perf(lexical/search): read facets from DocValues, not the stored document (#597)

[mosuka]

Summary

FacetCollector::collect_doc (facet.rs) fetched reader.document(doc_id) for every collected hit, decoding the variable-width rkyv stored-fields blob and cloning a Document just to read the facet field values via document.get(field).

The right data structure for facet aggregation — DocValues — already exists end to end:

• the writer stores every field into DocValues (InvertedIndexWriter::upsert_analyzed_document),
• the reader exposes get_doc_value(field, doc_id) / has_doc_values(field) (both overridden by the segment and multi-segment readers),
• and pub type FieldValue = crate::data::DataValue;, so get_doc_value returns exactly the DataValue the collector already matches on.

Since DocValues and the stored document derive from the same stored_fields, facet counts are identical.

Change

• collect_doc now reads each facet field from DocValues and skips reader.document() entirely when every facet field has a DocValues column (the common case). A field that lacks DocValues transparently falls back to the stored document (the synthetic error fallback is preserved on that path), so behaviour and counts are unchanged; readers that don't implement DocValues (has_doc_values defaults to false) keep the exact pre-existing path.
• The value → facet-path extraction is factored into a shared push_path_components helper used by both paths.
• Per-field DocValues availability is doc-independent, so it is resolved once and cached on the collector (field_has_dv) rather than re-probing the lock-guarded reader per hit. Phase 2 (interning + counter bumps) is unchanged.

Out of scope (follow-up): emitting per-field facet-ord columns at index time + dense Vec<u64> ordinal counters (the issue's larger "proposed" rewrite, which stacks with the columnar DocValues work in #547 / #688).

Tests

New unit tests in facet.rs (deterministic, via a configurable mock reader):

• facet_docvalues_counts_match_stored_doc — DocValues-path counts equal the stored-document-path counts for the same corpus (flat + hierarchical).
• facet_docvalues_skips_document_fetch — when all facet fields have DocValues, document() is never called (the mock panics if it is).
• facet_falls_back_to_document_without_docvalues — fields without DocValues are still counted via the stored document.
• facet_mixed_docvalues_and_stored — one field from DocValues, one from the stored document; both counted.

Verification

• cargo clippy -p laurus --all-targets -- -D warnings — clean
• cargo fmt --check — clean
• cargo test -p laurus --lib — 1108 passed (+4)
• markdownlint-cli2 (en/ja faceting.md) — 0 errors

Benchmark (facet_bench)

The bench mock now exposes DocValues and each document carries non-facet payload fields (title/body), modelling the asymmetry between a whole-document decode and a single-field DocValues read. Measured with git stash isolating facet.rs (before/after):

case	1k	10k	100k
flat_single	−42.5%	−29.2%	−17.7%
multi_field	−27.1%	−35.7%	−13.2%
hierarchical	−19.7%	−31.9%	−25.2%

All statistically significant (p < 0.05). The in-memory mock only models the Document clone avoidance; in production document() also decodes the variable-width rkyv blob (every stored field, not just facet fields) plus I/O, so the real-world win is expected to be larger.

Docs: added a "Performance" section to laurus/faceting.md (en/ja).

Closes #597