GPU & Memory
- Fixed GPU kernel OOM by routing all cudaMalloc through GPUMemPool with pager pinning — kills eviction races during Search.
- CUDA error checks after kernel launches; shared-memory overflow guard in TurboQuant v2.
SIMD & Kernels - Full AVX2/AVX512 dispatch for int32/uint32/int64/uint64 dot + Euclidean (+ fp16, float64, complex128).
- Raw AMX assembly kernels; Geo SIMD float32 haversine via SincosFloat32 batched passes.
- int8 L2Squared kernel, complex128 magnitude pruning.
Distributed & Graph - SWIM gossip for distributed boundary sync; LockFreeNeighborCache via sync.Map.
- GraphRAG activation propagation; visited-nodes budget for filtered search.
- Async temporal ingestion enabled by default; SegmentTree for temporal search.
ADBC Driver (new) - Full C-API binding with Cgo ABI, Makefile build, 100% statement coverage, cross-language smoke test.
Full Changelog: 0.2.1...0.2.2