case: SMLib fillet: segfault in SmFilletSolver::GetOtherFilletCorner (rig case fillets/1390) #453's back rim.
  • 41a39ec: the End at a smooth vertex (SM_BEND_AT_SMOOTH: EndAtSmoothVertex, MakeEndAtSmoothVertex, CalcEndCapEdges, the cap face in CreateEndPatchOfEnd, the cap surface in CalcRailIntEndSurface and the fillet geometry's side-face intersection); a rail past its side edge's far end refused naming the radius that reaches it (RefuseEndPastSideEdge, RadiusReachingSideEdgeEnd, FitOfEndPastSideEdges in the regularity check); the Join's refusal across more than one edge becomes SM_FILERR_NO_ROOM with no size; the executive keeps a corner's noted NO_ROOM.
  • 54a4ee7: a concave End's rail passing a side edge of the other convexity is refused, naming the largest radius (SmBlendCorner::BallTouchingAt).
  • 63c7962: the rail ending on its side edge continued across a seam's vertex (sm_SideEdgeuseAcrossSeamVertex, SideEdgeChainTo), the refusal taken over the whole chain.
    Not ported: the capi test commits (e4a4ac0, a889295, 17ea413, 4b8a1e7, ce52279, ad30d68); they are used only as the source of inputs and expected values (named in each test).
    Expected edits: crates/solidcore/src/blend/** (blend_corner, fillet_corner, fillet_executive, fillet_geom), regression tests under crates/solidcore/tests, a WR-697 addendum in plan/upstream-issues.
    Reconciliation: WR-674 (fillet_round_ends_at_tangent_continuation, Solidcore-only blend fix: a round ending at a tangent continuation of its edge, built by subtraction (fillet_round_ends_at_tangent_continuation, WR-674) #2326) and WR-706 (fillet_section_end_trim_at_creases, Solidcore-only blend fix: a section-ended round trimmed at its creases with other rounds (fillet_section_end_trim_at_creases, WR-706) #2327) cover the same rounds; with fillet_blend_corner the fork's smooth-vertex End is the construction, and the overlap is recorded in the addendum.
    Evidence: fork 18e96f1 (contains 619c2f0), built from source with -ffp-contract=off and run through SmApi* on the capi cases' inputs. Volumes are compared at 1e-10:
Case Fork 18e96f1 solidcore Test
#453 rim, round 3 6382.568447649069, 11 faces 6382.568447648735, 11 faces issue_453_bore_rim_smooth_end.rs
#453 rim, chamfer 3 6316.944421027456, 11 faces same, 11 faces same
#453 rim, blend corner off - 1001 (as pinned SMLib) same
boss tilt 20, r 1.5 1001 NO_ROOM, fit 0.262514174 same issue_2358_mixed_end_overflow.rs
boss tilt 20, r 0.25 / 0.2625 (removed) -0.576293097954 / -0.635789172127 -0.576293096943 / -0.635789174601 same
boss tilt 40, seam 125.13, r 1.5 10 faces, -33.432934277429, rail ends (-4.176377478, +-4.664695271, 0.893582223) 10 faces, -33.432934278928, same vertices to 1e-8 same
boss tilt 20, seam 170, r 1.5 / 0.262514174 1001 fit 0.262514174 / -0.636117736700, 10 faces same / -0.636117743787 (the union is already 7.7e-9 apart) same
#2176 tab top / rim / corner arc 0.214601836334623 / 7.958672926831696 / - equal / 4.5e-11 / 0.159135839575356 (also with WR-674 off) issue_2176_round_of_arc_between_tangent_straights.rs
#2176 tab side 0.344196265039045 (the fork's body has a Bad Gap) 0.344196254843155, check_brep clean (#2307's FlushCurve fix, solidcore's own) same

The 2358 tests turn off cylinder_plane_section_trims_exact and precise_properties_split_at_surface_knots, which the fork does not have. With them on, the union measures 3.6e-5 less.
issue_2223 and issue_2314 now use the fork's values with the blend corner on. Their tests without the blend corner keep WR-674 and WR-706 under test.
Wall split at 120/240 degrees (the fork's capi case junction_boss_tilt_40_wall_split_short_of_the_rail_end_refused, ad30d68): the wall split as the capi's smlib_split_faces cuts it (the plane section merged into the wall through the plane), rounded 1.5. Fork 18e96f1: 1001, SM_FILERR_NO_ROOM, fit 0.583942733779 (closed form (3 sin 40 - 2 cos 40) / (1 - sin 40 / 2) - 1e-5 = 0.583942733780). solidcore: 1001, fit 0.583942733779. New test tilt_40_wall_split_short_of_the_rail_end_refused in issue_2358_mixed_end_overflow.rs.

