feat(snake): add adaptive adversarial search

- Share duel search contexts and transpositions across candidate moves.
- Add aspiration windows, principal variation ordering, and body caches.
- Model simultaneous multiplayer responses with a compact beam rollout.
- Adapt search depth and response breadth to the remaining deadline.
- Add a deterministic arena benchmark with optional JSON reporting.
- Expose search metrics, document benchmarking, and bump Prism to 1.2.0.

README.md

@@ -68,9 +68,12 @@ Allowed values: `safe`, `balanced`, `aggressive`.
 
 ## PrismBattleSnake_GPT_5_6_Sol
 `PrismBattleSnake_GPT_5_6_Sol` is a separate snake that keeps Apex's strategy while
-accelerating hot spatial operations with a Python-integer bitboard engine. Its
+accelerating hot spatial operations with a Python-integer bitboard engine. It
-filename, class, and registry key include the model name, while its public
+also shares duel transpositions across candidate moves, uses principal-variation
-Battlesnake API name remains `PrismBattleSnake`.
+ordering and aspiration windows, and runs a compact adversarial multiplayer
+rollout with simultaneous enemy responses. Its filename, class, and registry
+key include the model name, while its public Battlesnake API name remains
+`PrismBattleSnake`.
 
 Run it with:
 ```sh
@@ -87,6 +90,17 @@ python scripts/benchmark_snakes_from_db.py \
 The benchmark opens SQLite read-only and reports mean, median, p95, and maximum
 move latency. Increase `--samples` for a broader but slower comparison.
 
+Run the deterministic CI-friendly arena benchmark without a gameplay database:
+```sh
+just bench-snake-arena positions=100
+```
+
+It reports latency, reached minimax depth, and move disagreements between Apex
+and Prism. Add `output=data/arena-report.json` to save a machine-readable report.
+For representative strategy evaluation, provide recorded positions to
+`scripts/benchmark_snake_arena.py --database /path/to/gameplay.sqlite3`, then run
+paired seeded games with the local Battlesnake CLI to measure win rate.
+
 ## Compact gameplay database
 New gameplay turns use normalized storage: the turn row stores food, hazards,
 move, and thinking data once; snake identity is stored once per game in

justfile

@@ -62,6 +62,14 @@ bench-best-snake iterations="1000":
 
   PYTHONPATH="{{justfile_directory()}}" python "{{justfile_directory()}}/tests/bench_best_battle_snake.py" --iterations "{{iterations}}"
 
+bench-snake-arena positions="100" output="":
+  #!/usr/bin/env bash
+  set -euo pipefail
+
+  args=(--positions "{{positions}}")
+  if [ -n "{{output}}" ]; then args+=(--json-output "{{output}}"); fi
+  PYTHONPATH="{{justfile_directory()}}" python "{{justfile_directory()}}/scripts/benchmark_snake_arena.py" "${args[@]}"
+
 build-battlesnake-cli:
   #!/usr/bin/env bash
   set -euo pipefail

scripts/benchmark_snake_arena.py

@@ -0,0 +1,117 @@
+#!/usr/bin/env python3
+"""Compare snake decisions and latency on deterministic synthetic positions.
+
+For outcome/win-rate tournaments use the local Battlesnake CLI. This harness is
+fast enough for CI and detects move disagreements, crashes, and latency changes.
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import sys
+from pathlib import Path
+from statistics import mean, median
+from time import perf_counter
+
+sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
+
+from scripts.benchmark_snakes_from_db import load_states, percentile
+from server.GameBoard import GameBoard
+from snakes import SnakeBuilder
+from tests.bench_best_battle_snake import build_game_state
+
+def synthetic_states(count: int) -> list[tuple[dict, dict]]:
+  states: list[tuple[dict, dict]] = []
+  for index in range(count):
+    payload = build_game_state()
+    payload["game"]["id"] = f"arena-{index}"
+    payload["turn"] = 20 + index
+    # Rotate food deterministically to exercise routing without creating
+    # invalid bodies or relying on private/real network values.
