Full/hierarchical Leiden: singleton local moves + multi-level aggregation

[ronaldmannak]

Motivation

The current LeidenCommunityDetector is a deliberately simplified single-level Leiden: local moving compares only stay/neighbor moves (never the empty-singleton candidate), and there's no multi-level aggregation. This was documented as a non-goal in the initial port (see the in-code note in Leiden.swift). Implementing it properly improves partition quality on connected graphs.

Scope

• Add the empty-singleton move to local-moving scoring (score Σ_tot = 0), so a connected over-merge can split back out when it improves modularity. Requires supporting community ids beyond 0..<n (e.g. dictionary-backed Σ_tot).
• Add multi-level aggregation: collapse communities into super-nodes and repeat local moving until modularity stops improving (classic Louvain/Leiden levels).
• Preserve determinism (stable ordering, sorted iteration) and the existing LeidenConfig knobs.
• Add a quality/regression test (e.g. a known graph where multi-level finds a better partition than single-level), plus the existing determinism tests must still pass.

Files (parallel-safe)

Sources/GraphRAG/Graph/Leiden.swift, Tests/GraphRAGTests/LightRAGTests.swift (or a new LeidenTests.swift). Isolated from LightRAG/retrieval work.

Acceptance criteria

• Deterministic across runs; leidenIsDeterministic and disconnected-cluster tests still pass.
• A test demonstrates multi-level/singleton refinement improving modularity vs. the single-level result.
• Modularity is non-decreasing across levels.
• CI green.

Note: best validated with the example/benchmark from the runnable-example issue. Context: follow-up to PR #2.