Expectations updated to the fork after smlib#124 (fork driver at b785227 = before smlib#124, 63c7962 and 18e96f1 = after; the reference is the fork):

🤖 Generated with Claude Code

https://claude.ai/code/session_01SdcgEtwSrdwZxqTQp5MfnU

…al, opt-in (smlib#81 port, #2358)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SdcgEtwSrdwZxqTQp5MfnU
@brett19 brett19 changed the title fix: the smooth-vertex End and the mixed End's overflow refusal, opt-in (smlib#81 port, #2358) fix: the smooth-vertex End and the mixed End's overflow refusal, built in solidcore (from smlib#81; fork WIP stops at its tag) Nov 2, 2026
…ounds it (smlib d657afa, fillet_blend_corner)

TrimOriginalSurfaces' leaking trim (a collected face is the rail's face on its
other side) was dropped, so the fillet's faces were merged into a body left
whole. With fixes.fillet_blend_corner it is SM_FILERR_BAD_MERGE ("the trim
leaks") and SM_ERR, as the fork's d657afa. DoFilletingSteps' trims also mark
every target edge of more than two faces before the trims (the bore wall
touching the bottom face along a line, tangency-labs#453): the faces
beyond it are not the blend's. trim_original_surfaces_with_solver returns
whether the trim leaked; the live session notes the error.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SdcgEtwSrdwZxqTQp5MfnU
@brett19 brett19 changed the title fix: the smooth-vertex End and the mixed End's overflow refusal, built in solidcore (from smlib#81; fork WIP stops at its tag) port: smlib#124 — the smooth-vertex End, the mixed End's overflow refusal and the seam-vertex continuation (Part of #2358) Nov 2, 2026
brett19 added a commit to brett19/solidcore-fork that referenced this pull request Nov 2, 2026
… rolling over onto the top face before its smooth End (failing)

The 851 arc-wall plate's top front edge rounded R2, then the middle edge between that
round and the arc wall rounded alone (the fork's capi
scenario_middle_edge_of_an_arc_tangent_chain_r0_25). The rail on the R2 round meets the
round's tangent edge with the top face before the top End's plane P, at the angle
asin(2 / (21 + r)) about the wall's axis, short of the vertex's asin(2 / 21): past it the
ball rolls on the top face. The tests check the construction against its closed forms:
the rolled rail on the top face's circle of radius 21 + r, the cap's edge on the top face
its line in P from that rail to the vertex, the end rib the ball's circle in P, the
rollover seam the ball's circle in the section plane at the crossing, G1 across the seam
and along the rails, the R2 round's rail on the round with its ball touching the wall, and
check_brep Level2. They fail today: the smooth-vertex End (smlib#124, tangency-labs#2364) is not in yet.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SdcgEtwSrdwZxqTQp5MfnU
brett19 and others added 6 commits November 2, 2026 17:19
…flow refusal and the seam-vertex continuation (fork 41a39ec, 54a4ee7, 63c7962; smlib#124)

Under fixes.fillet_blend_corner (WR-697), no new flag:

- 41a39ec: a round whose edge runs on G1 into an unblended edge of the
  same two faces ends at that smooth vertex (BlendEnd::AtSmooth), cut by
  the plane through it square to the edge (FilletCorner::end_cap,
  make_end_at_smooth_vertex, calc_end_cap_edges). A blend End's edges are
  the fork's CalcEndEdgeGeom (blend_corner::end::calc_end_edge_geom)
  instead of drop_end_rib; a corner that noted its reason keeps it.
- 54a4ee7: a concave End of mixed convexity whose rail passes over the
  side edge of the other convexity is refused (SM_FILERR_NO_ROOM), naming
  the largest radius whose rail still ends on the edge (BallTouchingAt,
  golden section, at most 200 steps) less the part's tolerance.
- 63c7962: a side edge split only by a seam ending on it is continued
  across that vertex (side_edgeuse_across_seam_vertex, side_edge_chain_to,
  at most 64 steps); the rail end search continues on the next edge
  (recursion bounded at 64) and the End's edge on it joins the chain.

