diff --git a/laurus/Cargo.toml b/laurus/Cargo.toml
index 91c3129b..9dd9b1fb 100644
--- a/laurus/Cargo.toml
+++ b/laurus/Cargo.toml
@@ -176,6 +176,10 @@ harness = false
 name = "pq_train_bench"
 harness = false
 
+[[bench]]
+name = "deletion_search_bench"
+harness = false
+
 [[bench]]
 name = "vector_multi_query_bench"
 harness = false
diff --git a/laurus/benches/deletion_search_bench.rs b/laurus/benches/deletion_search_bench.rs
new file mode 100644
index 00000000..73d3dc6a
--- /dev/null
+++ b/laurus/benches/deletion_search_bench.rs
@@ -0,0 +1,171 @@
+//! Deletion-aware HNSW search benchmark (Issue #625).
+//!
+//! Measures `VectorStore::search` (HNSW graph path) on a committed
+//! in-memory index in two states:
+//!
+//! - `del0pct` — no deletions: the pristine, bookkeeping-free traversal;
+//!   the no-regression guard for #625.
+//! - `del10pct` — 10% of docs soft-deleted: every admitted neighbour is
+//!   probed against the deletion set inside the bookkeeping loop. This
+//!   is the gate metric for #625 (per-query snapshot + lock-free
+//!   `contains` vs the previous per-probe bitmap read-lock).
+//!
+//! In-memory storage loads Eager, so distances run the int8 SIMD kernel
+//! and the traversal + deletion bookkeeping dominate. Setup is
+//! deterministic via `common::DEFAULT_SEED`. Uses only public APIs that
+//! also exist on `main`, so the same bench runs on the src-only
+//! `git stash` baseline.
+
+mod common;
+
+use std::any::Any;
+use std::collections::HashMap;
+use std::sync::Arc;
+
+use async_trait::async_trait;
+use criterion::{Criterion, criterion_group, criterion_main};
+use tokio::runtime::Runtime;
+
+use common::{DEFAULT_SEED, SAMPLE_SIZE_FAST, lcg_vec_unit};
+
+use laurus::lexical::LexicalIndexConfig;
+use laurus::storage::Storage;
+use laurus::storage::memory::{MemoryStorage, MemoryStorageConfig};
+use laurus::vector::core::distance::DistanceMetric;
+use laurus::vector::core::field::HnswOption;
+use laurus::vector::search::searcher::{VectorSearchParams, VectorSearchRequest};
+use laurus::vector::store::config::VectorFieldConfig;
+use laurus::vector::store::request::{QueryVector, VectorScoreMode};
+use laurus::vector::{FieldOption, Vector, VectorIndexConfig, VectorSearchQuery, VectorStore};
+use laurus::{DataValue, Document, EmbedInput, EmbedInputType, Embedder, LaurusError, Result};
+
+const DIM: usize = 128;
+/// Corpus size: large enough that ef-search traversal probes hundreds
+/// of neighbours per query.
+const N: u64 = 10_000;
+
+#[derive(Debug)]
+struct MockEmbedder;
+
+#[async_trait]
+impl Embedder for MockEmbedder {
+    async fn embed(&self, _input: &EmbedInput<'_>) -> Result<Vector> {
+        Err(LaurusError::invalid_argument(
+            "vectors are supplied directly",
+        ))
+    }
+    fn supported_input_types(&self) -> Vec<EmbedInputType> {
+        vec![EmbedInputType::Text]
+    }
+    fn name(&self) -> &str {
+        "mock"
+    }
+    fn as_any(&self) -> &dyn Any {
+        self
+    }
+}
+
+fn make_config() -> VectorIndexConfig {
+    let mut fields = HashMap::new();
+    fields.insert(
+        "vec".to_string(),
+        VectorFieldConfig {
+            vector: Some(FieldOption::Hnsw(HnswOption {
+                dimension: DIM,
+                distance: DistanceMetric::Cosine,
+                m: 16,
+                ef_construction: 200,
+                default_ef_search: None,
+                base_weight: 1.0,
+                quantizer: Default::default(),
+                rerank_storage: None,
+                embedder: None,
+            })),
+            lexical: None,
+        },
+    );
+    VectorIndexConfig {
+        fields,
+        embedder: Arc::new(MockEmbedder),
+        default_fields: vec!["vec".to_string()],
+        metadata: HashMap::new(),
+        deletion_config: laurus::DeletionConfig::default(),
+        shard_id: 0,
+        metadata_config: LexicalIndexConfig::default(),
+    }
+}
+
+/// Build a committed `VectorStore` holding `N` deterministic vectors.
+fn build_committed_store(rt: &Runtime) -> VectorStore {
+    let storage: Arc<dyn Storage> = Arc::new(MemoryStorage::new(MemoryStorageConfig::default()));
+    let store = VectorStore::new(storage, make_config()).unwrap();
+    let mut state = DEFAULT_SEED;
+    rt.block_on(async {
+        for id in 0..N {
+            let vec = lcg_vec_unit(&mut state, DIM);
+            let doc = Document::builder()
+                .add_field("vec", DataValue::Vector(vec))
+                .build();
+            store.upsert_document_by_internal_id(id, doc).await.unwrap();
+        }
+        store.commit().await.unwrap();
+    });
+    store
+}
+
+/// Top-10 request against the `vec` field (graph path).
+fn request() -> VectorSearchRequest {
+    let mut state = DEFAULT_SEED.wrapping_add(1);
+    VectorSearchRequest {
+        query: VectorSearchQuery::Vectors(vec![QueryVector {
+            vector: Vector::new(lcg_vec_unit(&mut state, DIM)),
+            weight: 1.0,
+            fields: Some(vec!["vec".to_string()]),
+        }]),
+        params: VectorSearchParams {
+            limit: 10,
+            score_mode: VectorScoreMode::WeightedSum,
+            ..Default::default()
+        },
+    }
+}
+
+fn bench_deletion_search(c: &mut Criterion) {
+    let rt = Runtime::new().unwrap();
+    let mut group = c.benchmark_group("Deletion Search");
+    group.sample_size(SAMPLE_SIZE_FAST);
+
+    let store = build_committed_store(&rt);
+
+    // Sanity: the graph path must return hits before timing.
+    let probe = store.search(request()).unwrap();
+    assert!(!probe.hits.is_empty(), "probe must return hits");
+
+    // 0% deletions: pristine traversal — the no-regression guard.
+    group.bench_function("top10/del0pct", |b| {
+        b.iter(|| store.search(request()).unwrap())
+    });
+
+    // Soft-delete 10% of docs (every 10th id), then measure the
+    // deletion-bookkeeping traversal. No commit needed: the bitmap is
+    // live and attached to freshly built readers.
+    rt.block_on(async {
+        for id in (0..N).step_by(10) {
+            store.delete_document_by_internal_id(id).await.unwrap();
+        }
+    });
+    let probe = store.search(request()).unwrap();
+    assert!(
+        !probe.hits.is_empty(),
+        "post-delete probe must still return hits"
+    );
+
+    group.bench_function("top10/del10pct", |b| {
+        b.iter(|| store.search(request()).unwrap())
+    });
+
+    group.finish();
+}
+
+criterion_group!(benches, bench_deletion_search);
+criterion_main!(benches);