v1.1.11 - Black Hole Sun

[v1.1.11] - Black Hole Sun - 7/9/2026

Added

• GPU-accelerated HNSW construction on CUDA and Vulkan.
  New pkg/search/hnsw_build_cuda.go and hnsw_build_vulkan.go add optional
  GPU build backends for HNSW graph construction, with matching stubs for
  builds without CUDA/Vulkan support. pkg/gpu/cuda/cuda_bridge.go gains the
  CUDA bridge functions; pkg/gpu/vulkan/compute.go adds a rewritten Vulkan
  compute pipeline with dedicated shaders (hnsw_build_cosine.comp,
  hnsw_build_topk_rows.comp and their compiled SPIR-V). Falls back to CPU
  build when GPU backends are unavailable. Follow-up commit tightens Vulkan
  shader dispatch/binding performance (perf(hnsw): optimize vulkan).
• Adjacency snapshot-isolation API for storage-backed graph reads.
  Added a new adjacency snapshot-isolation path across pkg/storage and
  pkg/cypher so traversal and delete-adjacent reads resolve visible edges
  against the correct MVCC view instead of relying on coarser snapshot
  behavior. This also threads through namespaced and WAL-wrapped engines and
  adds focused regression coverage in storage and server graph tests.

Fixed

• Multi-MATCH relationship variable bound in a later clause was silently
  dropped.
  MATCH (s) WHERE s.uid IN $u MATCH (s)-[rel]->() WHERE rel.evidence_source = $e DELETE rel deleted zero edges instead of the
  matching set, and the same shape used as a read (RETURN count(rel),
  RETURN rel, RETURN rel.prop, elementId(rel)) silently returned
  zero/nil for the relationship column while node columns in the same query
  resolved correctly. Root cause: pkg/cypher/match_multi.go's multi-match
  binding row (type binding map[string]*storage.Node) can only hold node
  values, so executeFirstMatch/executeChainedMatch never stored the
  relationship a clause's pattern bound, even though the PathResult
  computed by the traversal already carried it. Kept binding itself
  unchanged (a map[string]*storage.Node value-typed literal is
  constructed/indexed directly by ~10 existing binding-where test files) and
  instead threaded a parallel, index-aligned map[string]*storage.Edge per
  row through executeFirstMatch, executeChainedMatch, a new
  filterBindingsByWhereWithRels (relationship-aware WHERE filtering that
  reuses the unchanged node-only WHERE compiler via a read-only
  bindingWithRelView property adapter), and resolveBindingExpr/
  resolveBindingItem. Also fixed a related gap where executeMultiMatch
  never applied SKIP/LIMIT despite a comment claiming it did. Added
  pkg/cypher/multi_match_relationship_binding_bug_test.go
  (TestBug_MultiMatchRelationshipBindingLost + _Variations,
  TestMergeRelBindings, TestBindingWithRelView,
  TestResolveBindingExprUnboundVariable). Follow-up fixes now also restore
  MATCH ... WITH ... MATCH ... RETURN and ... DELETE rel pipeline shapes
  by delegating valid chained pipeline forms through the general pipeline
  executor, preserving row-bound node/relationship context across later MATCH
  clauses, and trimming trailing ORDER BY / SKIP / LIMIT correctly in
  pipeline RETURN projection.
• CREATE ... WITH ... query routing now distinguishes missing behavior from
  invalid syntax.
  Valid CREATE ... WITH ... RETURN, CREATE ... WITH ... MATCH ... RETURN,
  and related pipeline/fallback shapes now execute instead of failing as
  generically unsupported, while malformed tails now surface deterministic
  invalid-query errors and correctly roll back implicit single-statement
  transactions. This reuses the existing multiple-create executor as a strict
  fallback after pipeline routing and adds regression coverage for the rollback
  case.
• Cypher map keys containing : now parse correctly across map-literal
  surfaces.
  Quoted keys such as {'key:key': 'value'} were previously split at the
  first colon and mis-parsed in SET, MERGE, helper evaluators, APOC map
  parsing, and pipeline/map-literal call sites. Added a shared top-level
  key/value separator helper in pkg/cypher/pattern_parser.go and applied it
  across the affected map parsers, with parser-level and end-to-end regression
