SAH BVH, degenerate mesh filtering, robust signed distance + resolve cordic merge

crates/cord-cordic/src/compiler.rs

@@ -1,6 +1,7 @@
 use std::collections::BTreeMap;
 use cord_trig::{TrigGraph, TrigOp};
 use crate::config::{CordicConfig, CordicOp};
+use crate::lut::CordicTable;
 use crate::ops::*;
 
 /// A compiled CORDIC program ready for binary serialization or execution.
@@ -17,7 +18,7 @@ pub struct CORDICProgram {
     pub word_bits: u8,
     pub instructions: Vec<CORDICInstr>,
     pub output: u32,
-    /// One entry per unique iteration count.
+    /// One entry per unique iteration count (Arc-deduplicated via LutRegistry).
     pub tables: Vec<CordicTable>,
     /// Maps instruction index → table index. Only meaningful for CORDIC ops;
     /// non-CORDIC instructions have 0 here (ignored at eval time).
@@ -27,15 +28,6 @@ pub struct CORDICProgram {
     pub gain: i64,
 }
 
-/// Precomputed atan table + gain for a specific iteration count.
-#[derive(Debug, Clone)]
-pub struct CordicTable {
-    pub iterations: u8,
-    pub atan_table: Vec<i64>,
-    pub gain: i64,
-}
-
-#[derive(Debug, Clone)]
 pub struct CompileConfig {
     pub word_bits: u8,
     pub cordic: Option<CordicConfig>,
@@ -68,30 +60,21 @@ impl CORDICProgram {
                 c
             });
 
-        // Build tables for each unique iteration count
+        // Build tables for each unique iteration count, using LutRegistry for dedup
         let unique_counts = cordic_cfg.unique_iteration_counts();
         let mut count_to_table_idx: BTreeMap<u8, u8> = BTreeMap::new();
         let mut tables = Vec::new();
         for &iters in &unique_counts {
             count_to_table_idx.insert(iters, tables.len() as u8);
-            tables.push(CordicTable {
+            tables.push(CordicTable::get(iters, frac_bits));
-                iterations: iters,
-                atan_table: atan_table(iters),
-                gain: cordic_gain(iters, frac_bits),
-            });
         }
 
         let default_table_idx = count_to_table_idx
             .get(&cordic_cfg.default_f)
             .copied()
             .unwrap_or_else(|| {
-                // Ensure default has a table entry
                 let idx = tables.len() as u8;
-                tables.push(CordicTable {
+                tables.push(CordicTable::get(cordic_cfg.default_f, frac_bits));
-                    iterations: cordic_cfg.default_f,
-                    atan_table: atan_table(cordic_cfg.default_f),
-                    gain: cordic_gain(cordic_cfg.default_f, frac_bits),
-                });
                 idx
             });
 
@@ -163,12 +146,8 @@ impl CORDICProgram {
             }
         }).collect();
 
-        // Legacy single-table fields use the first table (or default)
+        let primary = tables.first().cloned()
-        let primary = tables.first().cloned().unwrap_or_else(|| CordicTable {
+            .unwrap_or_else(|| CordicTable::for_word_bits(config.word_bits));
-            iterations: config.word_bits,
-            atan_table: atan_table(config.word_bits),
-            gain: cordic_gain(config.word_bits, frac_bits),
-        });
 
         CORDICProgram {
             word_bits: config.word_bits,
@@ -176,7 +155,7 @@ impl CORDICProgram {
             output: graph.output,
             tables,
             instr_table_idx,
-            atan_table: primary.atan_table,
+            atan_table: primary.atan.to_vec(),
             gain: primary.gain,
         }
     }
@@ -248,11 +227,7 @@ impl CORDICProgram {
             pos += consumed;
         }
 
-        let tables = vec![CordicTable {
+        let tables = vec![CordicTable::get(word_bits, word_bits - 1)];
-            iterations: word_bits,
-            atan_table: atan_table.clone(),
-            gain,
-        }];
         let instr_table_idx = vec![0u8; instructions.len()];
 
         Some(CORDICProgram {

crates/cord-cordic/src/eval.rs

@@ -1,12 +1,12 @@
 use cord_trig::{TrigGraph, TrigOp};
-use crate::compiler::{CORDICProgram, CordicTable};
+use crate::compiler::CORDICProgram;
-use crate::ops::{atan_table, cordic_gain};
+use crate::lut::CordicTable;
 
