fix: honor active embed model in query vector lookups

[kuishou68]

Summary

• use the active llm embed model for query-time vector lookups instead of the hardcoded default alias
• keep hybrid and vector-only query paths aligned with the model used to build embeddings
• add regression coverage for vector and hybrid query routing

Closes #497

Testing

• corepack pnpm exec vitest run test/store.test.ts -t "vectorSearchQuery uses the active llm embed model for vector lookups" --reporter=dot
• corepack pnpm exec vitest run test/store.test.ts -t "hybridQuery uses the active llm embed model for precomputed vector lookups" --reporter=dot
• corepack pnpm build