+    payload["board"]["food"] = [
+      {"x": 1 + index % 3, "y": 9},
+      {"x": 9, "y": 1 + (index // 3) % 3},
+    ]
+    states.append((payload["board"], {
+      "game_id": payload["game"]["id"],
+      "source": "custom", "map": "standard",
+      "ruleset": payload["game"]["ruleset"], "turn": payload["turn"],
+      "you": payload["you"],
+    }))
+  return states
+
+def evaluate(name: str, states: list[tuple[dict, dict]]) -> tuple[list[str], dict]:
+  moves: list[str] = []
+  durations: list[float] = []
+  depths: list[int] = []
+  for index, (board_data, metadata) in enumerate(states):
+    snake = SnakeBuilder.build(name)
+    game_id = f"arena-{name}-{index}-{metadata['game_id']}"
+    board = GameBoard(
+      game_id=game_id, width=board_data["width"], height=board_data["height"],
+      ruleset=metadata["ruleset"], source=metadata["source"],
+      map=metadata["map"], snake_class=snake,
+    )
+    board.read_game_data({
+      "game": {
+        "id": game_id, "ruleset": metadata["ruleset"],
+        "source": metadata["source"], "map": metadata["map"], "timeout": 500,
+      },
+      "turn": metadata["turn"], "board": board_data, "you": metadata["you"],
+    })
+    started = perf_counter()
+    moves.append(snake.choose_move(board))
+    durations.append((perf_counter() - started) * 1000.0)
+    history = snake.get_history() if hasattr(snake, "get_history") else []
+    if history:
+      depths.append(int(history[-1].get("minimax_depth_reached", 0)))
+  return moves, {
+    "snake": name, "positions": len(states),
+    "mean_ms": mean(durations), "median_ms": median(durations),
+    "p95_ms": percentile(durations, 0.95), "max_ms": max(durations),
+    "mean_minimax_depth": mean(depths) if depths else 0.0,
+  }
+
+def main() -> None:
+  parser = argparse.ArgumentParser()
+  parser.add_argument("--snake", action="append", default=[])
+  parser.add_argument("--database")
+  parser.add_argument("--positions", type=int, default=100)
+  parser.add_argument("--stride", type=int, default=997)
+  parser.add_argument("--json-output")
+  args = parser.parse_args()
+
+  states = (
+    load_states(args.database, max(1, args.positions), max(1, args.stride))
+    if args.database else synthetic_states(max(1, args.positions))
+  )
+  if not states:
+    raise SystemExit("No benchmark positions found")
+
+  names = args.snake or ["ApexBattleSnake", "PrismBattleSnake_GPT_5_6_Sol"]
+  move_sets: dict[str, list[str]] = {}
+  reports: list[dict] = []
+  for name in names:
+    moves, report = evaluate(name, states)
+    move_sets[name] = moves
+    reports.append(report)
+    print(
+      f"{name}: mean={report['mean_ms']:.3f} ms, "
+      f"p95={report['p95_ms']:.3f} ms, max={report['max_ms']:.3f} ms, "
+      f"depth={report['mean_minimax_depth']:.2f}"
+    )
+  baseline = names[0]
+  disagreements = {
+    name: sum(a != b for a, b in zip(move_sets[baseline], move_sets[name]))
+    for name in names[1:]
+  }
+  if disagreements:
+    print(f"Move disagreements versus {baseline}: {disagreements}")
+
+  payload = {"reports": reports, "baseline": baseline, "disagreements": disagreements}
+  if args.json_output:
+    Path(args.json_output).write_text(json.dumps(payload, indent=2) + "\n")
+
+if __name__ == "__main__":
+  main()

snakes/PrismBattleSnake_GPT_5_6_Sol.py

@@ -26,23 +26,27 @@ Key speedups:
   S12: Duel minimax uses tuple bodies and bitboard move generation.
   S13: Iterative deepening reuses a transposition table and move-order hints.
   S14: Candidate duel moves and enemy replies resolve on the same root turn.
+  S15: Candidate moves share one duel transposition/search context per turn.
+  S16: Compact adversarial multiplayer rollout advances plausible enemy replies.
+  S17: Rollout memoization and adaptive depth spend time on ambiguous positions.
 """
 
 from __future__ import annotations
-from typing import Any
+
 from time import perf_counter
 
+from server.GameBoard import GameBoard
 from snakes.ApexBattleSnake import ApexBattleSnake
 from snakes.bitboard import BitBoard
 from snakes.bitboard_duel_search import BitboardDuelSearch
-from server.GameBoard import GameBoard
+from snakes.compact_survival_search import CompactSurvivalSearch
 
 # Direction offsets for coord-dict → tuple conversion
 _DIR_DELTAS = ((0, 1), (0, -1), (-1, 0), (1, 0))
 _DIR_NAMES = ("up", "down", "left", "right")
 
 class PrismBattleSnake_GPT_5_6_Sol(ApexBattleSnake):
-  VERSION = "1.1.0"
+  VERSION = "1.2.0"
 
   def __init__(self) -> None:
     super().__init__()
@@ -60,6 +64,11 @@ class PrismBattleSnake_GPT_5_6_Sol(ApexBattleSnake):
     self._enemy_attack_danger: int = 0      # tiles where enemy len >= our len
     self._enemy_attack_opportunity: int = 0 # tiles where enemy len < our len
 
+    # Shared per-turn search contexts. Candidate moves overlap heavily, so
+    # rebuilding their transposition tables wastes most iterative-deepening work.