  coverage.
• Bolt 4.x datetime/time encoding compatibility restored.
  pkg/bolt/packstream.go and pkg/bolt/server.go now negotiate older Bolt
  datetime encodings correctly, including the Rust-driver compatibility path
  and UTC/compatibility-sensitive record emission. Added focused packstream and
  server regressions for datetime structure decoding and negotiated time
  encoding.
• Bound relationship delete correctness hardened.
  Unified delete projection typing so relationship delete targets are preserved
  as relationships rather than flattened into mismatched row shapes, and
  normalized dangling-edge traversal semantics so stale adjacency rows are
  skipped consistently instead of surfacing inconsistent delete behavior. This
  also adds focused delete-helper and chained-traversal regressions.
• MVCC adjacency / visibility regressions corrected.
  Fixed several storage-layer correctness issues in the new snapshot-adjacency
  path, including edge visibility flags being dropped while copying materialized
  edges, pruning/order bugs that could tombstone visible adjacency incorrectly,
  and namespace filtering being applied too late during unprefixing.
• Fulltext query parser: field:"value" AND (term) now intersects correctly.
  db.index.fulltext.queryNodes / queryRelationships previously tokenized the
  query with a whitespace splitter that had no notion of parenthesized groups,
  field-scoped clauses, or Lucene escape sequences. The Graphiti integration
  shape group_id:"g" AND (<terms>) silently discarded the parenthesized
  default-field clause — a real term and a nonsense term returned the identical
  result set, making the lexical arm of hybrid search term-blind. Replaced the
  ad-hoc tokenizer with a proper Lucene-classic recursive-descent parser
  (pkg/cypher/fulltext_query.go) plus per-document evaluator that supports
  the full grammar: boolean AND/OR/NOT with parens and nesting, +/-
  mandatory/prohibited clause prefixes, phrase queries and proximity
  ("a b"~n), fuzzy (term~n, Levenshtein), range queries ([a TO b],
  {a TO b} with mixed inclusivity), boost (^n), wildcards (?, *
  including leading and mid-token), regex (/re/), and full Lucene escape
  rules (\X decodes to literal X for any X — fixes the Cloud\Trail
  zero-hits case). Reference implementation: Neo4j's MultiFieldQueryParser
  with setAllowLeadingWildcard(true). Added decodeCypherStringLiteral in
  pkg/cypher/call_fulltext.go so backslash escapes survive round-trip
  through Cypher parameter substitution.
• Post-WITH WHERE filtering and grouped aggregation restored.
  A cluster of Cypher WHERE evaluation issues against clean graph queries:
  • Split post-WITH WHERE from WITH projections so it is applied as a
    filter instead of being absorbed into the preceding alias.
  • Preserve grouped WITH aggregation semantics for OPTIONAL MATCH rows,
    including COUNT(c) over null optional targets.
  • Honor inline target labels/properties in relationship pattern predicates
    such as NOT (t)-[:R]->(:C {met:false}).
  • Evaluate bare boolean properties in traversal WHERE clauses so
    WHERE NOT c.met returns rows where c.met is false.
  • Allow multiple relationship CREATE clauses with inline endpoint nodes
    in one statement.
    Touched: pkg/cypher/clauses.go, create_pipeline_helpers.go,
    executor_mutations.go, match_rows.go, traversal.go plus new
    pkg/cypher/stats_query_test.go.
• DDL / CALL-tail / UNWIND / typed conversion cluster fixes.
  • Preserve cardinality constraint parser errors for malformed
    REQUIRE MAX COUNT DDL.
  • Avoid compiling CALL-tail regex predicates using =~ as equality
    comparisons.
  • Fix post-UNWIND WHERE boundary parsing so equality/inequality filters
    apply correctly.
  • Prefer explicit typed assignment conversions before generic reflect
    conversion.
  • Deterministic coverage tests added for schema DDL, CALL-tail predicates,
    helper branches, UNWIND filtering, and typed result assignment
    (pkg/cypher/coverage_lift_test.go +1163 lines).

