fix(algo): trim plane-plane section curves to mutual face overlap

[andymai]

Plane-plane FF section curves were trimmed to the combined extent of the two faces (min-of-mins/max-of-maxes in trim_t_range_to_aabb) instead of their mutual overlap. The per-face clip ranges were already computed by clip_line_to_face but only used to reject disjoint curves. Over-long chords crossed mid-edge with no vertices at the crossings, so coplanar cuts produced overlapping band sub-faces instead of a proper partition (e.g. coplanar_box_cut_d1a2: volume 0.667 vs 0.75 and zero same-domain pairs).

Changes

• New trim_raw_line shrinks each plane-plane section curve to the intersection of the two faces' clip ranges, recomputing endpoints/bbox/t_range; existing vertex snapping preserved.
• clip_line_to_face builds the face boundary polygon by chaining edges on shared vertex IDs instead of trusting stored is_forward flags (mis-ordered polygons silently truncated clip ranges).
• classify_2d::sample_interior_point rejects centroids lying on the opposing face boundary.
• Un-ignores coplanar_box_cut_d1a2 (correct frame+square partition, 2 SD pairs, volume 0.75) and example_custom_profile_extrusion (3/3).

Remaining coplanar failures (coincident_planes non-manifold cases, thin-shell fuse) trace to a separate defect — collinear split-edge images are not propagated to unsplit neighbor faces — and stay ignored.

Verification

Full corpus green: algo 94, operations 630 lib + 18 integration targets, wasm 206; fmt, clippy -D warnings, boundary check.

[github-actions]

WASM Binary Size

File	main	PR	Delta
brepkit_wasm.wasm	4.71 MB	4.71 MB	-5.4 KB (-.1%)

✅ Size decreased — nice!

[Copilot]

Pull request overview

This PR fixes incorrect trimming of plane–plane face/face (FF) section curves by shrinking Line intersection curves to the mutual overlap of the two faces’ clipped ranges, improving downstream face partitioning and enabling previously-ignored coplanar/operations examples to pass.

Changes:

• Trim plane–plane Line FF curves to the intersection of per-face clip ranges (trim_raw_line), recomputing endpoints/bbox/t-range.
• Make clip_line_to_face build boundary polygons by chaining edges via shared vertex IDs (instead of trusting potentially inconsistent is_forward flags).
• Improve coplanar/same-domain robustness: reject boundary-centroid “interior points”, treat containment-through-holes as non-overlap, and un-ignore two tests now passing.

Reviewed changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
crates/operations/tests/examples.rs	Un-ignores example_custom_profile_extrusion.
crates/operations/src/boolean/tests.rs	Un-ignores coplanar_box_cut_d1a2.
crates/algo/src/pave_filler/tests.rs	Updates box wire orientations; updates expected overlapping-fuse face counts.
crates/algo/src/pave_filler/phase_ff.rs	Implements plane–plane Line trimming to mutual overlap; rebuilds clip polygon via vertex chaining.
crates/algo/src/diagnostic.rs	Adjusts test box wire orientations.
crates/algo/src/classifier/ray_cast.rs	Adjusts test box wire orientations.
crates/algo/src/builder/same_domain.rs	Refines planar overlap logic to exclude containment through holes.
crates/algo/src/builder/classify_2d.rs	Makes interior-point sampling reject centroids that land on polygon boundaries.

---

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

[greptile-apps]

Greptile Summary

Fixes plane-plane face-face section curves being trimmed to the union of the two faces' extents instead of their intersection, which caused over-long chords that split faces incorrectly during coplanar boolean operations. The trimming is now done up-front in phase_ff to the mutual clip-range overlap, and both same_domain and classify_2d gain correctness guards for hole-boundary and boundary-touching centroid cases.

• phase_ff.rs: replaces the old gap-filter with a FaceClip-based mutual-overlap trim (trim_raw_line); adds a polygon-chaining wire builder that avoids relying on stored is_forward flags, and a polygon_is_convex guard to protect Cyrus-Beck clipping from non-convex outlines.
• same_domain.rs: refactors wire_points to accept a WireId, adds in_hole / footprint_in_holes closures to reject containment claims where the candidate sits inside a hole of the container face.
• classify_2d.rs: sample_interior_point now rejects centroids that land on the polygon boundary via an adaptive distance check, preventing unpredictable even-odd results at reflex corners.