+    self._duel_search_context: BitboardDuelSearch | None = None
+    self._survival_search_context: CompactSurvivalSearch | None = None
+
   # ── BitBoard accessor ────────────────────────────────────────────────────
 
   def _get_bb(self, width: int, height: int) -> BitBoard:
@@ -78,6 +87,8 @@ class PrismBattleSnake_GPT_5_6_Sol(ApexBattleSnake):
 
   def choose_move(self, game_data: GameBoard) -> str:
     bb = self._get_bb(game_data.get_width(), game_data.get_height())
+    self._duel_search_context = None
+    self._survival_search_context = None
 
     # S9: precompute enemy body / tail / attack bitboards for survival tree
     other_snakes = game_data.get_other_snakes()
@@ -128,7 +139,17 @@ class PrismBattleSnake_GPT_5_6_Sol(ApexBattleSnake):
     self._enemy_attack_danger = enemy_attack_danger
     self._enemy_attack_opportunity = enemy_attack_opportunity
 
-    return super().choose_move(game_data)
+    move = super().choose_move(game_data)
+    history = self.get_history()
+    if history:
+      thinking = history[-1]
+      if self._duel_search_context is not None:
+        thinking["prism_duel_nodes"] = self._duel_search_context.nodes
+        thinking["prism_duel_cache_hits"] = self._duel_search_context.cache_hits
+      if self._survival_search_context is not None:
+        thinking["prism_rollout_nodes"] = self._survival_search_context.nodes
+        thinking["prism_rollout_cache_hits"] = self._survival_search_context.cache_hits
+    return move
 
   # ── S1: Bitboard flood fill ──────────────────────────────────────────────
 
@@ -260,14 +281,16 @@ class PrismBattleSnake_GPT_5_6_Sol(ApexBattleSnake):
     self, food_set: set, hazard_set: set, hazard_count: dict,
     hazard_damage: int, width: int, height: int, deadline: float | None,
   ) -> BitboardDuelSearch:
-    return BitboardDuelSearch(
+    if self._duel_search_context is None:
-      board=self._get_bb(width, height),
+      self._duel_search_context = BitboardDuelSearch(
-      food=food_set,
+        board=self._get_bb(width, height),
-      hazards=hazard_set,
+        food=food_set,
-      hazard_count=hazard_count,
+        hazards=hazard_set,
-      hazard_damage=hazard_damage,
+        hazard_count=hazard_count,
-      deadline=deadline,
+        hazard_damage=hazard_damage,
-    )
+        deadline=deadline,
+      )
+    return self._duel_search_context
 
   def _minimax_candidate_id(
     self, my_body: list, enemy_body: list, my_target: tuple[int, int],
@@ -281,13 +304,19 @@ class PrismBattleSnake_GPT_5_6_Sol(ApexBattleSnake):
       food_set, hazard_set, hazard_count, hazard_damage,
       width, height, deadline,
     )
+    adaptive_depth = max_depth
+    remaining = self._remaining_ms(deadline)
+    if remaining > 250:
+      adaptive_depth = min(7, max_depth + 1)
+    elif remaining < 120:
+      adaptive_depth = min(max_depth, 2)
     return search.search_candidate(
       my_body=my_body,
       enemy_body=enemy_body,
       my_target=my_target,
       my_health=my_health,
       enemy_health=enemy_health,
-      max_depth=max_depth,
+      max_depth=adaptive_depth,
       previous_hazards=previous_hazard_set if previous_hazard_set is not None else hazard_set,
     )
 
@@ -331,7 +360,51 @@ class PrismBattleSnake_GPT_5_6_Sol(ApexBattleSnake):
       previous_hazards=previous_hazard_set if previous_hazard_set is not None else hazard_set,
     )
 
-  # ── S9: Optimised survival tree (bitboard-native) ────────────────────────
+  # ── S16/S17: compact adversarial survival rollout ───────────────────────
+
+  def _future_rollout_bonus(
+    self, move: str, safe_moves: dict, my_body: list, other_snakes: list,
+    food_set: set, is_constrictor: bool, width: int, height: int,
+    enemy_can_grow: dict, deadline: float | None,
+  ) -> float:
+    pos = safe_moves.get(move)
+    if pos is None:
+      return -250.0
+    # Duel minimax already advances the opponent exactly. Keep the much faster
+    # bitboard-native solo rollout here instead of paying for the same response
+    # model twice. Constrictor and multiplayer still use adversarial rollouts.
