feat: surface per-neighbor edge counts in graph_neighbors

[Evanfeenstra]

What

Follow-up to #1459. Each neighbor returned by graph_neighbors now carries an edges map ({EDGE_TYPE: count}):

[
  {
    "ref_id": "def-456",
    "node_type": "Function",
    "name": "handle_request",
    "edge_type": "CALLS",
    "direction": "forward",
    "edges": { "CALLS": 8, "CONTAINS": 1 }
  }
]

Why

#1459 added edge counts to graph_search and graph_get, but not graph_neighbors — it hits a different backend path (expand=edges → prediction_service.get_node_edges). This closes that gap so an agent can see how connected each hop target is and which edge types it can follow next, and choose where to traverse without a follow-up call per neighbor.

How

• Sends include_edge_counts=true on the expand=edges request so Jarvis attaches each neighbor's counts inline — one call, no per-node round trips.
• Reads the per-node edges map from the neighbor nodes and surfaces it on each returned neighbor.
• Adds an optional namespace param to scope the counts correctly.

Dependency

Requires the matching backend change in stakwork/jarvis-backend#2976 (now extends include_edge_counts to the expand=edges path). Until it deploys, neighbors carry edges: {}.

Tests

tsc --noEmit clean; all 25 tests in toolsJarvis.test.ts pass.