perf(snap): keep snap/hover responsive on block-heavy drawings

Snap and hover hit-testing scanned the whole model wire set every cursor move; a heavily block-instanced drawing explodes to millions of wires, so this stalled for over a second per move. Add a cached spatial grid (wire_grid) keyed on the geometry epoch and query only the cursor neighbourhood via Scene::hit_test_wires_near. Gate the O(in-range segments squared) intersection passes on the in-range POINT count, not the wire count — a handful of curved wires, each hundreds of tessellated points, blew the quadratic up while the wire count looked modest. Give zero-size entities away from the origin (POINTs) a real cullable AABB, and fall back to a block sub-wire's own world points when its entity bounding box is degenerate.

Co-Authored-By: Claude Opus 4.8 <[email protected]>

src/scene/pick/mod.rs

@@ -2,4 +2,5 @@ pub mod hit_test;
 pub mod selection_state;
 pub mod grip;
 pub mod quadtree;
+pub mod wire_grid;
 pub mod xclip;

src/scene/pick/wire_grid.rs

@@ -0,0 +1,178 @@
+//! Uniform spatial grid over a hit-test wire set's world-XY AABBs.
+//!
+//! Snap and hover hit-testing scan the full model wire set on every cursor
+//! move. On a heavily block-instanced drawing that set explodes to millions of
+//! wires, so an O(N) per-move scan — even with a cheap AABB reject — stalls the
+//! event loop. This grid is built once per geometry epoch (cached alongside the
+//! wire set) and answers "which wires are near this cursor point" in O(cells in
+//! range), turning the per-move cost from O(all wires) into O(local).
+//!
+//! Wires whose AABB is unbounded (previews / entities with no usable
+//! bounding box) or spans a large fraction of the grid (drawing-length lines)
+//! go into an always-checked `oversized` list rather than being smeared across
+//! thousands of cells.
+
+use crate::scene::model::wire_model::WireModel;
+
+/// A wire spanning more than this many cells is stored once in `oversized`
+/// instead of being inserted into every overlapped cell.
+const MAX_SPAN_CELLS: u64 = 64;
+
+pub struct WireGrid {
+    min: [f64; 2],
+    cell: f64,
+    cols: u32,
+    rows: u32,
+    cells: Vec<Vec<u32>>,
+    /// Indices always returned as candidates (unbounded / huge AABB).
+    oversized: Vec<u32>,
+}
+
+impl WireGrid {
+    fn finite_aabb(w: &WireModel) -> Option<[f64; 4]> {
+        let a = w.aabb;
+        (a[0].is_finite() && a[1].is_finite() && a[2].is_finite() && a[3].is_finite())
+            .then_some([a[0] as f64, a[1] as f64, a[2] as f64, a[3] as f64])
+    }
+
+    pub fn build(wires: &[WireModel]) -> Self {
+        // Overall finite extent.
+        let mut min = [f64::INFINITY; 2];
+        let mut max = [f64::NEG_INFINITY; 2];
+        for w in wires {
+            if let Some(a) = Self::finite_aabb(w) {
+                min[0] = min[0].min(a[0]);
+                min[1] = min[1].min(a[1]);
+                max[0] = max[0].max(a[2]);
+                max[1] = max[1].max(a[3]);
+            }
+        }
+        if !min[0].is_finite() {
+            // No finite wires — everything is oversized (always checked).
+            return Self {
+                min: [0.0, 0.0],
+                cell: 1.0,
+                cols: 1,
+                rows: 1,
+                cells: vec![Vec::new()],
+                oversized: (0..wires.len() as u32).collect(),
+            };
+        }
+        let ext_x = (max[0] - min[0]).max(1e-6);
+        let ext_y = (max[1] - min[1]).max(1e-6);
+        // Aim for ~8 wires/cell; a square-ish cell keeps queries tight.
+        let target_cells = ((wires.len() / 8).max(1)) as f64;
+        let cell = ((ext_x * ext_y) / target_cells).sqrt().max(1e-6);
+        let cols = (((ext_x / cell).ceil() as i64) + 1).clamp(1, 8192) as u32;
+        let rows = (((ext_y / cell).ceil() as i64) + 1).clamp(1, 8192) as u32;
+
+        let mut cells = vec![Vec::<u32>::new(); (cols as usize) * (rows as usize)];
+        let mut oversized = Vec::new();
+
