Multi-hop (cat4) regressed 0.132→0.104 after merge fix

[StigNorland]

The #16 concept-merge fix improved overall graph score (0.293→0.331) and 4/5 categories, but multi-hop (cat4) regressed 0.132 → 0.104 on sample 0 (199 QA).

Hypothesis: the pre-fix mega-hubs (3 L1 nodes absorbing everything) accidentally bridged some 2-hop reasoning chains — any two entities were ~2 hops apart through a hub. With 688 properly-separated concepts, those incidental bridges are gone, so some multi-hop answers are now less reachable during context assembly.

To investigate:

• Inspect the cat4 questions that flipped graph 1→0 between results/baseline_s0_full.json and results/postfix_s0_graph.json.
• Likely the real fix is proper multi-hop traversal in get_context (and/or the bi-temporal fact edges from v2 [Epic] Graph model v2 — bi-temporal relational facts #18 giving real relational paths), not reintroducing hubs.

Low priority vs the inference/adversarial gains; tracking so it isn't lost. Refs #16, #18.