Add the OverlayNG noding substrate (phase 1): exact noded arrangements with symbolic nodes

[asinghvi17]

Phase 1 of the OverlayNG port: the noding substrate — from two geometries to an exact noded arrangement, plus coordinate emission. Stacked on relateng (#416); merge after it. Design doc: docs/plans/2026-07-16-overlayng-noding-substrate.md (first commit on this branch), which also records the survey + spike evidence base (S1–S4) behind every decision here.

Design in one paragraph

Overlay is computed as an exact arrangement with symbolic nodes, rounded once at emission — the constructive extension of RelateNG's design D2. There is no snapping, no snap-rounding, no precision model, and no retry ladder, permanently: that JTS machinery patches an inexact-noding substrate this engine doesn't have. No constructed coordinate ever enters a decision; Float64 appears only inside certified filters (which escalate to the exact kernel predicates on uncertainty) and at final emission. The pipeline keeps JTS OverlayNG's shape (node → graph → label → build) so later phases stay file-by-file diffable against Java.

What's here

• src/methods/clipping/overlayng/noding/ (~600 SLOC, all internal, nothing exported):
  • noded_arrangement.jl — NodedArrangement{P} / NodedEdge: noded edges are (string, segment, node_lo, node_hi) references; a noded edge carries no geometry, so source metadata cannot desynchronize.
  • collect.jl — candidate enumeration via the existing payload RTree + dual_depth_first_search, classification via rk_classify_intersection; proper crossings become symbolic nodes on both parents, touch/collinear incidences become input-vertex nodes (asserted, not assumed). Accepts optional prebuilt trees (a prepared geometry's edge index is the same structure).
  • node_identity.jl — two-tier node identity: egal NodeKey Dict, then a float-sweep with certified radii proposing candidate coincidences confirmed by rk_nodes_coincide, union-find over confirmed pairs only (zero on all real data measured; exercised by degree-6 constructed cases).
  • split.jl — per-segment ordered/deduped interior node lists → noded edges, zero-length guarded.
  • emit.jl — the only lossy step, memoized: planar crossings via a certified double-double fast path (accepted iff provably correctly rounded, Rational{BigInt} fallback); spherical crossings via a conditioning-gated float direction (exact _sph_crossing_dir fallback); vertex nodes pass through bit-exact.
• Kernel substrate (in the relateng kernel files, where manifold predicates live):
  • rk_compare_along_segment — new rk_ predicate ordering nodes along a segment: certified float stage, lazy exact fallback (S1 measured the float stage resolving 100% of comparisons on real data, 240–590× vs exact-always in dense regimes).
  • _ring_material_interior_on_left — the material-interior orientation authority (the hole-side flip exists exactly once, next to _ring_interior_on_left), which phase 2's depth_delta derivation consumes.
  • One unified exact-crossing call shape serving ordering keys, coincidence, and emission.

Validation

• New tests: 7,490 assertions (test/methods/clipping/overlayng/noding.jl) + 422 (test/methods/relateng/kernel_along_segment.jl): the five arrangement invariants on constructed cases (degree-6 coincident node, crossing-on-third-string-vertex, vertex-identical shared borders → zero phantom crossings, a–b–a spikes), float-vs-exact-always ordering cross-check, emission certificate audit (certified result == rational result on every node, >7,000 nodes), and an env-gated NE-110m real-data subset asserting the invariants plus the rounded-arrangement audit (zero rounding-introduced crossings, both manifolds).
• Existing kernel gates re-run at their exact counts (spherical kernel conformance ~51,943; S2 crossings 4,022).
• Performance vs the spike budgets: planar emission 0.026 µs/node (budget ≤0.05); NE-110m single-pair arrangements 0.04–0.24 ms (budget ~1 ms); dense-synthetic full build ~30 ms for 63,500 nodes / 125,500 edges.

Not in this PR (phase 2/3, per the design doc)

Half-edge graph, OverlayLabel/labeller, ring builders (phase 2); the ops, symdifference, mixed dimensions, and the public opt-in OverlayNG{M} algorithm with differential validation vs GEOS and s2 (phase 3). Existing clipping behavior is completely untouched.

🤖 Generated with Claude Code