Tests: issue_453_bore_rim_smooth_end (the fork's volumes: chamfer exact,
round to 3.3e-10), issue_2358_mixed_end_overflow (tilt 20: refused naming
0.262514174, builds at 0.25 and 0.2625 within 2.5e-9 of the fork), and
issue_2176/2223/2314 updated to the fork's values, with WR-674 and WR-706
kept under test with the blend corner off. Not yet building: the boss
tilted 40 degrees with the seam at 125.13 degrees (the 63c7962 case),
which fails the corner's edges here.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SdcgEtwSrdwZxqTQp5MfnU
…fix/2358-smooth-end

issue_2176's tab side round keeps tangency-labs#2307's check_brep check; with its
FlushCurve fix (solidcore's own) the round removes 0.344196254843155,
1.0e-8 less than the fork's 18e96f1, whose body has the Bad Gap.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SdcgEtwSrdwZxqTQp5MfnU
… circle's arcs (fork 63c7962)

CalcEdgeOnEdge's chain copies each edge's curve as an SmBSplineCurve; a
circle or a line is one (SM_CAST_PTR), trimmed as its own class and joined
through its B-spline. The port took only a B-spline kind, so the boss
tilted 40 degrees with the seam at 125.13 degrees failed the corner's
edges (1001). It now builds as the fork's: 10 faces, -33.432934278928
removed (fork -33.432934277429), the rails' ends at the fork's vertices.

Tests (issue_2358_mixed_end_overflow): that case, and the boss tilted 20
degrees with the seam at 170 degrees (refused naming 0.262514174 over both
edges of the split rim; builds at that size within 7.1e-9 of the fork,
the union 7.7e-9 apart already).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SdcgEtwSrdwZxqTQp5MfnU
…ncy-labs#2175 chamfer; the wall split case

Run on the fork at 63c7962 and 18e96f1 (smlib#124) and at b785227 (before it):
- sphere_plane with every fix: the fork builds since smlib#124 (5 faces, volume
  3272.764885312161), as the port does; the mark becomes that volume. The closed form is
  3261.518762521043, recorded in the test.
- tangency-labs#2175's chamfer with the blend corner: the fork refuses it since smlib#124 (EDGE_PROBLEM,
  the through point at the bottom end fails); the port refuses it NO_ROOM at the
  smooth-vertex End because WR-641 skips that through point. Without WR-641 it returns the
  fork's EDGE_PROBLEM; both are now asserted.
- The wall split at 120/240 degrees (the fork's capi ad30d68 case): refused naming
  0.583942733779 on the fork and in the port; new test.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SdcgEtwSrdwZxqTQp5MfnU
@jyviko

jyviko commented Nov 2, 2026

Copy link
Copy Markdown
Collaborator

Two fixes to take in this port, coming in SMLib v1.1.0-alpha.1. Without them the port carries two defects of the fork's smooth-vertex End:

  1. No End where the run-on edge bounds a lead round. The overflow-first pass rounds one edge first. The next edge then runs on tangentially into that round's end, and the End caps it, leaving a square shoulder of side face with ΔV = r³(π/4 − 2/3). The fix: the first pass hands its new faces to the executive, and the End skips edges bounding them. Cases: SolidCraft rig 5158 (back to 1681.903764986) and 1080 (10198.718249146).
  2. A tangent seam isn't a smooth vertex. An edge ending on a flat/round tangent seam (G1, no convexity) gets no End. Case: rig 3369, 11 faces.

The alpha.1 commits will be linked here when they land.

