perf(vector/index): intern field names in memory as (u64, u16) + per-segment dictionary (#859)

[mosuka]

Summary

PR-B, the final piece of #633 (the on-disk dictionary landed in #860/#861): readers now keep one shared Arc<[Arc<str>]> field-name dictionary per segment and (doc_id, u16) pairs per record, instead of one heap String per record per retained structure.

• FieldInterner (format.rs): v3 segments use the header dictionary (fixed, ids validated); v1/v2 segments synthesize it at load in first-appearance order — legacy files get the same interned representation, and nothing branches on version after load.
• Readers ×3: vector_ids: Vec<(u64, u16)>; IVF cluster_to_vectors likewise (its eager path used to retain three String copies per record — the worst case found in the Phase 0 lifecycle sweep). Per-field caches and pool build inputs keep their String shapes (transient clones only; nothing per-record is retained).
• VectorStorage::get/contains take (doc_id, &str), and the OnDemand arm resolves the name against the segment dictionary by linear scan (1–3 entries in practice) — the per-candidate field_name.to_string() on the mmap-default f32 paths is gone structurally; OnDemand offsets are keyed (u64, u16).
• IVF searcher: probe results stay (u64, u16), the field filter resolves once per search and compares u16 per candidate; names rehydrate only for emitted top-k hits.
• Public trait surface unchanged: vector_ids(), iterators, and the other boundary methods rehydrate from the dictionary at return time. The VectorStorage method signatures changed (pub but not part of the reader/writer traits; no external callers).

RAM accounting (structural, deterministic)

A retained per-record String costs ~24B header + heap payload + allocator overhead, ×2 retained copies (HNSW/Flat) or ×3 (IVF eager); the interned pair costs 16B inline with zero per-record heap. At 1M records with short names that is roughly 90–135MB → a few MB of id-bearing structures.

Deviation from the plan, recorded honestly: pool build inputs keep the (u64, String, …) shape — the per-record Strings there are transient (consumed by per-field grouping, not retained), so converting the pipelines was all churn for no retained-memory gain.

Testing

• lib 1219/1219 (+ pq-fastscan 1225/1225); all 21 vector/hnsw/filter integration files green; 3 new FieldInterner/resolve_field_id unit tests; fmt / clippy (1.95 and 1.97 parity, workspace + wasm target) / wasm32 check clean.
• Eager gate bench neutral: Deletion Search del0 33.4µs / del10 29.3µs — inside the post-perf(vector/hnsw): segment-local u32 ordinal graph + LVS1 v2 format (#853) #854 cross-session envelope (33.0–37.4 / 29.0–31.7µs). The eager hot loop never touched these maps, so neutrality is the expected and required result; the structural wins live on the mmap/OnDemand paths (the ±6–19% bench class — claimed structurally per the perf(vector/hnsw): segment-local u32 ordinal graph + LVS1 v2 format (sub-issue of #686) #853/chore(vector/bench): resolve the mmap/OnDemand HNSW 50k bench class verdict after #853 #856 measurement policy).

Closes #859
Refs #633