Persist the dependency graph edge list (enables neighborhood/interaction-level risk hypotheses) #2992

Why

The temporal-crucible program falsified per-file, isolated, standing 'risk' (see #2991 and the #2982 record) and the open frontier is interaction-level hypotheses: a file's risk as a function of its graph neighbors' markers (concurrency markers meeting shared-state markers across an edge; allocation in file A with cleanup responsibility in file B; neighborhood marker load beyond own-file load). There is already a whisper of graph transmission in the data: the api_exposure spillover ripple on untouched files rides popularity/pagerank shifts.

The gap

The engine builds the full dependency graph in memory every scan (network_risk_sensor, GraphBuilder — pagerank, betweenness, in/out degree all computed from it) but persists only per-node degree summaries (popularity, import_count, internal_dependency_links, pagerank_score). The edge list is discarded. Every neighborhood hypothesis is therefore untestable from the scan DB.

Ask

Persist the per-snapshot edge list — e.g. an edge_data(repo_name, commit_hash, src_file_id, dst_file_id, edge_kind) table, or a compact adjacency JSON per snapshot. Constraints: keyed like the other tables ((repo_name, commit_hash), FK to file_data), tolerable size (curl-scale graphs are thousands of edges — trivial next to function_data), and written under the same --db-only path so history batches capture it. Note js/ts historically built no edges despite reporting dependency_links (comment at network_risk_sensor.py:153) — worth surfacing per-language edge coverage while at it.

Unblocks

INT-level hypotheses in the #2982 program (neighborhood marker load, cross-file discipline splits, marker×edge interactions), plus system-level graph statistics per snapshot (component structure over time). Related: #2991 (reframe), temporal-crucible#29 (history-enabled mode), #2982.