Remove vector cache; read vectors directly from disk

[vindwi]

Summary

Removes the in-memory VectorCache layer from HierarchicalNSW. Layer-0 vectors are now always fetched directly from disk through the existing vector fetcher; the cache that previously sat in front of those reads is gone.

In-memory upper-layer vectors (dataUpperLayer_) are unaffected — they're still served zero-copy. The old CacheReadView helper is renamed to ZeroCopyView since it no longer has anything to do with a cache.

Changes

• Delete src/hnsw/vector_cache.h and all VectorCache usage in hnswalg.h:
  • cache init in both constructors and the load path
  • cache lookup/insert in getDataByInternalId and getDataByInternalIdBatch
  • the 1024 striped cache locks (vectorCacheLocks_, getCacheMutex)
  • slot invalidation on the update path
  • cache memory in getApproxSizeGB
  • getCache() accessor
• Remove settings NDD_VECTOR_CACHE_PERCENTAGE / NDD_VECTOR_CACHE_MIN_BITS (and their DEFAULT_*)
• Docs: drop the vector-cache description from memory_management.md

Net: −369 lines.

Verification

• clang++ -std=c++20 -fsyntax-only of hnswalg.h (and all transitive includes) passes cleanly.
• Confirmed no remaining references to VectorCache / vector_cache_ / VECTOR_CACHE_* anywhere outside third_party/.

Note / trade-off

This trades memory for I/O: every layer-0 distance computation that previously hit the cache now performs a disk fetch (single or batched MDBX read). Search-heavy workloads on large indices will do more disk reads per query. Flagging in case a benchmark pass is wanted before merge.

🤖 Generated with Claude Code