refactor(rag): reduce retrieval to a vanilla vector baseline

Split the PR so this branch carries only what a working RAG needs: on-demand
embedding model, source schema + legacy/incompatibility migration, attach →
chunk → embed, vector-only retrieval (top-K by cosine + per-file cap +
neighbor expansion), context assembly, answer-time citation attribution and
the sources UI.

Everything that improves retrieval quality moves to a stacked follow-up
branch: keyword (FTS5/BM25) indexing, RRF fusion, MMR, adaptive-k, term
coverage, and Polish no-answer detection. hybridRetrieval/rankFusion/
keywordIndex and their tests are removed here; messageSources now calls the
vanilla retrieve().

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Author: kfaracik

__tests__/VectorStoreContext.test.tsx

@@ -32,10 +32,6 @@ jest.mock('../utils/embeddingModelMigration', () => ({
   migrateEmbeddingModelIfNeeded: jest.fn().mockResolvedValue(undefined),
 }));
 
-jest.mock('../database/keywordIndex', () => ({
-  ensureKeywordIndex: jest.fn().mockResolvedValue(undefined),
-}));
-
 jest.mock('../utils/embeddingModel', () => ({
   isEmbeddingModelDownloaded: jest.fn().mockResolvedValue(true),
 }));