• Fix infinite loop in GraphLayersBuilder::subgraph_connectivity when the selected entry point has no outgoing links on any layer. spent_budget is only incremented while iterating a point's outgoing links,
  so an isolated entry point leaves it stuck at 0 forever — the retry loop never observes spent_budget > SUBGRAPH_CONNECTIVITY_SEARCH_BUDGET. This was seen in production: the HNSW build thread spinning
  inside subgraph_connectivity → HNSWIndex::build → SegmentOptimizer::optimize, which in turn blocked the consensus thread (apply_entries waiting on the collection lock held by the optimizer), hanging
  the whole node.
• Break the retry loop when an iteration consumes zero budget. Sound because the graph is immutable across retries and the coin flip only gates following enumerated links, not enumerating them — if BFS
  enumerated zero edges once, it will every time. Return value in that case is 1.0 / points.len(), truthfully reflecting an isolated entry point.
• Add regression test test_subgraph_connectivity_isolated_entry_point_does_not_hang. Runs the call on a background thread with a 2s deadline so the test runner isn't wedged if the bug regresses. Fails in ~2s
  on the old code, passes in ~0.02s with the fix.

Test plan

• New regression test passes (cargo test -p segment --lib test_subgraph_connectivity_isolated_entry_point_does_not_hang)
• Full hnsw_index:: test suite passes (38/38), including test_graph_connectivity which exercises the normal, non-isolated path
• Manually verify on the affected deployment that optimizer-driven HNSW builds no longer hang