Confidence Score: 5/5

The fix is logically sound and all corpus tests pass; the two observations are edge cases (sub-millimetre geometry scale and inner-wire flag inconsistency) unlikely to fire in the existing test suite.

The core trimming logic, the polygon-chaining wire builder, the FaceClip enum, and the hole-containment guards are all correct for standard CAD geometry. The polygon_is_convex absolute threshold and the is_forward reliance in in_hole are edge-case concerns that do not affect currently tested inputs.

No files require blocking attention; the two observations are in phase_ff.rs (polygon_is_convex threshold) and same_domain.rs (wire_points is_forward reliance for inner wires).

Important Files Changed

Filename	Overview
crates/algo/src/pave_filler/phase_ff.rs	Core fix: replaces the old over-long section curve with mutual-overlap trimming. New trim_raw_line, FaceClip enum, polygon-chaining wire builder, and polygon_is_convex helper are all logically correct for normal-scale geometry; the convex-check threshold is absolute (1e-12) and can misfire on sub-millimetre faces.
crates/algo/src/builder/same_domain.rs	Adds in_hole / footprint_in_holes guards for frame-face holes and refactors wire_points to accept a WireId; guards are logically sound but wire_points still uses is_forward for both outer and inner wires, leaving the same potential mis-ordering issue the PR fixed in clip_line_to_face.
crates/algo/src/builder/classify_2d.rs	Adds adaptive boundary-distance guard to sample_interior_point; new helpers boundary_eps and distance_to_polygon_boundary are correct and use a scale-relative epsilon (diag * 1e-6).
crates/algo/src/classifier/ray_cast.rs	Test-only change: corrects oriented-edge direction flags in the box fixture to match the chained-vertex ordering; no production code touched.
crates/algo/src/diagnostic.rs	Test-only: same oriented-edge direction fix as in ray_cast.rs for the diagnostic crate's box fixture.
crates/algo/src/pave_filler/tests.rs	Updates the fuse face-count assertion from (10 or 24) to (10 or 12) to match the trimmed-curve output; narrows the accepted range, which is the correct expectation after the fix.
crates/operations/src/boolean/tests.rs	Un-ignores coplanar_box_cut_d1a2 — straightforward now that the root cause is fixed.
crates/operations/tests/examples.rs	Un-ignores example_custom_profile_extrusion — no logic change, just removes the ignore flag.

Prompt To Fix All With AI

Fix the following 2 code review issues. Work through them one at a time, proposing concise fixes.

---

### Issue 1 of 2
crates/algo/src/pave_filler/phase_ff.rs:1157-1173
**`polygon_is_convex` uses an absolute cross-product threshold that breaks at small scale**

The hardcoded `1e-12` tolerance for the 2D cross product in `polygon_is_convex` is an area-unit quantity (UV length²). For faces whose projected UV coordinates are in the sub-millimetre range (e.g. a 10μm × 10μm planar patch), edge vectors are ~1e-5 long and the cross product for any non-collinear turn is ~1e-10 — below this threshold — so every vertex is treated as collinear and every polygon passes the convex test. Cyrus-Beck clipping then runs on a non-convex outline and can return an over-trimmed interval, producing a section curve that's too short rather than discarded.

### Issue 2 of 2
crates/algo/src/builder/same_domain.rs:465-474
**`in_hole` still relies on `is_forward` for hole-wire vertex ordering**

The new `in_hole` closure calls `wire_points(wid)` for each inner-wire ID. `wire_points` uses `oe.is_forward()` to pick the vertex that starts each oriented edge, which is precisely the ordering path that `clip_line_to_face` was refactored away from in this PR (with the comment "wires from external builders may carry inconsistent `is_forward` flags"). If an inner-wire's oriented edges have inconsistent flags the collected vertices form a self-intersecting polygon, and the even-odd `point_in_polygon_2d` test will give unpredictable results — a point in a hole might be reported as outside it, and a face pair that should be rejected as overlapping-through-hole passes through to the SD list.

_{Reviews (2): Last reviewed commit: "fix(algo): disambiguate empty vs degrade..." | Re-trigger Greptile}