…lter (clippy)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SdcgEtwSrdwZxqTQp5MfnU
brett19 and others added 2 commits November 2, 2026 19:51
…m, bounds)

- issue_2314 and issue_2223: the corner arcs pinned again at the fork's figures (18e96f1 and
  63c7962, SmApiFilletEdges r 0.5): 0.159135839575356 and 0.159135839836381 to 1e-12, the
  rim and corner pair 8.117808766409325 to tangency-labs#2176's 1e-10 (8.7e-11 here, 4.5e-11 of it the
  rim's), the volume at the fork's 1e-10 accuracy. RIM_REMOVED is tangency-labs#2176's fork figure,
  7.958672926831696.
- WR-697 addendum: the tangency-labs#2175 code difference (8 through WR-641, the fork's 4), the
  sphere_plane gap to its closed form (tangency-labs#2441) and the wall split; FitOfEndPastSideEdges
  credited to 41a39ec. The sphere_plane test cites tangency-labs#2441.
- fillet_geom::p2's header names the seam-vertex recursion and its bound of 64.
- The indexed corner path notes SM_FILERR_VERTEX_PROBLEM when an End at a smooth vertex's
  cap patch fails, as the fork and the session path do.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SdcgEtwSrdwZxqTQp5MfnU
…k after origin/main's merge

origin/main's 2026-11-02 merges move each corner arc's round 3.5e-11 and 5.4e-11 from the
fork's figures (within 1e-12 before); the pins use tangency-labs#2176's 1e-10 for this arc.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SdcgEtwSrdwZxqTQp5MfnU
@brett19

brett19 commented Nov 2, 2026

Copy link
Copy Markdown
Collaborator Author

Thanks. We'll take both alpha.1 fixes (the lead-round End skip and the tangent seam) in this PR once the commits land on tangency-labs/smlib, with cases 5158, 1080 and 3369 as regression tests. The current head is 10331dc, with main merged.

Written by Claude Code (Anthropic) on behalf of @brett19.

brett19 and others added 3 commits November 2, 2026 20:57
… (fork smlib#124)

Under fillet_blend_corner, as the fork's SmFilletExecutive.cpp 3755-3759 and 1216-1219:
create_corner_vertices keeps a corner's SM_FILERR_NO_ROOM, or a CORNER_UNSUPPORTED it
noted itself, and preview_fillet_geometry keeps its NO_ROOM, instead of overwriting them
with VERTEX_PROBLEM "Unable to Make Topology". Only the CalculateCornerEdges site kept it
before.

Regression test: SolidCraft's drafted peg, front top edge and concave arc rounded r 2
together (its rig's input body): NO_ROOM, fit 0, the corner's comment (3 without the fix).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SdcgEtwSrdwZxqTQp5MfnU
…call by shared kernel (clippy)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SdcgEtwSrdwZxqTQp5MfnU
@brett19

brett19 commented Nov 2, 2026

Copy link
Copy Markdown
Collaborator Author

@jyviko, a question on a third case, from triaging SolidCraft's rig run of this PR. All results are at head 3dcc89a, compared with fork 18e96f1; the fork and this PR give the same body in every case.

  • Your lead-round defect (ΔV = r³(π/4 − 2/3), square cap with a 90° seam) also accounts for 1207-box-end-edge-into-unblended-round, every-fillet-after-fillet-curvature-front-then-side-r3, SolidCraft's a_round_next_to_a_round_of_another_radius and its 851 arc-then-front test. On 1207, before smlib#124 the fork built 7957.505245 with 8 faces; now it builds 7957.723438 with 9.
  • Your tangent-seam defect accounts for 3369.
  • The third case is every-block-cylinder-tangent-concave-arc-r2. The edge runs on into a concave arc that is not blended and has no lead round. The smooth-vertex End (41a39ec) caps it with a 90° G0 seam: 19085.770745, 9 faces. Before smlib#124 it was refused with 1001 (the tangent-propagation hint), as the case expects.

Does alpha.1 cover the third case? If not, would you refuse in EndAtSmoothVertex when the run-on edge is not blended, rather than cap it? We'll wait for your alpha.1 commits before changing anything here.

Written by Claude Code (Anthropic) on behalf of @brett19.

@Karl-VonRichter

Copy link
Copy Markdown
Contributor

📢 SolidCore follows the SMLib fork now. Please re-read AGENTS.md and CONTRIBUTING.md

SolidCore's reference is now our SMLib fork (tangency-labs/smlib), not NVIDIA's usd-brep v1.0.1.
• Where the fork replaced a construction (e.g. the generic rolling-ball blend), SolidCore matches the fork bit for bit. Differences from upstream there are the fork's changes, not port bugs.
• Where the fork hasn't changed anything, upstream and fork agree, and difftest keeps checking that.
• Flags only while a port is in progress. At parity, the flag becomes the default and the old path is deleted. No permanent fallbacks.
• Fix root causes, not single cases. No try/catch trains. Tests first.
• Kernel is the unit of concurrency: one per thread, no global state.

