Summary

• Dynamic iteration bounds — Phase 2 runs 8-50 iterations (was fixed 25), Phase 1 runs 5-30 (was fixed 15). Easy queries converge early and stop, hard queries get more exploration budget.
• Multi-signal convergence detection — Three independent stop signals: top-k node stability, reward variance stability, and confidence threshold. Any one triggers early exit after min_iterations floor.
• Branch pruning — Internal nodes with 3+ visits and avg reward below 0.25 get pruned along with their entire subtree. Select/expand skip pruned branches, no more wasting LLM calls on irrelevant sections.
• Exploration constant decay — UCB1's C parameter decays from 2.0 → 0.5 (linear or cosine) over the search lifetime. Explore broadly first, exploit high-scoring paths later.
• Batch parallel simulation (virtual loss trick) — Selects K=4 nodes per iteration using virtual visits for diversity (AlphaGo-style), fires K LLM calls in parallel via ThreadPoolExecutor, then backpropagates real rewards. Cuts wall-clock search time by ~4x without changing search quality.
• SearchStats transparency — New dataclass tracks iterations used, convergence reason, pruned branches, coverage %, mean reward, and variance. Wired through QueryResult and ChatMessage for full observability.
• Full backward compatibility — MCTS_ADAPTIVE=false falls back to exact prior behavior: fixed iteration counts, static exploration constant, original early stopping heuristic.

Files changed

Test plan

• pytest tests/ — 25 tests pass
• Manual test with MCTS_ADAPTIVE=true (default) — verify console shows convergence info and reduced iteration counts
• Manual test with MCTS_ADAPTIVE=false — verify identical behavior to previous version
• Manual test with large multi-doc folder — verify parallel batch simulation + per-doc stats collection
• Verify MCTS_SIMULATION_BATCH_SIZE=1 degrades gracefully to sequential mode