 /// CORDIC evaluator: evaluates a TrigGraph using only integer
 /// shifts, adds, and comparisons. No floating point trig.
 ///
 /// Proof that the entire pipeline compiles down to
-/// binary arithmetic — shift, add, compare, repeat.
+/// binary arithmetic -- shift, add, compare, repeat.
 pub struct CORDICEvaluator {
     frac_bits: u8,
     instr_table_idx: Vec<u8>,
@@ -19,11 +19,7 @@ impl CORDICEvaluator {
         CORDICEvaluator {
             frac_bits,
             instr_table_idx: Vec::new(),
-            tables: vec![CordicTable {
+            tables: vec![CordicTable::for_word_bits(word_bits)],
-                iterations: word_bits,
-                atan_table: atan_table(word_bits),
-                gain: cordic_gain(word_bits, frac_bits),
-            }],
         }
     }
 
@@ -76,8 +72,6 @@ impl CORDICEvaluator {
     /// Resolve the table for a given instruction index.
     fn table_for(&self, instr_idx: usize) -> &CordicTable {
         if self.instr_table_idx.is_empty() || self.tables.is_empty() {
-            // Fallback: use legacy single table via a static-lifetime trick
-            // won't work — build a table reference from self fields instead
             return &self.tables[0];
         }
         let tidx = self.instr_table_idx.get(instr_idx).copied().unwrap_or(0) as usize;
@@ -94,7 +88,7 @@ impl CORDICEvaluator {
             let d = if z >= 0 { 1i64 } else { -1 };
             let x_new = x - d * (y >> i);
             let y_new = y + d * (x >> i);
-            z -= d * table.atan_table[i];
+            z -= d * table.atan[i];
             x = x_new;
             y = y_new;
         }
@@ -118,7 +112,7 @@ impl CORDICEvaluator {
             let d = if y < 0 { 1i64 } else { -1 };
             let x_new = x - d * (y >> i);
             let y_new = y + d * (x >> i);
-            z -= d * table.atan_table[i];
+            z -= d * table.atan[i];
             x = x_new;
             y = y_new;
         }

crates/cord-cordic/src/lib.rs

@@ -11,7 +11,9 @@ pub mod compiler;
 pub mod config;
 pub mod ops;
 pub mod eval;
+pub mod lut;
 
 pub use compiler::CORDICProgram;
 pub use config::CordicConfig;
 pub use eval::CORDICEvaluator;
+pub use lut::CordicTable;

crates/cord-cordic/src/lut.rs

@@ -0,0 +1,114 @@
+use std::collections::HashMap;
+use std::sync::{Arc, Mutex, OnceLock};
+
+/// Pre-computed CORDIC table: atan values + gain for a given iteration count.
+#[derive(Debug, Clone)]
+pub struct CordicTable {
+    pub atan: Arc<[i64]>,
+    pub gain: i64,
+    pub iterations: u8,
+    pub frac_bits: u8,
+}
+
+/// Global table registry. Keyed by (iterations, frac_bits) so any
+/// combination of word width and iteration count shares a single
+/// allocation. Thread-safe via interior Mutex on the map only;
+/// once an Arc is cloned out, no lock is held during evaluation.
+struct LutRegistry {
+    tables: Mutex<HashMap<(u8, u8), CordicTable>>,
+}
+
+static REGISTRY: OnceLock<LutRegistry> = OnceLock::new();
+
+fn registry() -> &'static LutRegistry {
+    REGISTRY.get_or_init(|| LutRegistry {
+        tables: Mutex::new(HashMap::new()),
+    })
+}
+
+impl CordicTable {
+    /// Retrieve or compute the CORDIC table for the given parameters.
+    /// Repeated calls with the same (iterations, frac_bits) return
+    /// a clone of the same Arc'd data -- no recomputation.
+    pub fn get(iterations: u8, frac_bits: u8) -> Self {
+        let reg = registry();
+        let key = (iterations, frac_bits);
+
+        let mut map = reg.tables.lock().unwrap();
+        if let Some(cached) = map.get(&key) {
+            return cached.clone();
+        }
+
+        let table = Self::compute(iterations, frac_bits);
+        map.insert(key, table.clone());
+        table
+    }
+
+    /// Convenience: derive iterations and frac_bits from word_bits
+    /// using the standard CORDIC convention (iterations = word_bits,
+    /// frac_bits = word_bits - 1).
+    pub fn for_word_bits(word_bits: u8) -> Self {
+        Self::get(word_bits, word_bits - 1)
+    }
+
+    fn compute(iterations: u8, frac_bits: u8) -> Self {
+        let scale = (1i64 << frac_bits) as f64;
+
+        let atan: Arc<[i64]> = (0..iterations)
+            .map(|i| {
+                let angle = (2.0f64).powi(-(i as i32)).atan();
+                (angle * scale).round() as i64
+            })
+            .collect::<Vec<_>>()
+            .into();
+
+        let mut k = 1.0f64;
+        for i in 0..iterations {
+            k *= 1.0 / (1.0 + (2.0f64).powi(-2 * i as i32)).sqrt();
+        }
+        let gain = (k * scale).round() as i64;
+
+        CordicTable { atan, gain, iterations, frac_bits }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::ops;
+
+    #[test]
+    fn cached_matches_raw() {
+        for bits in [8u8, 16, 32, 48, 64] {
+            let table = CordicTable::for_word_bits(bits);
+            let raw_atan = ops::atan_table(bits);
+            let raw_gain = ops::cordic_gain(bits, bits - 1);
+
+            assert_eq!(table.atan.len(), raw_atan.len(), "len mismatch at {bits}");
+            assert_eq!(&*table.atan, &raw_atan[..], "atan mismatch at {bits}");
+            assert_eq!(table.gain, raw_gain, "gain mismatch at {bits}");
+        }
+    }
+
+    #[test]
+    fn deduplication() {
+        let a = CordicTable::for_word_bits(32);
+        let b = CordicTable::for_word_bits(32);
+        assert!(Arc::ptr_eq(&a.atan, &b.atan));
+    }
+
+    #[test]
+    fn different_widths_distinct() {
+        let a = CordicTable::for_word_bits(16);
+        let b = CordicTable::for_word_bits(32);
+        assert!(!Arc::ptr_eq(&a.atan, &b.atan));
+        assert_ne!(a.atan.len(), b.atan.len());
+    }
+
+    #[test]
+    fn custom_iteration_count() {
+        let t = CordicTable::get(16, 31);
+        assert_eq!(t.atan.len(), 16);
+        assert_eq!(t.frac_bits, 31);
+    }
+}