Everyone: re-read AGENTS.md and the new CONTRIBUTING.md before your next PR.
https://github.com/tangency-labs/solidcore/blob/main/AGENTS.md
Details: #2392

brett19 added a commit to brett19/solidcore-fork that referenced this pull request Nov 3, 2026
…abs#2410) into fix/2343-curved-setback

tangency-labs#2382's new scale test builds a `blend_setback::Face` literal; it takes
the curved fields' defaults (`..Face::default()`), as the other literals
of the plane-only face do since this branch added the quadric's fields.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SdcgEtwSrdwZxqTQp5MfnU
brett19 added a commit to brett19/solidcore-fork that referenced this pull request Nov 3, 2026
…y-labs#2550)

wr698's reopen wrote the rebuilt ProceduralBlend into the face slot
directly, so the surface had no owner and SmSurface::GetFace() was null.
Upstream every face surface is set through SmFace::SetSurface, which
makes the face its owner. Since tangency-labs#2364 the smooth-vertex End merges the
second round's rail onto the reopened round (MergeCurveOnSurfaces), and
the null face fails it with SM_ERR_NULL_POINTER (1006), as upstream
would. The old WR-674 path never merged a rail onto that face, so it
hid the missing owner. The plain B-spline run's surfaces were always
owned.

reopen now uses face::p1::set_surface and asserts that the face owns
the surface. a_round_on_a_reopened_round_lies_on_it passes with
--ignored: 1006 before this change.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SdcgEtwSrdwZxqTQp5MfnU
MarkusReichart pushed a commit to MarkusReichart/solidcore that referenced this pull request Nov 3, 2026
…ince tangency-labs#2364 (tangency-labs#2550)

tangency-labs#2510 makes it fast enough to run, but on main since tangency-labs#2364 the round on
the reopened round fails in merge_curve_on_surfaces with status 1006
(tangency-labs#2550). It runs again when that is fixed.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DtL5M41G8Nua7HvuFPZkmJ
@bherbruck

Copy link
Copy Markdown
Collaborator

A question on #2364 (the smlib#124 port), from #11360's line 2 (a_round_ending_in_an_earlier_round_matches_both_together). We reproduced it in solidcore with fillet_blend_corner on.

The part: a 41⅙ × 21 rectangle and, beside it, what a circle of radius 21 about (62⅙, −21) leaves against the top line it touches, extruded 10. That gives an arc wall square to the front, and a knife edge at its tip (62⅙, 0).

The case: each first edge is rounded at r 2, then the upright at the tip is rounded at r 2 on that result. There are two first edges:

  • the arc's top edge (the point (47.32, −6.15, 10) on it);
  • the back top edge (the point (30, 0, 10) on it).

What changed with #2364: at main just before it (fb937dd), both second rounds build. Their volumes match the same edges rounded together within 2.6e-5 and 4e-12, and check_brep Level 2 is clean. At #2364's merge (845bdb4), the second round is refused with status 8:

Fillet Error: the round's contact line runs past the end of its side edge at the corner (onto the faces beyond): a radius below 1.04762 fits

(1.05 for the back-edge case). The same edges rounded together at r 2 still build there. git log -S puts the message in 5a1cb65, the mixed End's overflow refusal (fork 54a4ee7).

Question: should that refusal fire at this end? The upright's round ends where the first round's surface (and, for the arc, the knife-edge tip) meets it, so I'd expect its "side edge" to be the earlier round's boundary rather than an original face's edge. We haven't traced which edge it is. Before #2364 the build was valid and matched rounding the edges together. If the refusal is right here, its named radius (1.05) should build; if it should only apply where the side face is an original face, could it be scoped to that?

@brett19

Written by Claude Code (Anthropic) on behalf of @bherbruck.

@brett19

brett19 commented Nov 3, 2026

Copy link
Copy Markdown
Collaborator Author

@bherbruck, we're working on this, in draft #2610.

