fix(search): lexical BM25 review follow-up — camelCase query + fallback + docs

[HumanBean17]

Summary

Follow-up to #432 (BM25 lexical fallback via LadybugDB FTS), applying the /requesting-code-review findings that came in after #432 merged. The review found no Critical issues; this PR fixes the one real bug that slipped into master, plus the test/doc gaps.

The bug now in master (B1 — recall regression)

Pasted camelCase queries return []. Symbol.search_text is built from _split_identifier tokens, but the raw query was passed straight to QUERY_FTS_INDEX and LadybugDB FTS's tokenizer does not split camelCase — so search DistributionChunkService matched nothing. The heuristic fallback didn't save it (it also doesn't split camelCase). Verified empirically during review. This is the most common code-search query pattern (pasting a class name).

Fix: a shared build_fts_query (search_scoring) pre-splits the query the same way search_text is indexed, so index/query tokenization finally agree. Regression test added.

Other fixes

• B2/B3: degenerate queries ("", stopword-only) and empty BM25 results now fall back to the heuristic scan — preserves the role-ranked listing for degenerate queries and is defensive against stemming / selective-filter-thinning cases.
• C1: test_cap_advisory_silent_below_cap now forces the heuristic path (it was passing vacuously on the BM25 default).
• C2: added a Symbol.search_text schema-anchor test (guards against a silently empty search_text breaking all searches with zero test signal).
• Docs: swept 6 stale ontology 18 references (AGENT-GUIDE, ARCHITECTURE×2, DESIGN, CONFIGURATION×2) to 19; fixed the re-rank module attribution (search_lexical, not search_scoring); added package to the search_text composition.

Review push-back (for the record)

One reviewer asserted the LadybugDB constraints behind #432's defensive build-side code (DROP TABLE refused while an FTS index references it; Symbol DML needs LOAD EXTENSION FTS) "don't reproduce" on the shipped version. Both were re-proved empirically — the DROP refusal reproduced in a fresh probe; the DELETE-needs-LOAD was the exact failure behind 22 incremental-test fixes in #432. The defensive code stays.

Verification

• 21 lexical tests green (incl. the new camelCase regression + search_text anchor).
• Bench (~/jrag-bench/shopizer): query mean ~26ms post-fix (pre-split adds ~2ms; still ~3.3× faster than the heuristic-only baseline of 82.8ms); build phases unchanged.
• Diff is a focused +101/−18 across 7 files (branched from current master, so no unrelated revert noise).

🤖 Generated with Claude Code