+    if len(other_snakes) == 1 and not is_constrictor:
+      return super()._future_rollout_bonus(
+        move, safe_moves, my_body, other_snakes, food_set, is_constrictor,
+        width, height, enemy_can_grow, deadline,
+      )
+    if self._survival_search_context is None:
+      remaining = self._remaining_ms(deadline)
+      enemy_branch = 2 if len(other_snakes) <= 2 and remaining > 100 else 1
+      self._survival_search_context = CompactSurvivalSearch(
+        board=self._get_bb(width, height),
+        food=food_set,
+        is_constrictor=is_constrictor,
+        deadline=deadline,
+        branch=self._planning_branch,
+        enemy_branch=enemy_branch,
+        response_cap=8 if remaining > 150 else 4,
+      )
+    remaining = self._remaining_ms(deadline)
+    depth = min(self._planning_depth, 2 if len(other_snakes) > 1 else 3)
+    if remaining < 90:
+      depth = min(depth, 2)
+    elif remaining > 250 and len(other_snakes) <= 2:
+      depth = min(4, depth + 1)
+    raw = self._survival_search_context.search_selected(
+      my_body=my_body,
+      enemies=other_snakes,
+      target=(pos["x"], pos["y"]),
+      depth=depth,
+    )
+    return raw * 0.15
+
+  # ── S9: Optimised survival tree (compatibility fallback) ────────────────
 
   def _future_position_score(
     self, my_body: list, other_snakes: list, food_set: set, is_constrictor: bool,
@@ -420,7 +493,6 @@ class PrismBattleSnake_GPT_5_6_Sol(ApexBattleSnake):
 
     bb = self._bb
     w = bb.width
-    h = bb.height
     head = my_body[0]
     hx, hy = head["x"], head["y"]
     head_idx = hy * w + hx

snakes/__init__.py

@@ -11,7 +11,7 @@ SNAKE_REGISTRY = {
   "UltimateBattleSnake": "4.5.0",
   "ApexBattleSnake": "1.0.0",
   "SupremeBattleSnake_ClaudeOpus4_6": "1.0.0",
-  "PrismBattleSnake_GPT_5_6_Sol": "1.1.0",
+  "PrismBattleSnake_GPT_5_6_Sol": "1.2.0",
 }
 
 DEFAULT_SNAKE_CONFIG = {

snakes/bitboard.py

@@ -326,15 +326,15 @@ class BitBoard:
     # index when several foods are equally close, which can change contested-
     # food scoring and therefore the selected move.
     queue = [start_idx]
-    distances = [0]
     seen = start_bit
     cursor = 0
+    layer_end = 1
+    dist = 0
     w = self.width
     size = self.size
 
     while cursor < len(queue):
       cell = queue[cursor]
-      dist = distances[cursor]
       cursor += 1
       x = cell % w
       candidates = (
@@ -357,6 +357,8 @@ class BitBoard:
           return dist + 1, neighbor
         seen |= bit
         queue.append(neighbor)
-        distances.append(dist + 1)
+      if cursor == layer_end:
+        dist += 1
+        layer_end = len(queue)
 
     return None, None

snakes/bitboard_duel_search.py

@@ -2,9 +2,9 @@
 
 from __future__ import annotations
 
+from collections.abc import Iterable
 from dataclasses import dataclass
 from time import perf_counter
-from typing import Iterable
 
 from snakes.bitboard import BitBoard
 
@@ -47,9 +47,10 @@ class BitboardDuelSearch:
     self.hazard_stacks = {
       board.idx(x, y): count for (x, y), count in hazard_count.items()
     }
-    self.transposition: dict[tuple[DuelState, int], tuple[float, str]] = {}
+    self.transposition: dict[tuple[DuelState, int], tuple[float, str, int | None]] = {}
     self.killer_moves: dict[int, int] = {}
     self.history: dict[int, int] = {}
+    self._body_bits_cache: dict[Body, int] = {}
     self.nodes = 0
     self.cache_hits = 0
 
@@ -79,7 +80,11 @@ class BitboardDuelSearch:
     for depth in range(1, max_depth + 1):
       if self._out_of_time(5.0):
         break
-      value, completed = self._search(state, depth, -float("inf"), float("inf"))
+      window = 80.0 if completed_depth else float("inf")
+      alpha, beta = result - window, result + window
+      value, completed = self._search(state, depth, alpha, beta)
+      if completed and window != float("inf") and (value <= alpha or value >= beta):
+        value, completed = self._search(state, depth, -float("inf"), float("inf"))
       if not completed:
         break
       result = value
@@ -119,9 +124,13 @@ class BitboardDuelSearch:
     for depth in range(1, max_depth + 1):
       if self._out_of_time(5.0):
         break
-      value, completed = self._search_selected_move(
+      window = 80.0 if completed_depth else float("inf")
-        state, target_idx, depth, -float("inf"), float("inf")
+      alpha, beta = result - window, result + window
-      )
+      value, completed = self._search_selected_move(state, target_idx, depth, alpha, beta)
+      if completed and window != float("inf") and (value <= alpha or value >= beta):
+        value, completed = self._search_selected_move(
+          state, target_idx, depth, -float("inf"), float("inf")
+        )
       if not completed:
         break
       result = value
@@ -200,7 +209,7 @@ class BitboardDuelSearch:
     cached = self.transposition.get(cache_key)
     if cached is not None:
       self.cache_hits += 1
-      cached_value, bound = cached
+      cached_value, bound, preferred_move = cached
       if bound == "exact":
         return cached_value, True
       if bound == "lower":
@@ -217,9 +226,10 @@ class BitboardDuelSearch:
     if not enemy_moves:
