MT3: Parallel collection scan (std::thread::scope)
- store.query_all() scans multiple collections in parallel
- Auto-skips for single-collection databases
- Used by FROM ANY queries and context search Tier 3

MT4: Parallel graph edge collection
- materialize_graph_lazy() collects edges from all collections in parallel
- Builds adjacency map after parallel collection phase

MT6: Parallel AI embedding persistence
- Batches >10 embeddings persisted in parallel threads
- Small batches (<= 10) stay sequential (less overhead)
- Error handling preserved — first error stops and reports

All using std::thread::scope — zero external dependencies.