Written by Claude Code (Anthropic) on behalf of @brett19.

brett19 added a commit to brett19/solidcore-fork that referenced this pull request Nov 3, 2026
…ilds (tangency-labs#2364 review)

With fillet_blend_corner, an End on its end face G whose rail on F runs
past the far end P of its side edge was refused NO_ROOM (the fork's
41a39ec RefuseEndPastSideEdge / FitOfEndPastSideEdges), also where G runs
on tangentially past P into the face H that F's next edge bounds. There
the rail ends on H, G and H are one smooth surface, and the global merge
(WR-303's dropped end rib) cuts the round exactly, as it did before tangency-labs#2364.

end_face_runs_on_past_side_edge is the narrow rule (P has exactly the
side edge, an F-H edge, and a tangential G-H edge). Where it holds,
calc_end_edge_geom deletes the end rib and returns EndEdges::GlobalMerge,
and the pre-corner fit check does not size that side edge. Every other
past-side-edge refusal is unchanged.

SolidCraft #11360's a_round_ending_in_an_earlier_round_matches_both_together:
the arc-wall plate's arc top (or back top) edge rounded r 2, then the tip's
upright r 2. Before this change it was refused naming 1.04762 / 1.05 (the
fork at v1.1.0-alpha.1 too, also with AddEarlierPassFace). It now builds:
9452.927851913 (2.7e-9 from rounding both together) and 9428.557343079
(equal), check_brep clean. That is main before tangency-labs#2364's result.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SdcgEtwSrdwZxqTQp5MfnU
@brett19

brett19 commented Nov 3, 2026

Copy link
Copy Markdown
Collaborator Author

@bherbruck Your expectation was right: the side edge is the earlier round's boundary, and the refusal was wrong here. Fixed in #2610 (ready; local gate passed). Reproduced on main ce1c7f9 with the rig set (every fix but the six it leaves off).

Which edge. The refusal is FitOfEndPastSideEdges (the fork's 41a39ec), at the upright round's End on its end face at the top of the tip, (62⅙, 0, 8). In both cases the side edge is the first round's boundary, not an original face's edge:

  • Arc top first: the first round's edge on the back face, from (62⅙, 0, 8) to P = (52.786, 0, 10).
  • Back top first: the first round's edge on the arc wall, from (62⅙, 0, 8) to P = (53.222, −2, 10).

In both, the rail at r 2 runs past P. But P is on the first round's top rail, where that round runs on tangentially into the top face. So the rail ends on the top face, and the end face is the smooth surface made of the first round and the top face.

add_earlier_pass_face alone doesn't fix it, in solidcore or in the fork. The fork's AddEarlierPassFace was given the first call's 5 faces (arc case) or 3 faces (back case), and both still refuse with the same fit. 9822462 skips earlier-pass edges only in EndAtSmoothVertex, and this corner isn't one: it's an End on its end face in both versions.

The named radius builds.

  • Arc: r 1.04762 gives 8 faces, 9519.736086559604.
  • Back: r 1.05 gives 8 faces, 9495.096950592540.
  • check_brep is clean for both, and the fork gives the same volumes, 9519.736086559600 and 9495.096950592537 (4e-16 relative).

The fork at v1.1.0-alpha.1 (8eac63c; built at 6843885, the same SMLib and SM_API source) refuses both at r 2: status 1001, SM_FILERR_NO_ROOM, fit 1.04761905 and 1.05, with the same message. It refuses with and without AddEarlierPassFace. Rounded together it builds 9452.927826170266 and 9428.557321606453.

Before #2364 (fb937dd) the end rib was dropped and the global merge cut the round (WR-303). That built 9452.927851913 (arc) and 9428.557343079312 (back), against 9452.927826170 and 9428.557343079308 rounded together.

The fix (#2610) is under fillet_blend_corner, with no new flag, since it turns a wrong refusal into a correct build. It is a narrow rule, end_face_runs_on_past_side_edge, which applies when the side edge's far end P has exactly three edges:

  • the side edge;
  • an edge bounding the rail's face and a face H;
  • an edge between the end face and H whose faces meet tangentially at P.

