feat(graph): personalized PageRank primitive [HELD]

[RioPlay]

HELD — do not merge. Draft for review/tracking. The PPR primitive is sound and tested, but not wired into ask: on the assembly_ab eval it tied direct relevance ordering on the 1-hop cases the harness covers, so wiring it is not yet justified. Hold pending a multi-hop assembly eval that can actually exercise PPR's multi-hop salience propagation.

What

aden-graph::personalized_pagerank — a Personalized PageRank primitive over the knowledge graph.

• personalized_pagerank<N: GraphNode, E: GraphEdge>(graph, seed: &HashMap<String, f32>) -> HashMap<String, f32>; power-iteration core (DAMPING = 0.85, EPS = 1e-6, MAX_ITERS = 100), no RNG → deterministic.
• 5 unit tests; assembly_ab extended to a 3-way structural vs relevance vs PPR comparison.
• No production-path wiring — cmd_ask is unchanged (the wiring was applied, then reverted).

Why held

The point of PPR is multi-hop salience (a node relevant because it sits near several relevant nodes). The current assembly_ab is largely 1-hop, where PPR and direct relevance coincide — hence the tie. Merging an unproven ranking into the production path would violate the gate discipline; the primitive lands behind the eval instead.

Validation

cargo test --workspace green; PPR unit tests pass. Inert in production (not called from any command).

Stack

Based on feat/dense-prose-foundation. Not part of the merge sequence.