fix(vector): fix 4 bugs causing Windows segfault + silent index corruption (HNSW)

[chiangchenghsin-hash]

Summary

Fixes 4 bugs in the VECTOR extension that cause SEGFAULT (0xC0000005) on Windows and silent index corruption on all platforms. Verified on Windows 11 x64 with MSVC 19.50, LadybugDB 0.18.0.

Supersedes #499 (closed without fix — bug still present in 0.18.0 source).

Bugs Fixed

Bug 1: MSVC const-correctness — create_hnsw_index.cpp:60

NodeTable::getNumTotalRows() is non-const. MSVC rejects const auto& table calling it (error C2662). Clang/GCC tolerate it, so this only blocks Windows source builds.

-    const auto& table =
+    auto& table =
         storage::StorageManager::Get(*context)->getTable(tableID)->cast<storage::NodeTable>();

Bug 2: Stale table cardinality — create_hnsw_index.cpp:64

getStats(transaction).getTableCard() returns a statistics snapshot that can be 0 or stale during uncommitted transactions / WAL replay. This caused numNodes=0 → HNSW graph allocated with wrong size.

-    auto numNodes = table.getStats(transaction).getTableCard();
+    auto numNodes = table.getNumTotalRows(transaction);

Bug 3: DUMMY_CHECKPOINT_TRANSACTION in finalize() — hnsw_index.cpp:672

OnDiskHNSWIndex::finalize() queried row count via the static dummy transaction, which returns 0 in non-checkpoint contexts. This made finalize() silently skip incremental writes — newly inserted nodes were never indexed. This is a platform-agnostic correctness bug (Linux silently loses index coverage; Windows additionally crashes downstream).

 void OnDiskHNSWIndex::finalize(main::ClientContext* context) {
     auto& hnswStorageInfo = storageInfo->cast<HNSWStorageInfo>();
-    const auto numTotalRows = nodeTable.getNumTotalRows(&DUMMY_CHECKPOINT_TRANSACTION);
+    auto transaction = Transaction::Get(*context);
+    const auto numTotalRows = nodeTable.getNumTotalRows(transaction);
     if (numTotalRows == hnswStorageInfo.numCheckpointedNodes) {
         return;
     }
-    auto transaction = Transaction::Get(*context);

Bug 4: INVALID_OFFSET out-of-bounds — hnsw_rel_batch_insert.cpp:127

Uninitialized CSR neighbor slots are marked INVALID_OFFSET. The batch insert loop passed them to graphToNodeOffset() without filtering → out-of-bounds access → 0xC0000005 on Windows (Linux silently writes garbage).

     for (const auto neighbourGraphOffset : neighbours) {
+        if (neighbourGraphOffset == common::INVALID_OFFSET) {
+            continue;
+        }
         const auto neighbourNodeOffset = selectionMap.graphToNodeOffset(neighbourGraphOffset);

Why these matter beyond Windows

Bugs 2, 3, and 4 are platform-agnostic correctness bugs. On Linux they don't crash (glibc's tolerant malloc) but silently produce:

• incomplete HNSW indexes (bug 3: finalize skips new nodes)
• corrupted graph edges (bug 4: garbage offsets written)
• wrong recall (cf. #351 — 18% recall vs Kuzu's 47%, likely caused by incomplete indexes)

Verification

Built from source on Windows 11 with MSVC 19.50 + CMake Ninja (BUILD_EXTENSIONS=vector). End-to-end test:

LOAD EXTENSION 'libvector.lbug_extension';
CREATE NODE TABLE v (id INT64, emb DOUBLE[3], PRIMARY KEY(id));
CREATE (v {id:1, emb:[0.1,0.2,0.3]});
CREATE (v {id:2, emb:[0.4,0.5,0.6]});
CREATE (v {id:3, emb:[0.7,0.8,0.9]});
CALL CREATE_VECTOR_INDEX('v','vi','emb',metric:='cosine');
CALL QUERY_VECTOR_INDEX('v','vi',CAST([0.15,0.25,0.35],'DOUBLE[3]'),2) RETURN node.id, distance;
-- Returns: id=1 dist=0.002585, id=2 dist=0.011805  ✅

Before the fix: LOAD VECTOR → STATUS_ACCESS_VIOLATION (deterministic crash, exit -1073741819).

Related

• Fixes #499 (SEGFAULT on LOAD VECTOR — Windows x64)
• Possibly related to #351 (poor vector recall — incomplete indexes)
• Possibly related to #431 (LOAD EXTENSION fts segfault — same loading path)

[adsharma]

Moved to #23 with CI