fix(boolean): restrict analytic FF curves + merge coincident junction edges

[andymai]

Summary

Two robustness improvements for coincident curved-surface fuses, surfaced while isolating the curved-loft socket family (fuse_shelled_box_with_socket_loft).

1. algo/phase_ff — restrict analytic FF curves to both faces' extent

compute_raw_curves intersects the unbounded surfaces, so a plane-analytic / algebraic ellipse (e.g. a cylinder vs a tilted wall) can reach far past the faces' shared region and slit the partner face's wire, leaving phantom inner wires. The new restrict_curves_to_faces builds a lightweight per-face extent (planar boundary polygon, or analytic v-range) and drops non-Line curves whose longest contiguous in-both sample run is under two samples — a tangency/point that never splits either face. Conservative: if either extent can't be built, curves pass through unchanged.

2. boolean — merge coincident duplicate junction edges

A coincident-junction fuse annihilates the shared cap but leaves each argument carrying its own copy of the junction-wire edges, differing by sub-micron numerical noise (independent construction). The tight-tolerance vertex merge leaves them distinct → each used once → free edges that open the shell. unify_coincident_boundary_edges snaps vertices at a looser tolerance and rebuilds every wire against a canonical-edge map keyed by unordered endpoints + curve type + geometric midpoint (so a line and an arc between the same endpoints stay distinct); degenerate edges are dropped. Gated on free-edge results so clean booleans keep exact topology.

Effect

Together these take the rrect-cap + frustum reproducer's raw GFA output from euler=−7 / 8 phantom holes / 30+ free edges down to euler=−2 with only the 4 arc-vs-chord corner mismatches left.

Known residual (not addressed here)

The 4 remaining corner mismatches are a loft limitation: loft samples each profile to a polygon and builds straight (Line) ring edges with planar side faces, so a rounded-rect loft becomes an octagonal frustum whose chord-cornered cap can't match a box's arc-cornered cap. Tracked by the new ignored fuse_coincident_rrect_cap_with_frustum test. Closing it needs a curve-preserving loft, not a change in the fuse pipeline.

Testing

Zero regressions: algo 113, operations 662, io 156, wasm 210. Clippy + boundaries clean. 5× flake gate on the boolean suite stable.

⚠️ Review note

phase_ff is core GFA (pave_filler) and runs on every analytic FF pair → HIGH-risk. Not set to auto-merge; requesting human review before merge.

[github-actions]

WASM Binary Size

File	main	PR	Delta
brepkit_wasm.wasm	4.92 MB	4.94 MB	+22.3 KB (+.4%)

⚠️ Size increased slightly.

[Copilot]

Pull request overview

This PR improves robustness of coincident curved-surface boolean fuses by (1) restricting unbounded face-face intersection (FF) curves to the mutual trimmed face extents and (2) merging numerically-near-duplicate junction boundary edges that can otherwise leave free edges and open shells. It also adds an ignored regression test capturing a known loft faceting limitation.

Changes:

• Add FF-curve restriction logic in pave_filler/phase_ff to drop analytic intersection curves that only graze the mutual face extents.
• Add a post-GFA healing pass in operations::boolean to unify coincident duplicate boundary edges when the result shell has free edges.
• Add an ignored integration test reproducer for a coincident rounded-rect cap fuse with a lofted frustum.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 5 comments.

File	Description
crates/algo/src/pave_filler/phase_ff.rs	Adds face-extent construction and FF-curve restriction to avoid untrimmed analytic intersections slitting partner face wires.
crates/operations/src/boolean/mod.rs	Adds a gated healing step to merge coincident duplicate boundary edges on open-shell results.
crates/operations/src/boolean/tests.rs	Adds an ignored reproducer test documenting the remaining loft faceting limitation.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[greptile-apps]

Greptile Summary

Two targeted robustness fixes for coincident curved-surface fuses: phase_ff now discards unbounded analytic intersection curves that only graze the mutual face extent (preventing phantom inner wires), and boolean now merges coincident duplicate junction edges left by sub-micron loft noise (closing free edges). Both paths are conservative and gated so clean booleans are unaffected.

