Skip to content

Blend End: the smooth-vertex End (smlib#124, #2364) removes half the material on the sphere_plane round (3272.765 vs closed form 3261.519) #2441

Closed

Description

@brett19

The exact-class case sphere_plane (crates/solidcore/tests/fork_099cf96_exact_class_cases.rs) builds with 5 faces and a volume of 3272.764885312161. The closed form for the rounded body is 3261.518762521043:

4000π/3 − 288π − π(40.5 − 2√60(π − √3.75/4 − 2 asin ¼)) = 3261.518762521043

The round should remove 22.4928 of material, but it removes only 11.2466, about half.

Expected: 3261.518762521 with the same 5 faces, or a typed refusal. A built body that is short by half the round's material is worse than a refusal.

Duplicate search: solidcore and smlib issues and PRs, any state, for "sphere_plane", "3272.76", "3261.5", "fork_099cf96", "sphere-plane", "smooth-vertex End" and "sphere round half material". Nearby but different: tangency-labs/solidcraft#13825 and #2406 (other smooth-vertex End cases), and smlib#81 and smlib#98.

Reported by the solidcore swarm from its #2364 work.

Filed by Claude Code for brett19.

Activity

  1. MarkusReichart commented on Nov 2, 2026

    @MarkusReichart
    Contributor

    Measured on Linux (WSL, Rust 1.98.0): with both edges of the rim selected, sphere_plane builds to the closed form. So 3272.765 looks like an effect of the test's edge selection, not a defect of the smooth-vertex End.

    The rim z = 4 is two edges, joined at (±√84, 0, 4). The test's one point (0, √84, 4) selects only the half y > 0, so the round has to end at the vertex where the unselected half carries on tangentially. That makes it a half-rim round, which can't be compared with the full rim's closed form. The fork changed its capi case for exactly this in tangency-labs/smlib@18bd2f5: it now selects a point on each edge, and its expected volume stays the same.

    sphere_plane, all fixes (solidcraft_kernel, as in mark_current), second point (0, −√84, 4)

    Branch Picks Faces Volume Off the closed form
    #2364 3dcc89a9f 1 point 5 3272.764885312 +11.246
    #2364 3dcc89a9f 2 points 4 3261.518763866 4.1e-10 rel
    main 552bbdd41 1 point – status 1001 –
    main 552bbdd41 2 points 4 3261.518763866 4.1e-10 rel
    • Closed form: 3261.518762521043; the test's gate is 1e-6 rel.
    • Every built body has 0 lines from check_brep at level 2, and its round faces are revolved and 2.6e-14 off the ball.
    • With two points, both branches return the same bits. The full rim never reaches the End.
    • fillet_exact_round_classes alone gives 3261.518786470 with two points on both branches (7.3e-9 rel), which also passes.

    cylinders_touching_at_a_point works the same way. Its concave arc is split at (5, 5√3, −5) by B's seam, and both of the test's points lie on the long piece, so it stays at 1001. With the fork's third point (5 cos h, y, −5 + 5 sin h), h = asin(0.6) / 2, it builds on both branches: 13 faces, 13442.703241790 (8.3e-9 rel), rational (12, 2) pieces 3.1e-13 off the ball, valid at level 2.

    Suggestion: select every edge of both rims in fork_099cf96_exact_class_cases.rs, as the fork has done since 18bd2f5. Both cases would then pass on main under mark_after and mark_current, and #2364 and #2410 would need no WrongVolume mark for sphere_plane. One side effect: on the default kernel (mark_before), the full rims give traced (3, 3) rounds instead of 1001 (1.40e-3 and 5.96e-4 off the ball), so that column's marks would need updating too.

    Not checked: whether the half-rim body is what the End should build. It removes 11.246635, a little more than half of 22.492758. That question is separate from the closed form.

    How it was measured: I appended a probe test to the file locally (not pushed). It runs each case with the test's points and with every rim edge, on the default kernel, the flag alone, and all fixes. cargo test -p solidcore --test fork_099cf96_exact_class_cases -- --nocapture exited 0 on both heads.

    Measured and written with Claude Code.

  2. added
    urgentNeeds attention first
    and removed
    urgentNeeds attention first
    on Nov 4, 2026
  3. MarkusReichart commented on Nov 5, 2026

    @MarkusReichart
    Contributor

    Taking this as the test follow-up @jyviko asked for in #2653 (comment) (#2653 landed as #2708).

    Plan, test-only: in fork_099cf96_exact_class_cases.rs, sphere_plane selects both edges of its rim z = 4 and is checked against the closed form 3261.518762521043; the pipes get a point on the short piece of their arc as well, as the fork does since tangency-labs/smlib@18bd2f5. Mark::WrongVolume goes. I'm re-measuring both cases on today's main first; if either misses its closed form, it becomes a known-failing test named for the defect and I'll post the numbers here. The half-rim End stays in tangency-labs/solidcraft#13784.

    Claude Code, for MarkusReichart.

  4. added 6 commits that reference this issue on Nov 5, 2026
    5fc7823
    1591cd1
    5afac03
    8b9f556
    ef578f1
    6c1e8f7
  5. added a commit that references this issue on Nov 6, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions