Core data-structure overhaul: interned simplices, facet adjacency, incremental hull (#5)

* refactor: split PyO3 surface out of triangulation.rs into src/py.rs

Pure mechanical move, no logic changes: the pure-Rust Triangulation core
stays in src/triangulation.rs; argument parsing, result conversion,
TriangulationError::into_pyerr, the proxy/iterator views, and the
PyTriangulation class move to the new src/py.rs. Prepares for reworking
the core data structures without PyO3 noise in the diff.

* perf: intern simplices, maintain facet adjacency and hull incrementally

Replace the FxHashSet<Vec<usize>>-everywhere layout (which cloned and
re-hashed whole simplices in every hot path) with interned storage:

- a slab assigns each simplex a small integer SimplexId; the sorted
  vertex list is stored once and referenced by id everywhere else
- vertex_to_simplices becomes vertex -> FxHashSet<SimplexId> (integer
  sets instead of sets of cloned Vecs)
- a new facet index maps every (dim-1)-face to the ids of its incident
  simplices, maintained by the single link/unlink pair of mutation
  primitives, together with the derived boundary-facet set (hull) and
  an overfull-facet count (corruption detector)

This turns the hot paths from set intersections over cloned vectors
into single hash lookups:

- locate_point walk steps and containing() facet queries are one
  facet-map lookup
- the bowyer_watson cascade finds neighbours via dim+1 facet lookups
  instead of unioning and scanning every simplex around all vertices
- extend_hull reads the maintained boundary set instead of recounting
  every face of every simplex per outside-hull insertion (was the
  O(n^2) bottleneck for hull-heavy workloads), and hull() is O(hull)
- reference_invariant() now also cross-checks the facet index,
  boundary set, and overfull count against a recount

Behavior is unchanged: the full pytest cross-validation suite against
the Python reference passes, and stress sweeps (mixed-scale,
off-origin, near-duplicate, random) show the same pass/fail profile
as main. A/B benchmark (best of 3, same machine, incremental
insertion): 2D 5000 pts 0.296s -> 0.143s, 3D 2000 pts 0.540s ->
0.163s, 4D 500 pts 1.959s -> 0.241s, hull-heavy 2D 3000 pts 3.271s ->
1.058s, hull-heavy 3D 1500 pts 3.395s -> 0.698s.

Author: basnijholt

Cargo.lock

@@ -22,6 +22,7 @@ numpy = "0.23"
 nalgebra = "0.33"
 thiserror = "2"
 rustc-hash = "2"
+smallvec = "1"
 
 [profile.release]
 lto = true

Cargo.toml

@@ -32,8 +32,7 @@ fn generate_points(dim: usize, count: usize, seed: u64) -> Vec<Vec<f64>> {
 
 fn first_simplex(triangulation: &Triangulation) -> Simplex {
     triangulation
-        .simplices
-        .iter()
+        .simplices()
         .next()
         .cloned()
         .expect("triangulation should not be empty")
@@ -72,9 +71,7 @@ fn outside_point(triangulation: &Triangulation) -> Vec<f64> {
 }
 
 fn insert_pending_vertex(triangulation: &mut Triangulation, point: Vec<f64>) -> usize {
-    triangulation.vertex_to_simplices.push(Default::default());
-    triangulation.vertices.push(point);
-    triangulation.vertices.len() - 1
+    triangulation.add_vertex(point)
 }
 
 fn constructor_benches(c: &mut Criterion) {
@@ -101,7 +98,7 @@ fn constructor_benches(c: &mut Criterion) {
                         let coords = points.clone();
                         let start = Instant::now();
                         let triangulation = Triangulation::new(coords).unwrap();
-                        black_box(triangulation.simplices.len());
+                        black_box(triangulation.num_simplices());
                         total += start.elapsed();
                     }
                     total
@@ -130,7 +127,7 @@ fn add_point_bench(c: &mut Criterion) {
                 for point in &to_insert {
                     triangulation.add_point(point.clone(), None, None).unwrap();
                 }
-                black_box(triangulation.simplices.len());
+                black_box(triangulation.num_simplices());
                 total += start.elapsed();
             }
             total
@@ -191,7 +188,7 @@ fn locate_and_containing_benches(c: &mut Criterion) {
     group.measurement_time(Duration::from_secs(8));
 
     let triangulation = Triangulation::new(generate_points(2, 2048, 0xFACE)).unwrap();
-    let simplices: Vec<Simplex> = triangulation.simplices.iter().take(256).cloned().collect();
+    let simplices: Vec<Simplex> = triangulation.simplices().take(256).cloned().collect();
     let queries: Vec<Vec<f64>> = simplices
         .iter()
         .map(|simplex| simplex_centroid(&triangulation, simplex))

benches/bench_triangulation.rs

@@ -4,6 +4,7 @@
 //! documented in [`tolerances`].
 
 pub mod geometry;
+pub mod py;
 pub mod tolerances;
 pub mod triangulation;
 
@@ -12,7 +13,7 @@ use pyo3::prelude::*;
 use pyo3::types::{PyAny, PyModule, PyTuple};
 
 use crate::geometry as geom;
-use crate::triangulation::{
+use crate::py::{
     numpy_linalg_error, parse_point, parse_points_sized, point_tuple, PyFacesIter,
     PySimplicesProxy, PyTriangulation, PyVertexToSimplicesIter, PyVertexToSimplicesProxy,
     PyVerticesIter, PyVerticesProxy,