crates/cord-decompile/src/bvh.rs

@@ -39,12 +39,54 @@ impl BVH {
     }
 
     /// Find the signed distance from point p to the mesh.
-    /// Sign is determined by the normal of the nearest triangle.
+    ///
+    /// Sign is determined by angle-weighted pseudonormal voting across all
+    /// triangles within a tolerance of the nearest distance. This handles
+    /// edge and vertex proximity correctly, where a single triangle's face
+    /// normal would give ambiguous or wrong sign.
     pub fn signed_distance(&self, mesh: &TriangleMesh, p: Vec3) -> f64 {
         let mut best_dist_sq = f64::INFINITY;
-        let mut best_sign = 1.0f64;
+        self.query_nearest_dist(mesh, p, 0, &mut best_dist_sq);
-        self.query_nearest(mesh, p, 0, &mut best_dist_sq, &mut best_sign);
+
-        best_dist_sq.sqrt() * best_sign
+        if best_dist_sq == f64::INFINITY {
+            return f64::INFINITY;
+        }
+
+        // Collect all triangles within a relative tolerance of the best distance.
+        // This catches co-nearest triangles sharing an edge or vertex.
+        let eps_factor = 1.0 + 1e-4;
+        let threshold_sq = best_dist_sq * eps_factor * eps_factor + 1e-20;
+        let mut neighbors: Vec<(usize, Vec3)> = Vec::new(); // (tri_idx, closest_point)
+        self.collect_near(mesh, p, 0, threshold_sq, &mut neighbors);
+
+        let sign = if neighbors.len() <= 1 {
+            // Single closest triangle: use its face normal directly
+            if let Some(&(tri_idx, closest)) = neighbors.first() {
+                let to_point = p - closest;
+                let normal = mesh.triangles[tri_idx].normal();
+                if to_point.dot(normal) >= 0.0 { 1.0 } else { -1.0 }
+            } else {
+                1.0
+            }
+        } else {
+            // Multiple co-nearest triangles: angle-weighted normal vote.
+            // Weight each triangle's contribution by the solid angle it subtends
+            // at p, approximated by the triangle area divided by distance squared.
+            let mut weighted_dot = 0.0f64;
+            for &(tri_idx, closest) in &neighbors {
+                let tri = &mesh.triangles[tri_idx];
+                let to_point = p - closest;
+                let normal = tri.normal();
+                let area = tri.area();
+                let d_sq = to_point.dot(to_point);
+                // Weight: area / (distance^2 + epsilon) — larger, closer faces dominate
+                let weight = area / (d_sq + 1e-20);
+                weighted_dot += to_point.dot(normal) * weight;
+            }
+            if weighted_dot >= 0.0 { 1.0 } else { -1.0 }
+        };
+
+        best_dist_sq.sqrt() * sign
     }
 
