perf(vector/search): `vector_ids()` clones full `Vec<(u64, String)>` per call

[mosuka]

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

[S] vector_ids() clones full Vec<(u64, String)> per call

• Where: laurus/src/vector/index/hnsw/reader.rs:560-562 — Ok(self.vector_ids.clone()).
  Used in hnsw/searcher.rs:206 (unfiltered brute fallback) and elsewhere.
• Current behavior: 10 M × 32 B ≈ 320 MB per call. Unconditional.
• Suggested direction: Return Arc<[(u64, String)]> or refactor to iterator. Already
  done for doc_ids_for_field (perf(vector): cache vector_ids per field at reader load #405).
• Risk / scope: Small.

---

ID: VS-29 — see ~/.claude/tasks/laurus/20260523_perf_round3_audit/task_list.md for the full Round-3 issue list.

[mosuka]

Partial resolution via #859 (PR #862): the hot-path cost this issue describes is largely gone — VectorStorage::get/contains now take (doc_id, &str) (no per-call owned-key materialization), and internal callers operate on the interned (u64, u16) representation. vector_ids() itself still clones (now rehydrating from the shared dictionary, same per-call cost as before) and its trait-signature change remains out of scope; its remaining callers are the unfiltered brute-force scans and count().