perf(vector/index): HNSW per-level fixed-stride neighbour packing during build (M0 / M known up front)

[mosuka]

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

[S] HNSW per-level fixed-stride neighbour packing during build (M0 / M known up front)

• Where: laurus/src/vector/index/hnsw/writer.rs:42-56,107-114.
• Current behaviour: each per-level neighbour list is a
  variable-length Vec<u64>; growth triggers reallocations during
  pruning even though we know the max length per level (M0 or M).
• Why it might be a bottleneck: extra Vec headers per layer
  (24 B), and the allocator pressure during pruning cycles. With M=16
  and 16 layers (max_level cap), 32 layers/node x N nodes of Vec
  headers is real RAM.
• Suggested direction: use a single flat
  Vec<SmallVec<[u32; M0]>> per node, or per-layer
  [u32; M0] arrays packed inside a global arena. Pruning then
  truncates instead of dropping a Vec. Pairs with Fix examples #3.
• Risk / scope: small, contained to writer; saves both heap and
  CPU during build.

---

---

ID: VI-23 — see ~/.claude/tasks/laurus/20260523_perf_round3_audit/task_list.md for the full Round-3 issue list.