       return self.WIN + depth, True
 
-    my_moves = self._ordered_moves(my_moves, state, depth, True)
+    my_moves = self._ordered_moves(my_moves, state, depth, True, preferred_move if cached is not None else None)
     enemy_moves = self._ordered_moves(enemy_moves, state, depth, False)
     best = -float("inf")
+    best_move: int | None = None
 
     for my_target in my_moves:
       worst = float("inf")
@@ -240,7 +250,9 @@ class BitboardDuelSearch:
           self.history[my_target] = self.history.get(my_target, 0) + depth * depth
           break
 
-      best = max(best, worst)
+      if worst > best:
+        best = worst
+        best_move = my_target
       alpha = max(alpha, best)
       if alpha >= beta:
         break
@@ -251,7 +263,7 @@ class BitboardDuelSearch:
       bound = "lower"
     else:
       bound = "exact"
-    self.transposition[cache_key] = (best, bound)
+    self.transposition[cache_key] = (best, bound, best_move)
     return best, True
 
   def _advance(self, state: DuelState, my_target: int, enemy_target: int) -> tuple[DuelState, float | None]:
@@ -308,7 +320,10 @@ class BitboardDuelSearch:
     """
     return list(self._iter_bits(self.board.neighbors_of(body[0])))
 
-  def _ordered_moves(self, moves: list[int], state: DuelState, depth: int, mine: bool) -> list[int]:
+  def _ordered_moves(
+    self, moves: list[int], state: DuelState, depth: int, mine: bool,
+    preferred: int | None = None,
+  ) -> list[int]:
     body = state.my_body if mine else state.enemy_body
     other = state.enemy_body if mine else state.my_body
     killer = self.killer_moves.get(depth)
@@ -320,8 +335,9 @@ class BitboardDuelSearch:
       food_bonus = 200.0 if (1 << target) & state.food_bits else 0.0
       space = self.board.flood_count(target, (self._body_bits(body[1:]) | self._body_bits(other[1:])) & ~(1 << target))
       center = -(abs(x - center_x) + abs(y - center_y))
+      preferred_bonus = 20_000.0 if target == preferred else 0.0
       killer_bonus = 10_000.0 if target == killer else 0.0
-      return killer_bonus + self.history.get(target, 0) + food_bonus + space * 2.0 + center, -target
+      return preferred_bonus + killer_bonus + self.history.get(target, 0) + food_bonus + space * 2.0 + center, -target
 
     # Our strongest-looking moves first; enemy ordering uses the same quality
     # estimate because dangerous enemy replies tend to gain space and food.
@@ -329,13 +345,25 @@ class BitboardDuelSearch:
 
   def _evaluate(self, state: DuelState) -> float:
     my_blocked = self._body_bits(state.my_body[1:]) | self._body_bits(state.enemy_body[1:])
-    my_space = self.board.flood_count(state.my_body[0], my_blocked)
+    my_head, enemy_head = state.my_body[0], state.enemy_body[0]
-    enemy_space = self.board.flood_count(state.enemy_body[0], my_blocked)
+    my_space = self.board.flood_count(my_head, my_blocked)
-    my_liberties = self.board.open_neighbor_count(state.my_body[0], my_blocked)
+    enemy_space = self.board.flood_count(enemy_head, my_blocked)
-    enemy_liberties = self.board.open_neighbor_count(state.enemy_body[0], my_blocked)
+    my_liberties = self.board.open_neighbor_count(my_head, my_blocked)
-    length_score = (len(state.my_body) - len(state.enemy_body)) * 18.0
+    enemy_liberties = self.board.open_neighbor_count(enemy_head, my_blocked)
-    health_score = (state.my_health - state.enemy_health) * 0.15
+    territory = self.board.territory(my_head, [enemy_head], my_blocked)
-    return (my_space - enemy_space) * 2.0 + (my_liberties - enemy_liberties) * 12.0 + length_score + health_score
+    my_tail_path = self.board.path_distance(my_head, state.my_body[-1], my_blocked)
+    enemy_tail_path = self.board.path_distance(enemy_head, state.enemy_body[-1], my_blocked)
+    tail_score = (12.0 if my_tail_path is not None else -24.0) - (12.0 if enemy_tail_path is not None else -24.0)
+    my_hazard = self._hazard_cost(my_head, state.previous_hazard_bits)
+    enemy_hazard = self._hazard_cost(enemy_head, state.previous_hazard_bits)
+    length_score = (len(state.my_body) - len(state.enemy_body)) * 20.0
+    health_score = (state.my_health - state.enemy_health) * 0.18
+    forced_score = (my_liberties > 1) * 10.0 - (enemy_liberties > 1) * 10.0
+    return (
+      (my_space - enemy_space) * 1.5 + territory * 1.2
+      + (my_liberties - enemy_liberties) * 14.0 + length_score + health_score
+      + tail_score + forced_score + (enemy_hazard - my_hazard) * 0.8
+    )
 
   def _hazard_cost(self, target: int, previous_hazard_bits: int) -> int:
     bit = 1 << target
@@ -351,11 +379,15 @@ class BitboardDuelSearch:
   def _tail_stacked(body: Body) -> bool:
     return len(body) >= 2 and body[-1] == body[-2]
 
-  @staticmethod
+  def _body_bits(self, body: Body) -> int:
-  def _body_bits(body: Body) -> int:
+    cached = self._body_bits_cache.get(body)
+    if cached is not None:
+      return cached
     bits = 0
     for cell in body:
       bits |= 1 << cell
+    if len(self._body_bits_cache) < 16_384:
+      self._body_bits_cache[body] = bits
     return bits
 
   @staticmethod

snakes/compact_survival_search.py

@@ -0,0 +1,263 @@
+"""Compact adversarial rollout for multiplayer Battlesnake positions."""
