Summary

Replaces the hnsw_rs HNSW backend with TurboQuant 4-bit compressed vector index for 8x memory reduction and faster search.

Problem

Current VectorStore uses hnsw_rs for in-memory search with f32 raw vectors. For large vaults:

• HNSW index build is slow (ef_build=200)
• Memory usage is high (10K vectors × 1024-dim ≈ 400MB)
• Cache eviction causes frequent full rebuilds from SQLite

Solution

Before: SQLite (raw f32 BLOBs) + hnsw_rs HNSW (in-memory, f32)
After:  SQLite (raw f32 BLOBs) + TurboQuant 4-bit (in-memory, 8x compressed)

TurboQuant benefits:

• 8x memory reduction (4-bit quantization)
• SIMD-accelerated ANN search
• Built-in search_filtered and remove (native, no SQLite workarounds)
• Index persistence via save/load — no rebuild on restart

Changes

Feature Gates

turboquant = llm-kernel-vector-index + rusqlite + llm-kernel/embedding  (new default)
vector    = hnsw_rs  (fallback, no TurboQuant deps)

Verification

790 tests passed (turboquant feature)
clippy: 0 warnings
fmt: clean
release build: OK

Notes

• [patch.crates-io] redirects llm-kernel to git — temporary until llm-kernel-vector-index is published on crates.io
• Test dimensions changed 3→8 (TurboQuant requires dim % 8 == 0)
• hnsw_rs remains available as fallback via vector feature