fix(search): match multi-word FTS queries

[neuromechanist]

Summary

Papers (and all knowledge sources) were synced but never found at query time. Root cause: _sanitize_fts5_query() wrapped the entire query in quotes, forcing an exact consecutive-phrase FTS5 match. Every multi-word query the agent sends returned zero rows.

Proven on the live dev eeglab.db (1940 papers):

Query	Old (phrase)	New (OR + rank)
ICA EEGLAB	0	497
papers about ICA artifact removal	0	710

Change

_sanitize_fts5_query now: tokenize -> drop stopwords + operator chars -> quote each term individually -> OR them. Results keep the existing BM25 rank ordering, so the best matches surface first.

• Still injection-safe: each term is individually quoted, so FTS5 operators (AND/OR/NEAR/wildcards) and punctuation are inert.
• Identifier tokens (e.g. pop_runica, clean_rawdata) are preserved as adjacency phrases.
• Falls back to a safe quoted phrase when a query is all stopwords/symbols.
• Fixes all 6 search call sites (papers, discussions, FAQ, docstrings, search_all), not just papers.

Tests

• Updated TestFTS5Sanitization for the new OR-of-terms output + injection safety.
• Added regression tests: a multi-word NL query with non-adjacent terms, and a full question with stopwords, both now return matching papers.
• 293 passed across knowledge + tool suites; ruff clean.

Closes #305

[neuromechanist]

Review (Sonnet, pr-review-toolkit) + resolutions

Passes: FTS5 injection safety sound (regex strips all operator chars, each term individually quoted); fallback branch correct for all-stopword/all-symbol input; single-term stopword queries fall back safely; regression genuinely fixed with real SQLite (no mocks); no new OperationalError paths.

Important findings addressed:

1. Operator test was too weak (only checked start/end quotes). Strengthened test_sanitize_fts5_operators to assert every OR-separated term is individually quoted and no bare operator text remains.
2. Stopwords could silently drop meaningful terms. Removed list and use from _FTS_STOPWORDS (they double as EEGLAB/MATLAB content nouns, e.g. 'list channels'). Added test_sanitize_keeps_command_noun_terms. Grammar/boolean words (and/or/not/the/what...) stay dropped on purpose — that's the precision win.

No critical issues. 41 passed, ruff clean.