+
+from __future__ import annotations
+
+from itertools import product
+from time import perf_counter
+
+from snakes.bitboard import BitBoard
+
+Body = tuple[int, ...]
+EnemyBodies = tuple[Body, ...]
+StateKey = tuple[Body, EnemyBodies, int, int]
+
+class CompactSurvivalSearch:
+  """Small paranoid beam search with simultaneous enemy responses.
+
+  It is deliberately narrower than full multiplayer minimax: each enemy keeps
+  only its most dangerous replies and the combined response beam is capped.
+  This models moving opponents without exhausting the request deadline.
+  """
+
+  DEATH = -5000.0
+
+  def __init__(
+    self,
+    board: BitBoard,
+    food: set[tuple[int, int]],
+    is_constrictor: bool,
+    deadline: float | None,
+    branch: int,
+    enemy_branch: int = 2,
+    response_cap: int = 8,
+  ) -> None:
+    self.board = board
+    self.food_bits = board.set_to_bits(food)
+    self.is_constrictor = is_constrictor
+    self.deadline = deadline
+    self.branch = max(1, branch)
+    self.enemy_branch = max(1, enemy_branch)
+    self.response_cap = max(1, response_cap)
+    self.cache: dict[StateKey, float] = {}
+    self.body_bits_cache: dict[Body, int] = {}
+    self.nodes = 0
+    self.cache_hits = 0
+
+  def body_from_dicts(self, body: list[dict]) -> Body:
+    return tuple(self.board.idx(segment["x"], segment["y"]) for segment in body)
+
+  def search_selected(
+    self,
+    my_body: list[dict],
+    enemies: list[dict],
+    target: tuple[int, int],
+    depth: int,
+  ) -> float:
+    mine = self.body_from_dicts(my_body)
+    enemy_bodies = tuple(self.body_from_dicts(enemy["body"]) for enemy in enemies)
+    target_idx = self.board.idx(*target)
+    if not self.board.neighbors_of(mine[0]) & (1 << target_idx):
+      return self.DEATH
+    return self._selected_root(mine, enemy_bodies, self.food_bits, target_idx, depth)
+
+  def _selected_root(
+    self, mine: Body, enemies: EnemyBodies, food_bits: int, target: int, depth: int,
+  ) -> float:
+    replies = self._enemy_responses(enemies, mine, target, food_bits)
+    if not replies:
+      replies = [()]
+    worst = float("inf")
+    for response in replies:
+      if self._out_of_time():
+        break
+      child = self._advance(mine, enemies, target, response, food_bits)
+      if child is None:
+        value = self.DEATH
+      else:
+        next_mine, next_enemies, next_food = child
+        value = self._evaluate(next_mine, next_enemies)
+        if depth > 1 and value > self.DEATH:
+          value += self._search(next_mine, next_enemies, next_food, depth - 1) * 0.72
+      worst = min(worst, value)
+    return self._evaluate(mine, enemies) if worst == float("inf") else worst
+
+  def _search(self, mine: Body, enemies: EnemyBodies, food_bits: int, depth: int) -> float:
+    self.nodes += 1
+    if self._out_of_time() or depth <= 0:
+      return 0.0
+    key = (mine, enemies, food_bits, depth)
+    cached = self.cache.get(key)
+    if cached is not None:
+      self.cache_hits += 1
+      return cached
+
+    occupied = self._occupied(mine, enemies)
+    targets = list(self._iter_bits(self.board.neighbors_of(mine[0])))
+    ranked: list[tuple[float, int]] = []
+    for target in targets:
+      # Collision legality is finalized simultaneously because eating controls
+      # whether tails vacate.
