The newmm BFS graph traversal had O(2^n) worst-case behavior on highly ambiguous input — the same bug fixed in Python via PyThaiNLP/pythainlp#1319. Without a visited set, every neighbor is unconditionally re-enqueued, causing the queue to grow exponentially when the dictionary produces many overlapping matches.

Changes

• bfs_paths_graph — add visited: HashSet<CharacterIndex>: Only enqueue nodes not yet seen. Reduces worst-case from O(2^n) to O(V+E).

  // Before: every neighbor unconditionally enqueued
  if position != goal {
      current_queue.push_back((position, appended_path));
  }
  
  // After: skip already-visited nodes
  } else if !visited.contains(&position) {
      visited.insert(position);
      current_queue.push_back((position, appended_path));
  }

• one_cut commit-point branch — call graph.clear(): Previously only graph_size was reset to 0; stale edges from resolved segments accumulated across BFS calls. Graph is now cleared at each commit point.

• one_cut no-candidate branch — remove unnecessary edge insertion: The non-dictionary-match branch inserted a graph edge that was never needed (token is yielded directly). Replaced with graph_size = 0; graph.clear(), consistent with the commit-point branch.

• Regression test (test_newmm_ambiguous_performance): 250-character maximally ambiguous input (1–3 char overlapping words covering every TCC boundary) must complete within 1 s. With the fix it runs in ~10 ms.