[bherbruck]

Moved from tangency-labs/solidcore#3167

Fixes #13650 (solidcraft#12785).

The kernel's own results carry a solid that touches itself along a line (a hole's bottom meeting two pocket walls along a segment of their corner edge) as one edge of all four faces, and the merge handles that form. The record format writes such an edge as one edge per face pair (split_touching), and the body SolidCraft hands its next feature comes back with two coincident edges of two faces each; it passes the check, but the merge of a tool whose edge lies along that line gives a body that is not a closed solid (default fixes) or fails in topology_intersector::p1::relate with 1001 (SolidCraft's fix set).

New opt-in flag Kernel::fixes.boolean_glues_self_touching_edges: merge_boolean_by first glues, in each operand, the coincident edges that share both vertices, the curve kind and their points (three samples within the B-rep tolerance) and whose faces are all distinct, into one edge (GlueEdgesGeneral, no gap, with region nesting), in one shell or in two shells touching along the line (the round trip leaves the self-touching solid in two), restoring the kernel's own form. Edges whose faces repeat (the kernel's crack edges, which its check refuses glued) are left alone. Off by default: upstream merges the operand as given. Record: plan/upstream-issues/WR-754--boolean-operand-touching-itself-along-two-coincident-edges.md.

SolidCraft follow-up once this merges: add the flag to its fix list (crates/smlib/src/core.rs).

Changes a default: no

Tests: crates/solidcore-ops/tests/boolean_join_fills_pocket_with_hole_at_its_corner.rs (committed first, failing): the reporter's block, pocket and hole built with the ops crate; the join of the cut's own body (passes without the flag), the join after a record round trip with the flag (fails without it: not a closed solid), and the control without the hole. Also run, laptop: clippy on solidcore and solidcore-ops (all targets) clean, cargo fmt clean on the changed files, the boolean unit tests 76/76, the ops boolean tests 5/5. Pending: the full suites and the kernel cases on the remote host; results follow in a comment.

@dudedagger42, this is your PR moved here. If you can't push to moved/solidcore-3167, open a PR from your fork against kernel that says 'Replaces #13576' (see CONTRIBUTING, 'Updating a PR that was moved from solidcore').

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

[dudedagger42]

On the solidcore tree before the move (0c222ff64; the same commits are on moved/solidcore-3167, rebased onto 9b559c4).

Kernel cases, whole corpus, with the harness's case budget raised to 1800 s (KERNEL_CASES_TIMEOUT_S), release, laptop, 5601 s: 1079 cases, 799 pass, 0 fail, 263 unsupported, 0 unported, 17 known-failing; fillets/cov-fillet-section-curvature-boss-root-r2 passes inside that budget. The harness still asserts, on one known-failing case whose result moved from the rig's recording: fillets/report-peg-front-and-concave-chamfers-together-d2, volume 9383.450962 against the rig's 9383.450660 (3e-8 relative). The glue code runs only behind boolean_glues_self_touching_edges, which the corpus's fixes.json does not set, so that movement is the base's (the solidcore main this branch sits on), not this PR's.

fillets/cov-fillet-section-curvature-boss-root-r2 alone (1800 s budget, release, laptop, back to back in one run): with boolean_glues_self_touching_edges off (the corpus's fixes.json as committed) 616 s; with the name appended to that list, on, 810 s. Both ran while another release build and a corpus run shared the laptop (the on run overlapped more of it), so the gap is not a measurement of the pre-pass; the release timing of the pre-pass on a 1600-sided prism (above) showed none. A quiet repeat on an idle machine is listed for the kernel lanes. The harness's fix set is the corpus's fixes.json (the rig's 297 names), so the flag is off in every corpus run as committed; the case's slowness is solidcore main's since 99ed57c (tangency-labs/solidcore#3238, filed before the move).

[brett19]

@bherbruck, #16392 carries this PR's three commits, cherry-picked with their authorship and credited. We did this under the solidcore stall rule: we poked you on #13650 at 2026-11-13 22:34Z and had no reply within 15 minutes. On top of your commits, #16392 removes the boolean_glues_self_touching_edges flag, so the glue is always on. It also replaces the 3-sample coincidence test with a bound over each edge's whole span: an end-point bound for lines, certified_distance for non-rational B-splines, rounding included, and NaN refuses. Any other pair is not glued. Your test moves into solidcore-ops/tests/boolean.rs.

Both PRs close #13650, so the duplicate-PR check flags #16392. Could you close #13576 in favour of #16392? If you want both kept open, please add not-a-duplicate instead.

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

[bherbruck]

Closing in favour of #16392, which carries these commits and finishes the fix (the glue is decided by a proof over the whole span, and there's no flag). Thanks @brett19.