perf(vector/search): HNSW `id_to_index: AHashMap<u64, usize>` — `usize` is 8 bytes when u32 would suffice

[mosuka]

Round-3 perf push sub-issue (tracked under umbrella #536).

[S] HNSW id_to_index: AHashMap<u64, usize> — usize is 8 bytes when u32 would suffice

• Where: laurus/src/vector/index/hnsw/graph.rs:21 —
  id_to_index: AHashMap<u64, usize>.
• Current behavior: Two pointer-width values per entry, 24 bytes including HashMap
  overhead.
• Suggested direction: AHashMap<u64, u32> (or IntMap after a nohash switch).
  Saves 4 bytes / entry; at 10 M nodes that's 40 MB.
• Risk / scope: Trivial. u32::try_from(...).expect("doc id fits in u32") shim.

---

ID: VS-38 — see ~/.claude/tasks/laurus/20260523_perf_round3_audit/task_list.md for the full Round-3 issue list.

[mosuka]

Obsolete after #853 (PR #854): the search path no longer holds any doc_id → slot map — OrdinalHnswGraph addresses nodes by u32 ordinals directly, the ordinal → doc_id direction is a plain array, and the doc_id → ordinal direction is a binary search used only on cold paths. The remaining doc_id-keyed maps live in the writer's build path (out of scope here by design, see #853).