     /// Find the nearest triangle index and unsigned distance.
@@ -55,13 +97,13 @@ impl BVH {
         (best_idx, best_dist_sq.sqrt())
     }
 
-    fn query_nearest(
+    /// Find the minimum squared distance from p to any triangle.
+    fn query_nearest_dist(
         &self,
         mesh: &TriangleMesh,
         p: Vec3,
         node_idx: usize,
         best_dist_sq: &mut f64,
-        best_sign: &mut f64,
     ) {
         match &self.nodes[node_idx] {
             BVHNode::Leaf { bounds, first, count } => {
@@ -71,13 +113,10 @@ impl BVH {
                 for i in *first..(*first + *count) {
                     let tri_idx = self.tri_indices[i];
                     let tri = &mesh.triangles[tri_idx];
-                    let (closest, dist) = tri.closest_point(p);
+                    let (_, dist) = tri.closest_point(p);
                     let dist_sq = dist * dist;
                     if dist_sq < *best_dist_sq {
                         *best_dist_sq = dist_sq;
-                        let to_point = p - closest;
-                        let normal = tri.normal();
-                        *best_sign = if to_point.dot(normal) >= 0.0 { 1.0 } else { -1.0 };
                     }
                 }
             }
@@ -91,16 +130,49 @@ impl BVH {
                 let dr = right_bounds.distance_to_point(p);
 
                 if dl < dr {
-                    self.query_nearest(mesh, p, *left, best_dist_sq, best_sign);
+                    self.query_nearest_dist(mesh, p, *left, best_dist_sq);
-                    self.query_nearest(mesh, p, *right, best_dist_sq, best_sign);
+                    self.query_nearest_dist(mesh, p, *right, best_dist_sq);
                 } else {
-                    self.query_nearest(mesh, p, *right, best_dist_sq, best_sign);
+                    self.query_nearest_dist(mesh, p, *right, best_dist_sq);
-                    self.query_nearest(mesh, p, *left, best_dist_sq, best_sign);
+                    self.query_nearest_dist(mesh, p, *left, best_dist_sq);
                 }
             }
         }
     }
 
+    /// Collect all (triangle_index, closest_point) pairs within threshold_sq of p.
+    fn collect_near(
+        &self,
+        mesh: &TriangleMesh,
+        p: Vec3,
+        node_idx: usize,
+        threshold_sq: f64,
+        out: &mut Vec<(usize, Vec3)>,
+    ) {
+        match &self.nodes[node_idx] {
+            BVHNode::Leaf { bounds, first, count } => {
+                if bounds.distance_to_point(p).powi(2) > threshold_sq {
+                    return;
+                }
+                for i in *first..(*first + *count) {
+                    let tri_idx = self.tri_indices[i];
+                    let tri = &mesh.triangles[tri_idx];
+                    let (closest, dist) = tri.closest_point(p);
+                    if dist * dist <= threshold_sq {
+                        out.push((tri_idx, closest));
+                    }
+                }
+            }
+            BVHNode::Internal { bounds, left, right } => {
+                if bounds.distance_to_point(p).powi(2) > threshold_sq {
+                    return;
+                }
+                self.collect_near(mesh, p, *left, threshold_sq, out);
+                self.collect_near(mesh, p, *right, threshold_sq, out);
+            }
+        }
+    }
+
     fn query_nearest_idx(
         &self,
         mesh: &TriangleMesh,
@@ -195,6 +267,120 @@ fn point_in_aabb(p: Vec3, b: &AABB) -> bool {
 }
 
