perf(vector/index): read + write LVS v3 field-dictionary records in HNSW/Flat/IVF (#858)

[mosuka]

Summary

Second half of the #633 field-name dictionary (PR-A2+A3 of the plan there; the format layer landed in #860). The three vector index types now write LVS version-3 segments whose records carry a u16 field_id into the per-segment dictionary instead of the inline u32 len + UTF-8 name, and read v1/v2/v3 uniformly.

Readers

• The 8 record-parse branches (HNSW SQ eager/lazy, PQ, FastScan; Flat eager/lazy; IVF eager/lazy) collapse into the shared read_record_field(); record strides become version-dependent via record_prefix_size().
• Flat/IVF quant matches go by-ref — previously a partial move meant the parsed header did not survive to the parse loops (latent, harmless at v1/v2, load-bearing at v3).
• Latent bug fixed: the Flat and HNSW lazy paths computed the record-section seek by rebuilding a fresh header instead of measuring the parsed one — that reconstruction would have silently omitted the v3 dictionary.
• OnDemand offsets now point at the record payload, so VectorStorage::get seeks straight to the vector data instead of re-reading the variable-width name prefix on every access (net-negative LOC in storage.rs; a structural win for the mmap-default path).

Writers

• Per-segment dictionary built in first-appearance order over the exact emission order (HNSW doc_id-sorted, Flat buffer order, IVF cluster order — its existing flatten pass provides the walk for free); records emit u16 ids; HNSW's five header stamps go 2→3 and Flat/IVF stamp a version for the first time.
• The three load() re-parsers translate ids back for the doc_id-keyed in-memory state, so compaction/merge upgrade v1/v2 segments to v3 in place.

Compatibility

v1/v2 read paths are byte-identical (all pre-existing fixture tests pass unmodified). Older builds reading v3 get the clean range-gate IncompatibleFormat — the same deliberate forward break as #853/#854, documented in persistence.md.

Testing

• Byte-exact file-size assertion (the deterministic headline gate): total-size formula per component plus the v1-vs-v3 delta formula (n·(k+2) − dict bytes saved; −18.9%/record at the issue's k=32 example).
• v1→v3 upgrade-on-rewrite proof (writer loads a hand-built v1 fixture, rewrites, version bytes asserted, reader loads, search results identical); v3 hand-built fixture load + search (eager and lazy); out-of-range field_id and empty-dict-with-records rejected as corruption; multi-field and empty-segment cases.
• lib 1216/1216 (+ pq-fastscan 1222/1222); all 23 vector/hnsw/filter integration test files green; fmt / clippy (1.95 and 1.97 parity, workspace + wasm target) / wasm32 check clean; markdownlint clean.
• Eager gate bench neutral as required (deletion bench del0 33.2µs / del10 31.6µs — 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 touches names).
• Docs: module layout docs ×4, persistence.md and vector_indexing.md (en/ja).

Closes #858
Refs #633