+      ate = bool((1 << target) & food_bits)
+      own_tail_blocked = self.is_constrictor or ate
+      body_blocked = self._body_bits(mine if own_tail_blocked else mine[:-1])
+      enemy_blocked = 0
+      for enemy in enemies:
+        enemy_blocked |= self._body_bits(enemy[:-1] if not self.is_constrictor else enemy)
+      if (1 << target) & (body_blocked | enemy_blocked):
+        continue
+      free_space = self.board.flood_count(target, occupied & ~(1 << target))
+      ranked.append((free_space + (20 if ate else 0), target))
+    ranked.sort(reverse=True)
+
+    if not ranked:
+      return self.DEATH
+
+    best = self.DEATH
+    for _, target in ranked[:self.branch]:
+      replies = self._enemy_responses(enemies, mine, target, food_bits) or [()]
+      worst = float("inf")
+      for response in replies:
+        if self._out_of_time():
+          break
+        child = self._advance(mine, enemies, target, response, food_bits)
+        if child is None:
+          value = self.DEATH
+        else:
+          next_mine, next_enemies, next_food = child
+          value = self._evaluate(next_mine, next_enemies)
+          if depth > 1 and value > self.DEATH:
+            value += self._search(next_mine, next_enemies, next_food, depth - 1) * 0.72
+        worst = min(worst, value)
+      if worst != float("inf"):
+        best = max(best, worst)
+
+    if not self._out_of_time() and len(self.cache) < 16_384:
+      self.cache[key] = best
+    return best
+
+  def _enemy_responses(
+    self, enemies: EnemyBodies, mine: Body, my_target: int, food_bits: int,
+  ) -> list[tuple[int, ...]]:
+    if not enemies:
+      return []
+    choices: list[list[int]] = []
+    my_length_after = len(mine) + int(bool((1 << my_target) & food_bits))
+    for enemy in enemies:
+      ranked: list[tuple[float, int]] = []
+      for target in self._iter_bits(self.board.neighbors_of(enemy[0])):
+        ate = bool((1 << target) & food_bits)
+        enemy_length_after = len(enemy) + int(ate)
+        score = 0.0
+        if target == my_target:
+          score += 1000.0 if enemy_length_after >= my_length_after else -1000.0
+        tx, ty = self.board.coord(target)
+        mx, my = self.board.coord(my_target)
+        score -= abs(tx - mx) + abs(ty - my)
+        score += self.board.open_neighbor_count(target, self._occupied(mine, enemies)) * 3.0
+        score += 20.0 if ate else 0.0
+        ranked.append((score, target))
+      ranked.sort(reverse=True)
+      choices.append([target for _, target in ranked[:self.enemy_branch]])
+
+    responses: list[tuple[int, ...]] = []
+    for response in product(*choices):
+      responses.append(response)
+      if len(responses) >= self.response_cap:
+        break
+    return responses
+
+  def _advance(
+    self,
+    mine: Body,
+    enemies: EnemyBodies,
+    my_target: int,
+    enemy_targets: tuple[int, ...],
+    food_bits: int,
+  ) -> tuple[Body, EnemyBodies, int] | None:
+    my_ate = bool((1 << my_target) & food_bits)
+    next_mine = self._advance_body(mine, my_target, my_ate)
+    next_enemies = tuple(
+      self._advance_body(body, target, bool((1 << target) & food_bits))
+      for body, target in zip(enemies, enemy_targets)
+    )
+
+    # Body and self collisions after all tails have moved.
+    if my_target in next_mine[1:]:
+      return None
+    if any(my_target in enemy[1:] for enemy in next_enemies):
+      return None
+
+    surviving: list[Body] = []
+    for index, enemy in enumerate(next_enemies):
+      target = enemy[0]
+      dead = target in enemy[1:] or target in next_mine[1:]
+      dead = dead or any(
+        target in other[1:] for other_index, other in enumerate(next_enemies)
+        if other_index != index
+      )
+      if target == my_target:
+        if len(enemy) >= len(next_mine):
+          return None
+        dead = True
+      if not dead:
+        # Enemy/enemy head collisions remove equal-length snakes and the shorter.
+        for other_index, other in enumerate(next_enemies):
+          if other_index != index and target == other[0] and len(enemy) <= len(other):
+            dead = True
+            break
+      if not dead:
+        surviving.append(enemy)
+
+    eaten = (1 << my_target) if my_ate else 0
+    for body, target in zip(enemies, enemy_targets):
+      if (1 << target) & food_bits:
+        eaten |= 1 << target
+    return next_mine, tuple(surviving), food_bits & ~eaten
+
+  def _evaluate(self, mine: Body, enemies: EnemyBodies) -> float:
+    blocked = self._occupied(mine, enemies) & ~(1 << mine[0])
+    space = self.board.flood_count(mine[0], blocked)
+    liberties = self.board.open_neighbor_count(mine[0], blocked)
+    if liberties == 0 or space < len(mine):