 const MAX_LEAF_SIZE: usize = 8;
+const SAH_NUM_BINS: usize = 16;
+const SAH_TRAVERSAL_COST: f64 = 1.0;
+const SAH_INTERSECT_COST: f64 = 1.5;
+
+/// Surface area of an AABB (used as probability proxy in SAH).
+fn aabb_surface_area(b: &AABB) -> f64 {
+    let e = b.extent();
+    2.0 * (e.x * e.y + e.y * e.z + e.z * e.x)
+}
+
+/// SAH bin for accumulating triangle bounds during split evaluation.
+struct SAHBin {
+    bounds: AABB,
+    count: usize,
+}
+
+impl SAHBin {
+    fn empty() -> Self {
+        Self { bounds: AABB::empty(), count: 0 }
+    }
+}
+
+/// Evaluate SAH cost for splitting at each bin boundary along the given axis.
+/// Returns (best_split_bin, best_cost). Split puts bins [0..split] left, [split..N] right.
+fn sah_best_split(
+    triangles: &[Triangle],
+    indices: &[usize],
+    centroids: &[Vec3],
+    start: usize,
+    end: usize,
+    parent_bounds: &AABB,
+    axis: usize,
+) -> (usize, f64) {
+    let get_axis = |v: Vec3| match axis {
+        0 => v.x,
+        1 => v.y,
+        _ => v.z,
+    };
+
+    // Centroid bounds (tighter than triangle bounds) for bin mapping
+    let mut centroid_min = f64::INFINITY;
+    let mut centroid_max = f64::NEG_INFINITY;
+    for &idx in &indices[start..end] {
+        let c = get_axis(centroids[idx]);
+        centroid_min = centroid_min.min(c);
+        centroid_max = centroid_max.max(c);
+    }
+
+    let span = centroid_max - centroid_min;
+    if span < 1e-12 {
+        // All centroids coincide on this axis; can't split here
+        return (SAH_NUM_BINS / 2, f64::INFINITY);
+    }
+    let inv_span = SAH_NUM_BINS as f64 / span;
+
+    // Bin triangles
+    let mut bins: Vec<SAHBin> = (0..SAH_NUM_BINS).map(|_| SAHBin::empty()).collect();
+
+    for &idx in &indices[start..end] {
+        let c = get_axis(centroids[idx]);
+        let b = ((c - centroid_min) * inv_span) as usize;
+        let b = b.min(SAH_NUM_BINS - 1);
+        bins[b].bounds = bins[b].bounds.union(&AABB::from_triangle(&triangles[idx]));
+        bins[b].count += 1;
+    }
+
+    // Sweep from left: accumulate bounds and counts for splits [0..i+1] | [i+1..N]
+    let mut left_area = vec![0.0f64; SAH_NUM_BINS - 1];
+    let mut left_count = vec![0usize; SAH_NUM_BINS - 1];
+    let mut running_bounds = AABB::empty();
+    let mut running_count = 0usize;
+
+    for i in 0..(SAH_NUM_BINS - 1) {
+        running_bounds = running_bounds.union(&bins[i].bounds);
+        running_count += bins[i].count;
+        left_area[i] = aabb_surface_area(&running_bounds);
+        left_count[i] = running_count;
+    }
+
+    // Sweep from right
+    let mut right_area = vec![0.0f64; SAH_NUM_BINS - 1];
+    let mut right_count = vec![0usize; SAH_NUM_BINS - 1];
+    running_bounds = AABB::empty();
+    running_count = 0;
+
+    for i in (0..(SAH_NUM_BINS - 1)).rev() {
+        running_bounds = running_bounds.union(&bins[i + 1].bounds);
+        running_count += bins[i + 1].count;
+        right_area[i] = aabb_surface_area(&running_bounds);
+        right_count[i] = running_count;
+    }
+
+    let parent_area = aabb_surface_area(parent_bounds);
+    let inv_parent = if parent_area > 1e-12 { 1.0 / parent_area } else { 1.0 };
+
+    let mut best_cost = f64::INFINITY;
+    let mut best_bin = SAH_NUM_BINS / 2;
+
+    for i in 0..(SAH_NUM_BINS - 1) {
+        if left_count[i] == 0 || right_count[i] == 0 {
+            continue;
+        }
+        let cost = SAH_TRAVERSAL_COST
+            + SAH_INTERSECT_COST * inv_parent
+                * (left_area[i] * left_count[i] as f64
+                    + right_area[i] * right_count[i] as f64);
+        if cost < best_cost {
+            best_cost = cost;
+            best_bin = i;
+        }
+    }
+
+    (best_bin, best_cost)
+}
 
 fn build_recursive(
     triangles: &[Triangle],
@@ -206,7 +392,6 @@ fn build_recursive(
 ) -> usize {
     let count = end - start;
 
-    // Compute bounds
     let mut bounds = AABB::empty();
     for &idx in &indices[start..end] {
         bounds = bounds.union(&AABB::from_triangle(&triangles[idx]));
@@ -218,24 +403,82 @@ fn build_recursive(
         return node_idx;
     }
 
-    // Split along longest axis at centroid median
+    // SAH split: evaluate all three axes, pick lowest cost
-    let axis = bounds.longest_axis();
+    let leaf_cost = SAH_INTERSECT_COST * count as f64;
-    let mid = start + count / 2;
+    let mut best_axis = 0usize;
+    let mut best_bin = SAH_NUM_BINS / 2;
+    let mut best_cost = f64::INFINITY;
 
