Add antimeridian_split on the OverlayNG face surface

[asinghvi17]

Adds antimeridian_split — splitting lon/lat polygons at the antimeridian (or any seam, in any rotated frame) — built directly on the OverlayNG exact arrangement. Stacked on #443 (→ #442 → #416 → #434); merge in stack order.

What's here

Face enumeration on the OverlayNG graph (polygon_builder.jl +66, overlay_labeller.jl +11): the non-dissolving companion to the set-op pipeline. _build_faces links every half-edge to its face-ring successor (onext ∘ sym) through the existing next_result field and reuses the unchanged _OverlayEdgeRing/free-hole machinery; _face_ring_location reads a face's per-input location off the shared labels (the face lies on the right of every ring half-edge); _build_face_polygons keeps the faces a keep(loc_a, loc_b) predicate selects. _is_result_of_op gains a predicate-dispatch overload, so custom result semantics can also ride the dissolving pipeline. Pure refactor for existing behavior — the phase-2 engine and graph suites pass unchanged. This is the general surface that later serves cut, polygonize-from-linework, and coverage cleanup (design doc §4 / spike S6).

antimeridian_split(geom; antimeridian = 180.0, north_pole = nothing, pole_spacing = 5.0) (src/transformations/antimeridian_split.jl, ~375 lines with literate docs): cuts a Polygon/MultiPolygon along a pole-to-pole meridian into a MultiPolygon whose pieces each stay within one 360°-wide longitude branch. An encoding repair, not a geometry correction — the spherical region is unchanged, and Polygon → MultiPolygon is why this is a transformation rather than a GeometryCorrection.

• Exact arrangement, no heuristics: the polygon and the 2-segment cutting meridian go through the phase-1 noder and phase-2 labeller; faces with loc_A == INTERIOR are kept. Seam vertices take their longitude branch from the incident cutting-arc half-edge's direction; an enclosed pole (a degree-1 dangle in the arrangement) emits the two-branch pole pair structurally — pole enclosure is proved by the labeller, never by winding heuristics.
• Pole-edge resampling (pole_spacing, default 5°): the pole edge is a zero-length geodesic that pseudocylindrical projections map to a line, so emission densifies it into a strictly monotone constant-latitude row between the (always-present) branch corners. The infill points are the same kernel point — exactly zero spherical area, collinear in plate carrée; nothing emits corners only.
• Arbitrary seam (antimeridian = λ): the cut runs along meridian λ with output longitudes in the closed branch [λn−360, λn] (lips at λn / λn−360). Input coordinates are untouched — only the cutting curve and the output encoding move — so exactness is fully preserved; the default seam is bit-identical to the fixed-±180 form (asserted).
• Rotated pole (north_pole = (λp, φp)): the input is rotated into the CF rotated_latitude_longitude / PROJ +proj=ob_tran frame (north_pole_grid_longitude = 0) and split there; output coordinates are in the rotated frame, deliberately — a cutting curve that dead-ends inside the polygon needs the target frame's dual longitude encoding to emit valid (non-slit) pieces. The rotation is the only coordinate-moving step in the stack.
• Fast path: per polygon part, a spherical open-hemisphere-convexity test on kernel points (an arc with both endpoints strictly on one side of the seam great circle, or both in the far hemisphere, cannot reach the seam) — no false negatives, unlike planar |Δlon| heuristics, which miss near-pole crossing edges; non-interacting parts pass through verbatim.
• Documented composition rule: run antimeridian_split after segmentize (spherical segmentize of an on-meridian edge emits interpolated points whose atan2 longitude is +180 regardless of branch).

Validation

• test/methods/clipping/overlayng/faces.jl (96 assertions): predicate-vs-enum equivalence on all 4 ops × 16 location pairs, offset-squares face census (A∩B/A\B/B\A/outer with labels, orientations, areas), polygon-with-hole faces, dangle doubling, the one-extraction-per-graph contract.
• test/transformations/antimeridian_split.jl (2,604 assertions): seam-crossing box (exact lip vertex sets both branches); south/north caps with 73-point pole rows (strictly monotone, step ≤ 5°, corners exactly (±180, ±90)); corners-only mode with bit-equal areas; vertex-at-pole-without-enclosure (no row); touch-without-crossing; fast-path verbatim passthrough; seam-crossing and one-sided holes; arbitrary-seam cases (Greenwich, Pacific 330°≡−30°, default-vs-explicit-180 bit-identity); rotated-pole cap with a rotated-frame pole row and isometry-conserved area (measured ~2e-15, asserted 1e-12); Python antimeridian package fixtures; @testset_implementations across GeoInterface/LibGEOS/ArchGDAL; unsupported-trait errors.
• Every emitted piece, in every case: LibGEOS-valid, no |Δlon| > 180° between consecutive vertices anywhere (no pole exemption), longitudes within the documented branch, spherical area conservation at rtol 1e-12.
• Natural Earth 110m (data-availability-gated, 2,069 assertions): Russia → 14 pieces, Fiji → 3, Antarctica → 8 with the pole row, conservation ≤ 1e-13, all valid; rotated-pole Russia valid + conserved. (NE Antarctica is deliberately not rotated: its antimeridian-slit encoding is only OGC-valid because the slit sits on ±180 — a finding the tests document.)
• Phase-2 regression: overlay_ng.jl (154) and overlay_graph.jl (37,279) unchanged after the face-enumeration refactor.

Test plan

• CI green on the tip a46d56570 (full package suite + docs, 5 Julia matrices) — run 29543622992, all 7 jobs

🤖 Generated with Claude Code