fix: improve search quality and query deduplication

[gzenz]

Summary

• FTS5 multi-word AND logic: Queries now use "graph" AND "store" instead of phrase matching, so "graph store" finds GraphStore and related nodes
• Deduplicated query results: callers_of/callees_of/inheritors_of no longer return duplicate nodes when multiple call-site edges exist between the same pair
• Ambiguous query auto-resolution: Bare-name queries with multiple matches auto-resolve to the production function when exactly one non-test candidate exists
• Test function deprioritization: Search results apply 0.5x score penalty to test functions so production code ranks higher
• search_nodes FTS5 fast path: GraphStore.search_nodes() now tries FTS5 first, falling back to LIKE only when FTS5 returns no results
• Composite edge index: v6 migration adds idx_edges_composite on (kind, source_qualified, target_qualified, file_path, line) for faster upsert_edge

Changed files

File	Change
code_review_graph/graph.py	search_nodes rewritten: FTS5 fast path with LIKE fallback
code_review_graph/search.py	FTS5 AND logic for multi-word queries, test score penalty
code_review_graph/tools/query.py	Deduplication via seen_qn set, ambiguous auto-resolution
code_review_graph/migrations.py	v6 migration with composite edge index

Test plan

• All 589 tests pass
• Multi-word search "graph store" returns GraphStore (was empty before)
• callers_of with multiple call sites returns unique nodes
• Bare "build" query resolves to production build() not test_build()

[tirth8205]

Good search quality improvements. Note: the v6 migration in this PR conflicts with the v6 migration that's already on main. You'll need to rebase and renumber to v8 (v7 is taken by PR #127).

Please rebase on latest main.