-    // Partial sort: partition around the median centroid on the chosen axis
+    for axis in 0..3 {
+        let (bin, cost) = sah_best_split(
+            triangles, indices, centroids, start, end, &bounds, axis,
+        );
+        if cost < best_cost {
+            best_cost = cost;
+            best_bin = bin;
+            best_axis = axis;
+        }
+    }
+
+    // If SAH says a leaf is cheaper, or no valid split found, fall back to median
+    let use_median = best_cost >= leaf_cost || best_cost == f64::INFINITY;
+
+    let mid = if use_median {
+        // Median split along longest axis
+        let axis = bounds.longest_axis();
         let get_axis = |v: Vec3| match axis {
             0 => v.x,
             1 => v.y,
             _ => v.z,
         };
-
+        indices[start..end].sort_unstable_by(|&a, &b| {
-    // Simple partition: sort the range by centroid along axis
+            get_axis(centroids[a])
-    let index_slice = &mut indices[start..end];
+                .partial_cmp(&get_axis(centroids[b]))
-    index_slice.sort_unstable_by(|&a, &b| {
+                .unwrap_or(std::cmp::Ordering::Equal)
-        let ca = get_axis(centroids[a]);
-        let cb = get_axis(centroids[b]);
-        ca.partial_cmp(&cb).unwrap_or(std::cmp::Ordering::Equal)
         });
+        start + count / 2
+    } else {
+        // Partition around the SAH split bin
+        let get_axis = |v: Vec3| match best_axis {
+            0 => v.x,
+            1 => v.y,
+            _ => v.z,
+        };
+
+        let mut centroid_min = f64::INFINITY;
+        let mut centroid_max = f64::NEG_INFINITY;
+        for &idx in &indices[start..end] {
+            let c = get_axis(centroids[idx]);
+            centroid_min = centroid_min.min(c);
+            centroid_max = centroid_max.max(c);
+        }
+        let span = centroid_max - centroid_min;
+        let inv_span = SAH_NUM_BINS as f64 / span;
+
+        // Partition: triangles in bins [0..=best_bin] go left
+        let slice = &mut indices[start..end];
+        let mut lo = 0usize;
+        let mut hi = slice.len();
+        while lo < hi {
+            let c = get_axis(centroids[slice[lo]]);
+            let b = ((c - centroid_min) * inv_span) as usize;
+            let b = b.min(SAH_NUM_BINS - 1);
+            if b <= best_bin {
+                lo += 1;
+            } else {
+                hi -= 1;
+                slice.swap(lo, hi);
+            }
+        }
+
+        let mid = start + lo;
+        // Guard: if partition degenerates, fall back to midpoint
+        if mid == start || mid == end {
+            start + count / 2
+        } else {
+            mid
+        }
+    };
 
