fix(udf): remove sqrt from l2_distance to match USearch L2sq metric

l2_distance UDF was computing actual L2 (with sqrt) while USearch and
the rewritten execution paths all use L2sq (no sqrt). This caused the
same query to return different numeric distance values depending on
whether the optimizer rewrote it.

Remove sqrt to match USearch's MetricKind::L2sq and DuckDB VSS's
array_distance behavior. All paths now return consistent L2sq values.

Author: anoop-narang

README.md

@@ -272,11 +272,11 @@ All three distance functions are **lower-is-closer**:
 
 | SQL function | Index metric | Kernel |
 |---|---|---|
-| `l2_distance(a, b)` | `L2sq` | `sqrt(sum((a_i - b_i)^2))` (UDF) / `sum((a_i - b_i)^2)` (index) |
+| `l2_distance(a, b)` | `L2sq` | `sum((a_i - b_i)^2)` |
 | `cosine_distance(a, b)` | `Cos` | `1 - dot(a,b) / (norm(a) * norm(b))` |
 | `negative_dot_product(a, b)` | `IP` | `-(a . b)` |
 
-Note: `l2_distance` UDF returns actual L2 (with sqrt) for human-readable distances; USearch uses L2sq internally (no sqrt). The sort order is identical.
+`l2_distance` returns squared L2 (no sqrt), matching USearch's `MetricKind::L2sq`. This ensures numeric consistency between the UDF, the rewritten index path, and the brute-force path.
 
 ### Running tests
 

src/lib.rs

@@ -93,7 +93,7 @@ use datafusion::prelude::SessionContext;
 /// Register all extension components with a DataFusion [`SessionContext`].
 ///
 /// Registers:
-/// - `l2_distance(col, query)`          — Euclidean distance (L2)
+/// - `l2_distance(col, query)`          — squared Euclidean distance (L2sq)
 /// - `cosine_distance(col, query)`      — cosine distance
 /// - `negative_dot_product(col, query)` — negated inner product
 /// - `vector_usearch(table, query, k)`  — explicit ANN table function

src/udf.rs

@@ -27,7 +27,6 @@ fn l2_kernel(a: &[f32], b: &[f32]) -> f32 {
         .zip(b.iter())
         .map(|(x, y)| (x - y) * (x - y))
         .sum::<f32>()
-        .sqrt()
 }
 
 fn cosine_kernel(a: &[f32], b: &[f32]) -> f32 {