Before improving the ranker, inspect the graph.

A change-impact system has at least two jobs: discover relevant relationships, then decide which consequences deserve attention. It is tempting to focus on the second because ranking produces a clean metric. But a missing edge is a different failure from a badly ranked edge.

As a general evaluation principle, missing evidence and poor decisions need different explanations. Keeping those questions separate makes a system easier to inspect and a result easier to interpret.

The research question is therefore more precise than whether a graph works. Which evidence source adds useful coverage? What does it cost? Once that evidence exists, does a more complex ranker improve on reachability or name matching? Those questions can produce different answers, and the record should preserve each one.

Evaluate what the system can see separately from how it orders what it sees.