     let node_idx = nodes.len();
     nodes.push(BVHNode::Leaf { bounds: AABB::empty(), first: 0, count: 0 }); // placeholder
@@ -246,3 +489,93 @@ fn build_recursive(
     nodes[node_idx] = BVHNode::Internal { bounds, left, right };
     node_idx
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn make_unit_cube_mesh() -> TriangleMesh {
+        // 12 triangles forming a unit cube centered at origin
+        let h = 0.5;
+        let corners = [
+            Vec3::new(-h, -h, -h), Vec3::new( h, -h, -h),
+            Vec3::new( h,  h, -h), Vec3::new(-h,  h, -h),
+            Vec3::new(-h, -h,  h), Vec3::new( h, -h,  h),
+            Vec3::new( h,  h,  h), Vec3::new(-h,  h,  h),
+        ];
+        // CCW winding viewed from outside each face
+        let faces: [(usize, usize, usize); 12] = [
+            (0,2,1), (0,3,2), // -Z face: normal points -Z
+            (4,5,6), (4,6,7), // +Z face: normal points +Z
+            (0,1,5), (0,5,4), // -Y face: normal points -Y
+            (2,3,7), (2,7,6), // +Y face: normal points +Y
+            (0,4,7), (0,7,3), // -X face: normal points -X
+            (1,2,6), (1,6,5), // +X face: normal points +X
+        ];
+        let mut triangles = Vec::new();
+        let mut bounds = AABB::empty();
+        for &(a, b, c) in &faces {
+            let tri = Triangle { v: [corners[a], corners[b], corners[c]] };
+            for v in &tri.v {
+                bounds.min = bounds.min.component_min(*v);
+                bounds.max = bounds.max.component_max(*v);
+            }
+            triangles.push(tri);
+        }
+        TriangleMesh { triangles, bounds }
+    }
+
+    #[test]
+    fn bvh_builds_for_cube() {
+        let mesh = make_unit_cube_mesh();
+        let bvh = BVH::build(&mesh);
+        assert!(!bvh.nodes.is_empty());
+    }
+
+    #[test]
+    fn signed_distance_inside_cube() {
+        let mesh = make_unit_cube_mesh();
+        let bvh = BVH::build(&mesh);
+        let d = bvh.signed_distance(&mesh, Vec3::new(0.0, 0.0, 0.0));
+        assert!(d < 0.0, "origin should be inside cube, got {d}");
+    }
+
+    #[test]
+    fn signed_distance_outside_cube() {
+        let mesh = make_unit_cube_mesh();
+        let bvh = BVH::build(&mesh);
+        let d = bvh.signed_distance(&mesh, Vec3::new(2.0, 0.0, 0.0));
+        assert!(d > 0.0, "point far from cube should be outside, got {d}");
+    }
+
+    #[test]
+    fn signed_distance_on_edge_correct_sign() {
+        // Point near a cube edge: the multi-triangle voting should give
+        // consistent sign even though two face normals are involved.
+        let mesh = make_unit_cube_mesh();
+        let bvh = BVH::build(&mesh);
+        // Slightly outside the cube near the +X/+Y edge
+        let d = bvh.signed_distance(&mesh, Vec3::new(0.6, 0.6, 0.0));
+        assert!(d > 0.0, "point outside near edge should be positive, got {d}");
+        // Slightly inside near the same edge
+        let d2 = bvh.signed_distance(&mesh, Vec3::new(0.4, 0.4, 0.0));
+        assert!(d2 < 0.0, "point inside near edge should be negative, got {d2}");
+    }
+
+    #[test]
+    fn signed_distance_on_vertex_correct_sign() {
+        // Point near a cube vertex: three face normals are involved.
+        let mesh = make_unit_cube_mesh();
+        let bvh = BVH::build(&mesh);
+        let d = bvh.signed_distance(&mesh, Vec3::new(0.7, 0.7, 0.7));
+        assert!(d > 0.0, "point outside near vertex should be positive, got {d}");
+    }
+
+    #[test]
+    fn nearest_triangle_finds_closest() {
+        let mesh = make_unit_cube_mesh();
+        let bvh = BVH::build(&mesh);
+        let (_, dist) = bvh.nearest_triangle(&mesh, Vec3::new(1.5, 0.0, 0.0));
+        assert!((dist - 1.0).abs() < 0.01, "distance to +X face should be ~1.0, got {dist}");
+    }
+}

