feat(search): a full-text index over chunks and a global search palette

[mrsibe]

What does this PR do?

Adds chunks_fts (FTS5/BM25) and a Ctrl/Cmd+K search palette that shows literal and semantic results as two separate signals. This is the sparse index #77 will fuse with dense.

Why?

Both chat retrieval and the search box were vector-only. Embeddings are measurably bad at exact tokens — model names, ids, symbols, numbers — so there was no way to find a literal string, and a note could not be searched at all. #96 also builds the index #77 needs for BM25/hybrid, so the two share an index but not a ranking.

Related issue

Fixes #96
Related to #154, #77

What changed

• chunks_fts (FTS5). ensureChunksFts() creates it and backfills missing rows on every call, so a library indexed before FTS existed — or a crash between the chunk write and the FTS write — heals itself.
• Maintained with the chunks: written in the ingestion pipeline, deleted with the document, so re-index and delete stay consistent.
• KnowledgeService.searchText() returns the same SearchResult shape as the vector path, so provenance and the jump-to-source chain are not duplicated.
• Injection-proof queries: user text is quoted term by term before MATCH, so -, NEAR(, col:value, a AND b or an unbalanced quote return no rows instead of a syntax error — a syntax error reads as "no results", not "bad query".
• SearchPalette (Ctrl/Cmd+K): matching text and related passages as labelled groups, never fused. A hit opens the reader at its passage (the locator travels with the result).
• KnowledgeSearchResult gained an optional locator, and the preload API a searchText method.

How was this tested?

• npm run typecheck — passes.
• npm test — 391 pass, including FTS5 integration against the bundled SQLite: BM25 ordering, notebook isolation, document-scoped filter, operator-looking queries not throwing, snippet extraction.
• npm run check:design — no violations.
• npm run build — passes.
• npm run eval — output byte-identical to baseline-v1.5.json (FTS does not touch dense retrieval).
• electron . --smoke-test — PASS, 27 checks, including a new one: a literal search finds the term after indexing, and finds exactly one row after a re-index.

Not verified

• The palette was not clicked in a running app. The FTS path is covered by unit tests and the smoke test; the palette markup only by typecheck and the design guard.

Checklist

• I have reviewed my own changes.
• npm run typecheck passes.
• npm run build passes.
• I have tested the affected user workflow. (FTS via smoke; palette noted above.)
• I have not included unrelated changes.
• I have updated documentation when necessary.

Desktop / build changes

• Not applicable (the FTS table is created at runtime with IF NOT EXISTS; no drizzle migration).