perf(vector/index): stop persisting the HNSW graph as redundant manual u64s; pack as CSR + variable-byte deltas

[mosuka]

Round-3 perf push sub-issue (tracked under umbrella #535).

[M] Stop persisting the HNSW graph as redundant manual u64s; pack as CSR + variable-byte deltas

• Where: laurus/src/vector/index/hnsw/writer.rs:1259-1297
  (write path) and :417-485 (read path).

• Current behaviour: graph dump is node_count u64 followed by
  per-node [doc_id u64, layer_count u32, per layer [neighbour_count u32, neighbours u64*]]. Every edge takes 8 bytes even though
  neighbour IDs are local to the segment and almost always <2^32. No
  compression at all.

• Why it might be a bottleneck: I/O size is doubled vs Lucene's
  packed graph; cold-start mmap loads more pages than necessary. For a
  10 M-node segment with avg ~24 neighbours per node, the graph dump is
  ~2 GB at u64.

• Reference precedent: Lucene99HnswVectorsFormat writes layer-0 as
  [delta(n) varint]* per node, sorted neighbours, plus a sparse upper
  layer dump. Qdrant uses bincode with u32 IDs over a length-prefixed
  layout.

• Proposed layout (LVS2 graph section):

  per layer:
    offsets : Vec<u64> length n+1     // mmap-friendly random access
    edges   : Vec<u32>                // (sorted within node, delta-varint
                                      //  optional later)
  

• Suggested direction: convert IDs to u32 (already proposed for
  CSR); store sorted-within-node neighbours so a future v3 can introduce
  delta-varint codes without breaking compatibility; align offsets
  array to 8 B for direct mmap. Combine with Add parallel search #1 and Add parallel index #2.

• Risk / scope: format break; needs LVS2 magic; pairs naturally
  with the in-memory CSR change.

---

---

ID: VI-09 — see ~/.claude/tasks/laurus/20260523_perf_round3_audit/task_list.md for the full Round-3 issue list.

[mosuka]

Substantially addressed by #853 (PR #854): the LVS1 v2 graph block stores u32 ordinals and drops the per-node u64 doc id entirely — roughly halving the on-disk graph, which was the bulk of this issue's redundancy claim. The search-side CSR mirror this issue paired with was measured NO-GO in #679 (contiguity does not pay on random-hop traversal). The residual idea (variable-byte deltas over ordinals) buys marginal bytes at per-load decode cost; reopen a scoped issue if segment disk size becomes a driver.