crates/cord-decompile/src/mesh.rs

@@ -237,6 +237,53 @@ pub struct TriangleMesh {
 }
 
 impl TriangleMesh {
+    /// Remove degenerate triangles: zero/near-zero area, NaN/Inf vertices,
+    /// and collapsed edges. Threshold is relative to the mesh diagonal.
+    fn sanitize(&mut self) {
+        if self.triangles.is_empty() {
+            return;
+        }
+        let diag = self.bounds.diagonal();
+        // Minimum area: square of 1e-7 times diagonal — catches slivers from
+        // CAD exports without rejecting intentionally thin geometry.
+        let min_area = (diag * 1e-7).powi(2) * 0.5;
+        // Minimum edge length squared
+        let min_edge_sq = (diag * 1e-8).powi(2);
+
+        self.triangles.retain(|tri| {
+            // Reject NaN/Inf vertices
+            for v in &tri.v {
+                if !v.x.is_finite() || !v.y.is_finite() || !v.z.is_finite() {
+                    return false;
+                }
+            }
+            // Reject degenerate area
+            if tri.area() < min_area {
+                return false;
+            }
+            // Reject collapsed edges (two vertices coincident)
+            let e0 = tri.v[1] - tri.v[0];
+            let e1 = tri.v[2] - tri.v[1];
+            let e2 = tri.v[0] - tri.v[2];
+            if e0.dot(e0) < min_edge_sq
+                || e1.dot(e1) < min_edge_sq
+                || e2.dot(e2) < min_edge_sq
+            {
+                return false;
+            }
+            true
+        });
+
+        // Recompute bounds after filtering
+        self.bounds = AABB::empty();
+        for tri in &self.triangles {
+            for v in &tri.v {
+                self.bounds.min = self.bounds.min.component_min(*v);
+                self.bounds.max = self.bounds.max.component_max(*v);
+            }
+        }
+    }
+
     pub fn from_stl(path: &Path) -> Result<Self> {
         let data = std::fs::read(path)
             .with_context(|| format!("reading {}", path.display()))?;
@@ -389,12 +436,102 @@ impl TriangleMesh {
         let ext = path.extension()
             .and_then(|e| e.to_str())
             .unwrap_or("");
-        match ext.to_ascii_lowercase().as_str() {
+        let mut mesh = match ext.to_ascii_lowercase().as_str() {
             "stl" => Self::from_stl(path),
             "obj" => Self::from_obj(path),
             "3mf" => Self::from_3mf(path),
             _ if ext.is_empty() => anyhow::bail!("no file extension"),
             _ => anyhow::bail!("unsupported mesh format: .{ext}"),
-        }
+        }?;
+        mesh.sanitize();
+        anyhow::ensure!(!mesh.triangles.is_empty(), "mesh contains no valid triangles after sanitization");
+        Ok(mesh)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn sanitize_removes_degenerate_triangles() {
+        let good = Triangle {
+            v: [Vec3::new(0.0, 0.0, 0.0), Vec3::new(1.0, 0.0, 0.0), Vec3::new(0.0, 1.0, 0.0)],
+        };
+        // Zero-area triangle: all three vertices collinear
+        let degenerate = Triangle {
+            v: [Vec3::new(0.0, 0.0, 0.0), Vec3::new(1.0, 0.0, 0.0), Vec3::new(2.0, 0.0, 0.0)],
+        };
+        // Collapsed edge: two vertices coincident
+        let collapsed = Triangle {
+            v: [Vec3::new(0.0, 0.0, 0.0), Vec3::new(0.0, 0.0, 0.0), Vec3::new(0.0, 1.0, 0.0)],
+        };
+        let mut mesh = TriangleMesh {
+            triangles: vec![good, degenerate, collapsed],
+            bounds: AABB {
+                min: Vec3::new(0.0, 0.0, 0.0),
+                max: Vec3::new(2.0, 1.0, 0.0),
+            },
+        };
+        mesh.sanitize();
+        assert_eq!(mesh.triangles.len(), 1, "only the good triangle should survive");
+    }
+
+    #[test]
+    fn sanitize_removes_nan_vertices() {
+        let good = Triangle {
+            v: [Vec3::new(0.0, 0.0, 0.0), Vec3::new(1.0, 0.0, 0.0), Vec3::new(0.0, 1.0, 0.0)],
+        };
+        let bad = Triangle {
+            v: [Vec3::new(f64::NAN, 0.0, 0.0), Vec3::new(1.0, 0.0, 0.0), Vec3::new(0.0, 1.0, 0.0)],
+        };
+        let mut mesh = TriangleMesh {
+            triangles: vec![good, bad],
+            bounds: AABB {
+                min: Vec3::new(0.0, 0.0, 0.0),
+                max: Vec3::new(1.0, 1.0, 0.0),
+            },
+        };
+        mesh.sanitize();
+        assert_eq!(mesh.triangles.len(), 1);
+    }
+
+    #[test]
+    fn sanitize_keeps_valid_mesh() {
+        let t1 = Triangle {
+            v: [Vec3::new(0.0, 0.0, 0.0), Vec3::new(1.0, 0.0, 0.0), Vec3::new(0.0, 1.0, 0.0)],
+        };
+        let t2 = Triangle {
+            v: [Vec3::new(1.0, 0.0, 0.0), Vec3::new(1.0, 1.0, 0.0), Vec3::new(0.0, 1.0, 0.0)],
+        };
+        let mut mesh = TriangleMesh {
+            triangles: vec![t1, t2],
+            bounds: AABB {
+                min: Vec3::new(0.0, 0.0, 0.0),
+                max: Vec3::new(1.0, 1.0, 0.0),
+            },
+        };
+        mesh.sanitize();
+        assert_eq!(mesh.triangles.len(), 2, "both valid triangles should survive");
+    }
+
+    #[test]
+    fn sanitize_recomputes_bounds() {
+        let good = Triangle {
+            v: [Vec3::new(0.0, 0.0, 0.0), Vec3::new(1.0, 0.0, 0.0), Vec3::new(0.0, 1.0, 0.0)],
+        };
+        // This triangle has vertices way out at x=100 but is degenerate
+        let degenerate = Triangle {
+            v: [Vec3::new(100.0, 0.0, 0.0), Vec3::new(100.0, 0.0, 0.0), Vec3::new(100.0, 0.0, 0.0)],
+        };
+        let mut mesh = TriangleMesh {
+            triangles: vec![good, degenerate],
+            bounds: AABB {
+                min: Vec3::new(0.0, 0.0, 0.0),
+                max: Vec3::new(100.0, 1.0, 0.0),
+            },
+        };
+        mesh.sanitize();
+        assert!(mesh.bounds.max.x < 2.0, "bounds should not include removed triangle");
     }
 }