Where it holds, the End's end rib is dropped for the global merge, as before #2364, and the pre-corner fit check doesn't size that edge. Every other past-side-edge refusal stays. The boss's split wall is still refused at 0.583942734: past P its end face meets a crease. The mixed End's overflow refusal (54a4ee7) is unchanged.

The result:

The regression test is issue_2364_round_ending_in_an_earlier_round.rs, which checks both against rounding the edges together. It fails without the fix. The fork shares the refusal; the WR-697 addendum in #2610 records it.

Written by Claude Code (Anthropic) on behalf of @brett19.

bherbruck pushed a commit that referenced this pull request Nov 3, 2026
…builds (#2364 review)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SdcgEtwSrdwZxqTQp5MfnU
bherbruck added a commit that referenced this pull request Nov 3, 2026
fix: an End whose end face runs on tangentially past its side edge builds (#2364 review)
brett19 added a commit to brett19/solidcore-fork that referenced this pull request Nov 4, 2026
… the end face, not capped (#2656)

The End at a smooth vertex (smlib#124, tangency-labs#2364) capped every blended edge that runs on
tangentially into an unblended edge of its convexity. Where the continuing edge turns into
the two edges' shared face away from the blended edge continued (an earlier round's end
arc, a round between two faces of a corner), the blend run on runs out of the part, and
the End on the end face builds that, as before tangency-labs#2364. `run_on_leaves_the_part` compares the
two edges' heights along the shared face's binormal past the vertex (second and third
order, over the shorter chord, against the face's zone tolerance). tangency-labs#2176's corner arcs and
tangency-labs#453's rim keep their caps.

Under fillet_blend_corner (it restores results built before tangency-labs#2364). Regression tests from
SolidCraft's cases: ladder 012 and 021, fillets/1207, 5158 and 1080 in two passes without
add_earlier_pass_face, its picked-edge and earlier-round tests, the curvature side round;
and the tangency-labs#2369 outside-setback corner with SolidCraft's calls (builds on main). WR-697
addendum.

Part of #2656

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SdcgEtwSrdwZxqTQp5MfnU
bherbruck pushed a commit that referenced this pull request Nov 4, 2026
…(#2656)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SdcgEtwSrdwZxqTQp5MfnU
bherbruck added a commit that referenced this pull request Nov 4, 2026
fix: the blend regressions SolidCraft sees from #2364 and #2369 (#2656)
jyviko added a commit that referenced this pull request Nov 4, 2026
blend (round, curvature, chord, equal chamfer; corners set back, filled or opened, the holes
returned), chamfer_sides (a distance along a named face, a second distance or an angle),
fillet_variable (a radius law from a vertex) and edge_faces_along (for the application's side
choice), each a sequence of FilletSession calls computing no geometry.

When some rounds overflow their faces across cliffs and others fit, the overflowing ones go
first on the untouched input and the others then round what is left of their edges in a second
pass that is handed the first pass's faces (FilletSession::add_earlier_pass_face, #2364), so it
meets them as one blend instead of ending at a smooth vertex. Tested on the plate and the
notched block of issue 2364: 8 and 13 faces, the kernel's two-pass volumes.

Setbacks keep the add_set_back_vertex + set_corner_setback sequence on N x N convex or
Setback-event corners. The law's corner check is the kernel's FilletLaw::corner_parameters.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NvoxKMbMFSQU8tzRtoJhtT
ablasik pushed a commit to ablasik/solidcore that referenced this pull request Nov 4, 2026
… retired (tangency-labs#2647)

`UpstreamFixes::fillet_round_on_round_as_definition` is gone and its path is the
only one: a round of one radius on a face that is a round of one radius by its
definition is built as its definition and certified on it, and a definition
evaluated at one of its stations takes the side it is asked for. It still needs
`fillet_round_as_definition` and `fillet_gn_section_generator`, which build the inner
round as its definition.

The test of the old path (`without_wr714_the_round_on_the_round_is_built_as_before`)
is deleted with it; the SmoothInteriorPlan opt-in test now turns off
`fillet_round_as_definition`, the one opt-in left; tangency-labs#2364's rig kernels and the
corpus's fixes.json drop the name.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DtL5M41G8Nua7HvuFPZkmJ
@jyviko jyviko added this to the Blends complete milestone Nov 6, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

Status: Done

Development

Successfully merging this pull request may close these issues.

4 participants