chore[turboquant]: cleanup from second simplify pass

- Use vortex_utils::aliases::dash_map::DashMap instead of raw dashmap,
  matching codebase convention (removes direct dashmap dependency)
- Fix stale doc comment on gen_random_signs (leftover from deleted
  apply_inverse_srht_from_bits function)
- Move function-scoped `use crate::rotation::RotationMatrix` to test
  module top per CLAUDE.md
- Optimize hot loop: replace padded.fill(0.0) every row with
  conditional [..dim] zeroing only when norm==0. The tail
  [dim..padded_dim] is zeroed once at allocation and never overwritten,
  saving padded_dim×4 bytes of unnecessary stores per row.

Signed-off-by: Will Manning <will@spiraldb.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Signed-off-by: Will Manning <will@willmanning.io>

Author: lwwmanning

Cargo.lock

@@ -17,7 +17,6 @@ version = { workspace = true }
 workspace = true
 
 [dependencies]
-dashmap = { workspace = true }
 prost = { workspace = true }
 rand = { workspace = true }
 vortex-array = { workspace = true }

encodings/turboquant/Cargo.toml

@@ -11,9 +11,9 @@
 
 use std::sync::LazyLock;
 
-use dashmap::DashMap;
 use vortex_error::VortexResult;
 use vortex_error::vortex_bail;
+use vortex_utils::aliases::dash_map::DashMap;
 
 /// Number of numerical integration points for computing conditional expectations.
 const INTEGRATION_POINTS: usize = 1000;
@@ -25,7 +25,7 @@ const CONVERGENCE_EPSILON: f64 = 1e-12;
 const MAX_ITERATIONS: usize = 200;
 
 /// Global centroid cache keyed by (dimension, bit_width).
-static CENTROID_CACHE: LazyLock<DashMap<(u32, u8), Vec<f32>>> = LazyLock::new(DashMap::new);
+static CENTROID_CACHE: LazyLock<DashMap<(u32, u8), Vec<f32>>> = LazyLock::new(DashMap::default);
 
 /// Get or compute cached centroids for the given dimension and bit width.
 ///

encodings/turboquant/src/centroids.rs

@@ -109,12 +109,15 @@ pub fn turboquant_encode_mse(
         let norm = l2_norm(x);
         norms_buf.push(norm);
 
-        padded.fill(0.0);
+        // Normalize and write into [..dim]; tail [dim..padded_dim] stays zero
+        // from initialization and is never overwritten.
         if norm > 0.0 {
             let inv_norm = 1.0 / norm;
             for (dst, &src) in padded[..dim].iter_mut().zip(x.iter()) {
                 *dst = src * inv_norm;
             }
+        } else {
+            padded[..dim].fill(0.0);
         }
         rotation.rotate(&padded, &mut rotated);
 
@@ -225,12 +228,13 @@ pub fn turboquant_encode_qjl(
         let norm = l2_norm(x);
 
         // Reproduce the same quantization as MSE encoding.
-        padded.fill(0.0);
         if norm > 0.0 {
             let inv_norm = 1.0 / norm;
             for (dst, &src) in padded[..dim].iter_mut().zip(x.iter()) {
                 *dst = src * inv_norm;
             }
+        } else {
+            padded[..dim].fill(0.0);
         }
         rotation.rotate(&padded, &mut rotated);
 

encodings/turboquant/src/compress.rs

@@ -132,6 +132,7 @@ mod tests {
     use vortex_session::VortexSession;
 
     use crate::TurboQuantConfig;
+    use crate::rotation::RotationMatrix;
     use crate::turboquant_encode_mse;
     use crate::turboquant_encode_qjl;
 
@@ -536,8 +537,6 @@ mod tests {
     /// produce identical output.
     #[test]
     fn stored_rotation_signs_produce_correct_decode() -> VortexResult<()> {
-        use crate::rotation::RotationMatrix;
-
         let fsl = make_fsl(20, 128, 42);
         let config = TurboQuantConfig {
             bit_width: 3,

encodings/turboquant/src/lib.rs

@@ -207,12 +207,6 @@ impl RotationMatrix {
     }
 }
 
-/// Apply the inverse SRHT using sign bits stored in a raw byte slice.
-///
-/// This is the hot-path function for decompression. The `signs_bytes` buffer
-/// contains `3 * padded_dim` bits in inverse-application order `[D₃ | D₂ | D₁]`.
-/// Convention: bit set (1) = +1, bit unset (0) = -1 (negate).
-///
 /// Generate a vector of random ±1 signs.
 fn gen_random_signs(rng: &mut StdRng, len: usize) -> Vec<f32> {
     (0..len)