• phase_ff.rs: New FaceExtent abstraction + restrict_curves_to_faces drops non-Line curves whose longest in-both-faces sample run spans fewer than 2 segments, treating them as tangency/grazing points that never split either face.
• boolean/mod.rs: New unify_coincident_boundary_edges, invoked only when has_free_edges fires, rebuilds all wires against a canonical-edge map keyed by endpoint quanta + midpoint + curve type; inner-wire bail-out now matches the outer-wire behaviour (prior review concern addressed).
• tests.rs: Adds an #[ignore] reproducer documenting the residual loft arc-vs-chord corner limitation out of scope for this PR.

Confidence Score: 5/5

Safe to merge after a human second-read of phase_ff — the filter is conservative and all 1141 tests pass with no regressions.

Both new code paths degrade gracefully to the original topology on any failure. The ecanon midpoint key using pre-snap positions is a minor robustness gap well outside the sub-micron noise range targeted here. Orphaned edges on bail-out are an arena-lifetime leak, not a correctness hazard. The previous inner-wire-loss concern is now properly addressed.

phase_ff.rs — core GFA path touched by every analytic FF pair; worth a second human read of FaceExtent::contains and the N=24 sampling threshold.

Important Files Changed

Filename	Overview
crates/algo/src/pave_filler/phase_ff.rs	Adds FaceExtent enum + restrict_curves_to_faces to discard unbounded analytic intersection curves that only graze the mutual face region. Logic is conservative.
crates/operations/src/boolean/mod.rs	Adds unify_coincident_boundary_edges gated on has_free_edges. Two minor P2 points: midpoint key uses original vertex positions; edges before Wire::new bail-out become orphaned.
crates/operations/src/boolean/tests.rs	Adds an #[ignore] regression test documenting the residual loft-faceting limitation. Construction and assertions look correct.

_{Reviews (3): Last reviewed commit: "fix(boolean): bail on inner-wire rebuild..." | Re-trigger Greptile}

[andymai]

Addressed the Copilot review in 872af0f:

• Arc endpoint sampling (phase_ff boundary polygon): now samples 0..=16 so the curve segment's t1 endpoint is included — no artificial closing chord.
• Inner wires / holes: FaceExtent::Plane now builds hole polygons from the face's inner wires and contains() returns false for points strictly inside a hole (beyond the boundary margin), so a curve through a hole is no longer treated as inside the trimmed face.
• Drop-threshold wording: clarified the doc + inline comment — b1 - b0 < 2 keeps runs of ≥2 segments (≥3 in-both samples); the text now says "fewer than two segments" rather than "under two samples".
• changed always-true bug: the unify pass now reuses the original edge when its endpoints didn't move, only allocating (and flagging a change) on a genuine merge / degenerate drop / vertex snap. It's now a true no-op on a free-edge shell that has no coincident duplicates, instead of rebuilding the whole outer shell every time.
• Silently-dropped error: the unify call now logs at debug level on error (still best-effort, doesn't abort the boolean).

Zero regressions (algo 113, operations 662), clippy clean, flake-stable.

[andymai]

Thanks — addressed the Greptile review:

Issue 1 (inner-wire silent loss) — fixed in 286a09b. unify_coincident_boundary_edges now bails with Ok(false) when an inner wire fails to rebuild, consistent with the outer-wire path, so a hole can't be silently dropped (which would change topology and remove free edges, undetectable downstream).

Issue 2 (drop-threshold comment) — already corrected in 872af0f: the doc/comment now say the filter drops runs spanning "fewer than two segments (at most two consecutive in-both samples)", matching b1 - b0 < 2.

Issue 3 (hardcoded 1e-20 in point_to_polygon_dist) — keeping this one as-is. The 1e-20 is a div-by-zero guard for a degenerate (zero-length) polygon segment, not a tolerance decision. Substituting tol.linear² (≈1e-14) would treat legitimately short ~1e-7-length boundary segments as degenerate points and skip them in the nearest-point search, which is less correct for this purpose. The guard intentionally only excludes truly-zero-length segments.

Test job green on the prior commit; CI re-running on 286a09b.