+      return self.DEATH
+    enemy_pressure = 0.0
+    for enemy in enemies:
+      enemy_blocked = blocked & ~(1 << enemy[0])
+      enemy_space = self.board.flood_count(enemy[0], enemy_blocked)
+      enemy_liberties = self.board.open_neighbor_count(enemy[0], enemy_blocked)
+      enemy_pressure += max(0, 3 - enemy_liberties) * 18.0
+      if len(mine) > len(enemy):
+        enemy_pressure += max(0, 8 - enemy_space) * 8.0
+    return space * 1.9 + liberties * 32.0 + enemy_pressure - len(enemies) * 4.0
+
+  def _occupied(self, mine: Body, enemies: EnemyBodies) -> int:
+    occupied = self._body_bits(mine)
+    for enemy in enemies:
+      occupied |= self._body_bits(enemy)
+    return occupied
+
+  def _body_bits(self, body: Body) -> int:
+    cached = self.body_bits_cache.get(body)
+    if cached is not None:
+      return cached
+    bits = 0
+    for cell in body:
+      bits |= 1 << cell
+    if len(self.body_bits_cache) < 16_384:
+      self.body_bits_cache[body] = bits
+    return bits
+
+  def _advance_body(self, body: Body, target: int, ate: bool) -> Body:
+    if self.is_constrictor or ate:
+      return (target,) + body
+    return (target,) + body[:-1]
+
+  @staticmethod
+  def _iter_bits(bits: int):
+    while bits:
+      bit = bits & -bits
+      yield bit.bit_length() - 1
+      bits ^= bit
+
+  def _out_of_time(self) -> bool:
+    return self.deadline is not None and perf_counter() >= self.deadline

tests/snakes/test_PrismBattleSnake_GPT_5_6_Sol.py

@@ -3,9 +3,10 @@ from time import perf_counter
 
 from snakes import SnakeBuilder, get_snake_version
 from snakes.ApexBattleSnake import ApexBattleSnake
-from snakes.PrismBattleSnake_GPT_5_6_Sol import PrismBattleSnake_GPT_5_6_Sol
 from snakes.bitboard import BitBoard
 from snakes.bitboard_duel_search import BitboardDuelSearch
+from snakes.compact_survival_search import CompactSurvivalSearch
+from snakes.PrismBattleSnake_GPT_5_6_Sol import PrismBattleSnake_GPT_5_6_Sol
 
 class TestBitBoard(unittest.TestCase):
 
@@ -51,8 +52,8 @@ class TestPrismBattleSnake_GPT_5_6_Sol(unittest.TestCase):
     snake = PrismBattleSnake_GPT_5_6_Sol()
 
     self.assertEqual(snake.name, "PrismBattleSnake")
-    self.assertEqual(snake.version, "1.1.0")
+    self.assertEqual(snake.version, "1.2.0")
-    self.assertEqual(get_snake_version("PrismBattleSnake_GPT_5_6_Sol"), "1.1.0")
+    self.assertEqual(get_snake_version("PrismBattleSnake_GPT_5_6_Sol"), "1.2.0")
     self.assertIsInstance(SnakeBuilder.build("PrismBattleSnake_GPT_5_6_Sol"), PrismBattleSnake_GPT_5_6_Sol)
 
   def test_bitboard_primitives_match_apex(self):
@@ -156,6 +157,55 @@ class TestPrismBattleSnake_GPT_5_6_Sol(unittest.TestCase):
 
     self.assertIn(board.idx(1, 1), advanced)
 
+  def test_candidate_duel_searches_share_the_same_context(self):
+    snake = PrismBattleSnake_GPT_5_6_Sol()
+    kwargs = {
+      "my_body": [{"x": 0, "y": 1}, {"x": 0, "y": 0}],
+      "enemy_body": [{"x": 3, "y": 1}, {"x": 3, "y": 0}],
+      "food_set": set(), "hazard_set": set(),
+      "my_health": 100, "enemy_health": 100,
+      "hazard_damage": 15, "hazard_count": {},
+      "width": 4, "height": 3, "max_depth": 2,
+      "alpha": -1e9, "beta": 1e9, "deadline": perf_counter() + 1.0,
+    }
+
+    snake._minimax_candidate_id(my_target=(1, 1), **kwargs)
+    first_context = snake._duel_search_context
+    snake._minimax_candidate_id(my_target=(0, 2), **kwargs)
+
+    self.assertIs(snake._duel_search_context, first_context)
+    self.assertGreater(first_context.nodes, 0)
+
+  def test_compact_rollout_models_lethal_enemy_head_response(self):
+    board = BitBoard(3, 3)
+    search = CompactSurvivalSearch(
+      board=board, food=set(), is_constrictor=False,
+      deadline=perf_counter() + 1.0, branch=2,
+    )
+    mine = [{"x": 0, "y": 1}, {"x": 0, "y": 0}]
+    enemies = [{
+      "body": [{"x": 2, "y": 1}, {"x": 2, "y": 0}, {"x": 1, "y": 0}],
+    }]
+
+    value = search.search_selected(mine, enemies, (1, 1), depth=1)
+
+    self.assertEqual(value, search.DEATH)
+
+  def test_compact_rollout_reuses_transpositions(self):
+    board = BitBoard(5, 5)
+    search = CompactSurvivalSearch(
+      board=board, food=set(), is_constrictor=False,
+      deadline=perf_counter() + 1.0, branch=2,
+    )
+    mine = [{"x": 1, "y": 1}, {"x": 1, "y": 0}]
+    enemies = [{"body": [{"x": 3, "y": 3}, {"x": 3, "y": 4}]}]
+
+    search.search_selected(mine, enemies, (2, 1), depth=3)
+    hits_before = search.cache_hits
+    search.search_selected(mine, enemies, (2, 1), depth=3)
+
+    self.assertGreater(search.cache_hits, hits_before)
+
   def test_bitboard_duel_search_reuses_transpositions(self):
     board = BitBoard(5, 5)
     search = BitboardDuelSearch(