+        let col_of = |x: f64| (((x - min[0]) / cell).floor()).clamp(0.0, (cols - 1) as f64) as u32;
+        let row_of = |y: f64| (((y - min[1]) / cell).floor()).clamp(0.0, (rows - 1) as f64) as u32;
+
+        for (idx, w) in wires.iter().enumerate() {
+            let Some(a) = Self::finite_aabb(w) else {
+                oversized.push(idx as u32);
+                continue;
+            };
+            let c0 = col_of(a[0]);
+            let c1 = col_of(a[2]);
+            let r0 = row_of(a[1]);
+            let r1 = row_of(a[3]);
+            let span = (c1 - c0 + 1) as u64 * (r1 - r0 + 1) as u64;
+            if span > MAX_SPAN_CELLS {
+                oversized.push(idx as u32);
+                continue;
+            }
+            for r in r0..=r1 {
+                let base = (r * cols) as usize;
+                for c in c0..=c1 {
+                    cells[base + c as usize].push(idx as u32);
+                }
+            }
+        }
+
+        Self {
+            min,
+            cell,
+            cols,
+            rows,
+            cells,
+            oversized,
+        }
+    }
+
+    /// Candidate wire indices whose cell overlaps the `radius`-padded box around
+    /// `(x, y)`. Always includes the oversized list. May contain a wire more
+    /// than once is avoided by dedup.
+    pub fn query(&self, x: f64, y: f64, radius: f64) -> Vec<u32> {
+        let r = radius.max(0.0);
+        let mut out = self.oversized.clone();
+        // Query box fully outside the grid → only oversized candidates.
+        if x + r < self.min[0]
+            || y + r < self.min[1]
+            || x - r > self.min[0] + self.cols as f64 * self.cell
+            || y - r > self.min[1] + self.rows as f64 * self.cell
+        {
+            return out;
+        }
+        let clampc = |v: f64| v.clamp(0.0, (self.cols - 1) as f64) as u32;
+        let clampr = |v: f64| v.clamp(0.0, (self.rows - 1) as f64) as u32;
+        let c0 = clampc(((x - r) - self.min[0]) / self.cell);
+        let c1 = clampc(((x + r) - self.min[0]) / self.cell);
+        let r0 = clampr(((y - r) - self.min[1]) / self.cell);
+        let r1 = clampr(((y + r) - self.min[1]) / self.cell);
+        for row in r0..=r1 {
+            let base = (row * self.cols) as usize;
+            for col in c0..=c1 {
+                out.extend_from_slice(&self.cells[base + col as usize]);
+            }
+        }
+        out.sort_unstable();
+        out.dedup();
+        out
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn wire(x: f32, y: f32) -> WireModel {
+        let mut w = WireModel::solid_f64("t".into(), vec![[x as f64, y as f64, 0.0]], [0.0; 4], false);
+        w.aabb = [x, y, x, y];
+        w
+    }
+
+    #[test]
+    fn query_is_local_and_hits_the_point() {
+        // 50×50 lattice of point-wires spaced 10 apart (2500 wires).
+        let mut wires = Vec::new();
+        for gx in 0..50 {
+            for gy in 0..50 {
+                wires.push(wire(gx as f32 * 10.0, gy as f32 * 10.0));
+            }
+        }
+        let n = wires.len();
+        let grid = WireGrid::build(&wires);
+        // Tight query around one lattice point.
+        let hits = grid.query(250.0, 250.0, 5.0);
+        assert!(hits.len() < n / 10, "query too broad: {} of {n}", hits.len());
+        let idx = (25u32 * 50) + 25; // gx=25, gy=25
+        assert!(hits.contains(&idx), "the on-point wire must be a candidate");
+        // A far point must not drag in the whole lattice.
+        let far = grid.query(-9999.0, -9999.0, 5.0);
+        assert!(far.len() < n / 10, "far query too broad: {}", far.len());
+    }
+
+    #[test]
+    fn unbounded_wire_is_always_a_candidate() {
+        let mut wires = vec![wire(0.0, 0.0), wire(1_000_000.0, 1_000_000.0)];
+        wires[1].aabb = WireModel::UNBOUNDED_AABB;
+        let grid = WireGrid::build(&wires);
+        let hits = grid.query(0.0, 0.0, 1.0);
+        assert!(hits.contains(&1), "unbounded wire must always be returned");
+    }
+}