Performance

• IN-list-anchored relationship traversal now index-seeds instead of
  scanning the whole label.
  MATCH (s:Label)-[rel]->() WHERE s.uid IN $list ... (with or without an
  additional AND predicate on the bound relationship, e.g.
  rel.evidence_source = $e) previously fell through every start-node
  pruning branch in executeMatchWithRelationshipsWithPath
  (pkg/cypher/traversal.go) — none of them recognized an IN [...] /
  IN $param predicate — straight to loadNodesWithTemporalViewport, an
  O(all nodes of the label) scan, even though the equivalent node-only
  MATCH (s:Label) WHERE s.uid IN $list RETURN s already used the schema
  property index via tryCollectNodesFromPropertyIndexIn(Literal)
  (pkg/cypher/match_index_seek.go, already wired into match.go,
  clauses.go, and executor_mutations.go). Added
  tryCollectNodesFromPropertyIndexInCompound, which wires those existing
  index-seek helpers into the traversal start-node pruning chain —
  including when the IN-list is one conjunct of an AND-combined WHERE
  clause, mirroring tryCollectNodesFromIDEqualityCompound's conjunct
  handling. Correctness is unaffected: filterPathsByWhere still
  re-evaluates the full WHERE clause after seeding, so pruning from one
  recognized conjunct can only over-fetch, never under-fetch.
  BenchmarkInListAnchoredRelMatch (50k-node label, 100-node target
  sublist; 5k was tried first but the algorithmic difference is inside
  measurement noise for an in-process MemoryEngine at that size) on this
  branch: 109,930,838 ns/op → 301,410 ns/op (~365x), 83,268,576 B/op →
  228,696 B/op (~364x), 1,388,688 allocs/op → 2,800 allocs/op (~496x).
  Added pkg/cypher/inlist_start_node_index_seed_bug_test.go
  (TestBug_InListStartNodeDoesNotIndexSeed + scan-budget and DELETE
  variants, TestTryCollectNodesFromPropertyIndexInCompound,
  BenchmarkInListAnchoredRelMatch).
• Match/merge hot path from Graphify workloads.
  pkg/cypher/match_multi.go, merge.go, and pkg/storage/schema.go gain a
  fast pattern-property index lookup so multi-pattern MATCH/MERGE
  statements hit the schema index directly instead of scanning candidates.
  New benchmark (graphify_push_profile_bench_test.go) and
  pattern-property regression tests (match_pattern_property_index_test.go)
  guard the change.
• Bound relationship deletes now use targeted source lookup and cheaper
  snapshot adjacency reads.
  The delete hot path now resolves bound relationship candidates using indexed
  and node-local transaction-snapshot reads instead of broader adjacency scans,
  reducing the cost of relationship delete workloads while keeping the delete
  projection contract intact.
• Vulkan compute shaders tuned.
  Dispatch/binding refactor in pkg/gpu/vulkan/compute.go following the GPU
  HNSW build feature.

Changed

• Dependency refresh — Go modules.
  • google.golang.org/api v0.286.0 → v0.287.0
  • google.golang.org/grpc v1.81.1 → v1.82.0
  • github.com/googleapis/enterprise-certificate-proxy v0.3.16 → v0.3.17
    (indirect)
  • google.golang.org/genproto/googleapis/rpc bumped to 20260622175928
    (indirect)
• Dependency refresh — UI npm packages.
  • @ornery/ui-grid-core / -react / -vanilla ^1.0.6 → ^1.0.8
  • lucide-react ^1.22.0 → ^1.23.0
  • neo4j-driver ^6.1.0 → ^6.2.0
  • react-router-dom ^7.18.0 → ^7.18.1
  • three ^0.184.0 → ^0.185.0
• Dependabot workflow versions refreshed across .github/workflows/
  (cd-llama-cpu.yml, cd-llama-cuda.yml, cd.yml, ci.yml,
  docs-pages.yml, release-macos.yml).
• Documentation. README.md copy tweak. ORM/Neo4j-compatible streaming
  driver plan (plans/) expanded with implementation detail (+556/-189).

Tests

• New pkg/cypher/coverage_lift_test.go (+1163 lines) exercising DDL,
  CALL-tail predicates, UNWIND filtering, and typed assignment.
• New pkg/cypher/multi_match_relationship_binding_bug_test.go and
  pkg/cypher/inlist_start_node_index_seed_bug_test.go covering the
  multi-MATCH relationship-binding regression, the follow-up MATCH/WITH/MATCH
  pipeline behavior, and IN-list traversal index-seeding correctness,
  scan-budget, DELETE, and benchmark variants.
• New pkg/cypher/kalman_functions_test.go covering the Kalman helper
  branches.
• New pkg/cypher/fulltext_query_test.go and
  pkg/cypher/call_fulltext_parser_test.go covering the full Lucene-classic
  grammar surface plus e2e regressions against the Graphiti bug repro.
• New pkg/cypher/stats_query_test.go, match_pattern_property_index_test.go,
  graphify_push_profile_bench_test.go guarding the WHERE-semantics and
  hot-path match/merge changes.
• pkg/search/hnsw_build_gpu_test.go extended for GPU-build coverage.
• New Bolt compatibility regressions in pkg/bolt/packstream_into_test.go and
  pkg/bolt/server_test.go covering older PackStream datetime structures,
  negotiated datetime emission, and Rust-driver compatibility.

What's Changed

• feat(cypher,storage): adding adjacency snapshot isolation api by @orneryd in #253
• Speed up bound relationship deletes by @linuxdynasty in #230
• cover the bound relationship delete fast-path branches (follow-up to #230) by @linuxdynasty in #254
• fix(cypher): bind relationships across multi-MATCH clauses; perf(cypher): index-seed IN-list traversal starts by @linuxdynasty in #259

Full Changelog: v1.1.10...v1.1.11