fix(sqlite): WAL checkpoint + add view_index for mmap loading

src/registry.rs

@@ -87,7 +87,7 @@ impl USearchIndexConfig {
             .map_err(|e| DataFusionError::Execution(format!("USearch Index::new failed: {e}")))
     }
 
-    /// Load a previously saved index from `path`.
+    /// Load a previously saved index from `path` into memory.
     ///
     /// Uses the same `IndexOptions` as `build_index()`.  The options must
     /// match those used when the index was originally built — passing wrong
@@ -101,6 +101,26 @@ impl USearchIndexConfig {
         Ok(index)
     }
 
+    /// Memory-map a previously saved index from `path`.
+    ///
+    /// Unlike [`load_index`], this does not copy the index into RAM. The OS
+    /// pages data in on demand, keeping resident memory proportional to the
+    /// working set rather than the full index size. Prefer this for the
+    /// reload-from-disk path where the index file is already local.
+    ///
+    /// The returned [`Index`] is fully functional for search; the backing
+    /// file must remain on disk for the lifetime of the index.
+    ///
+    /// [`load_index`]: Self::load_index
+    pub fn view_index(&self, path: &str) -> Result<Index> {
+        let index = Index::new(&self.to_index_options())
+            .map_err(|e| DataFusionError::Execution(format!("USearch Index::new failed: {e}")))?;
+        index
+            .view(path)
+            .map_err(|e| DataFusionError::Execution(format!("USearch index view failed: {e}")))?;
+        Ok(index)
+    }
+
     fn to_index_options(&self) -> IndexOptions {
         IndexOptions {
             dimensions: self.dimensions,

src/sqlite_provider.rs

@@ -252,6 +252,12 @@ impl SqliteLookupProvider {
             )?;
         }
 
+        // Checkpoint WAL so the data is flushed to the main database file.
+        // Without this, data written during build may only exist in the WAL
+        // and can be lost if the process exits before a passive checkpoint.
+        conn.execute_batch("PRAGMA wal_checkpoint(TRUNCATE);")
+            .map_err(|e| DataFusionError::Execution(format!("WAL checkpoint failed: {e}")))?;
+
         let mut conns = vec![conn];
         for _ in 1..pool_size {
             conns.push(open_conn(db_path)?);

tests/optimizer_rule.rs

@@ -419,3 +419,52 @@ async fn test_qualified_ref_where_clause_rewrites() {
         "qualified ref + WHERE → filter absorbed, rule must fire\nPlan: {plan:?}"
     );
 }
+
+// ═══════════════════════════════════════════════════════════════════════════════
+// SELECT only distance — no base columns projected
+// ═══════════════════════════════════════════════════════════════════════════════
+//
+// When the SELECT list contains only the distance UDF (no base table columns),
+// the Projection node has a single computed expression. The optimizer must still
+// recognise the pattern and rewrite to USearchNode.
+
+/// Bare table, SELECT only distance alias, ORDER BY alias — rule must fire.
+#[tokio::test]
+async fn test_select_only_distance_bare_rewrites() {
+    let ctx = make_ctx(MetricKind::L2sq).await;
+    let sql =
+        format!("SELECT l2_distance(vector, {Q}) AS dist FROM items ORDER BY dist ASC LIMIT 5");
+    let plan = optimized_plan(&ctx, &sql).await;
+    assert!(
+        contains_usearch_node(&plan),
+        "SELECT only distance (bare) → rule must fire\nPlan: {plan:?}"
+    );
+}
+
+/// Qualified table, SELECT only distance alias, ORDER BY alias — rule must fire.
+#[tokio::test]
+async fn test_select_only_distance_qualified_rewrites() {
+    let ctx = make_ctx_qualified(MetricKind::L2sq).await;
+    let sql = format!(
+        "SELECT l2_distance(vector, {Q}) AS dist FROM datafusion.public.items ORDER BY dist ASC LIMIT 5"
+    );
+    let plan = optimized_plan(&ctx, &sql).await;
+    assert!(
+        contains_usearch_node(&plan),
+        "SELECT only distance (qualified) → rule must fire\nPlan: {plan:?}"
+    );
+}
+
+/// Bare table, SELECT only distance (no alias), ORDER BY the UDF directly.
+#[tokio::test]
+async fn test_select_only_distance_no_alias_rewrites() {
+    let ctx = make_ctx(MetricKind::L2sq).await;
+    let sql = format!(
+        "SELECT l2_distance(vector, {Q}) FROM items ORDER BY l2_distance(vector, {Q}) ASC LIMIT 5"
+    );
+    let plan = optimized_plan(&ctx, &sql).await;
+    assert!(
+        contains_usearch_node(&plan),
+        "SELECT only distance (no alias, ORDER BY UDF) → rule must fire\nPlan: {plan:?}"
+    );
+}