feat: add gameplay stats and restore Zenith rollout depth
Build and Push Docker Container / build-and-push (push) Successful in 3m6s

- Add read-only gameplay statistics with filters and opponent comparisons.
- Report move-trace latency percentiles and external timeout counts.
- Pair engine latency with the preceding turn to separate timing stages.
- Adjust Zenith rollout deepening for its larger timeout reserve.
- Bump Zenith to 1.2.0 and document the fix and snake-stats command.
- Cover statistics, trace aggregation, rollout depth, and verdict budgets.
This commit is contained in:
2026-11-05 08:07:15 +01:00
parent a2414f0b3d
commit 5e121c20cf
10 changed files with 498 additions and 7 deletions
@@ -1,5 +1,8 @@
import unittest
from time import perf_counter
from unittest import mock
from snakes.engine.survival_search import CompactSurvivalSearch
from snakes import SnakeBuilder, get_snake_version
from snakes.strategies.prism import PrismBattleSnake_GPT_5_6_Sol
@@ -14,8 +17,8 @@ class TestZenithRegistration(unittest.TestCase):
snake = ZenithBattleSnake_ClaudeOpus5_5()
self.assertEqual(snake.name, "ZenithBattleSnake")
self.assertEqual(snake.version, "1.1.0")
self.assertEqual(get_snake_version("ZenithBattleSnake_ClaudeOpus5_5"), "1.1.0")
self.assertEqual(snake.version, "1.2.0")
self.assertEqual(get_snake_version("ZenithBattleSnake_ClaudeOpus5_5"), "1.2.0")
self.assertIsInstance(snake, PrismBattleSnake_GPT_5_6_Sol)
self.assertIsInstance(
SnakeBuilder.build("ZenithBattleSnake_ClaudeOpus5_5"),
@@ -80,5 +83,56 @@ class TestZenithVerdicts(unittest.TestCase):
self.assertEqual(verdicts, {})
def _rollout_depth(snake, remaining_ms: float) -> int:
"""Depth requested by a 3-snake rollout with ``remaining_ms`` left."""
snake._get_bb(11, 11)
snake._survival_search_context = None
enemies = [
{"body": [{"x": 8, "y": 8}, {"x": 8, "y": 9}, {"x": 8, "y": 10}]},
{"body": [{"x": 2, "y": 8}, {"x": 2, "y": 9}, {"x": 2, "y": 10}]},
]
with mock.patch.object(CompactSurvivalSearch, "search_selected", return_value=0.0) as search:
snake._future_rollout_bonus(
move="up", safe_moves={"up": {"x": 5, "y": 3}},
my_body=[{"x": 5, "y": 2}, {"x": 5, "y": 1}, {"x": 5, "y": 0}],
other_snakes=enemies, food_set=set(), is_constrictor=False,
width=11, height=11, enemy_can_grow={},
deadline=perf_counter() + remaining_ms / 1000,
)
return search.call_args.kwargs["depth"]
class TestZenithRolloutDepth(unittest.TestCase):
def test_gate_is_lowered_by_the_extra_timeout_reserve(self):
zenith = ZenithBattleSnake_ClaudeOpus5_5()
prism = PrismBattleSnake_GPT_5_6_Sol()
extra = zenith._get_timeout_buffer_ms() - prism._get_timeout_buffer_ms()
self.assertEqual(zenith._rollout_deepen_ms, prism._rollout_deepen_ms - extra)
def test_three_snake_rollout_deepens_after_same_elapsed_time_as_prism(self):
# Zenith's budget is 50 ms shorter, so 220 ms left equals 270 ms for Prism.
self.assertEqual(_rollout_depth(ZenithBattleSnake_ClaudeOpus5_5(), 220), 3)
self.assertEqual(_rollout_depth(PrismBattleSnake_GPT_5_6_Sol(), 270), 3)
self.assertEqual(_rollout_depth(PrismBattleSnake_GPT_5_6_Sol(), 220), 2)
def test_low_budget_still_caps_rollout_depth(self):
self.assertEqual(_rollout_depth(ZenithBattleSnake_ClaudeOpus5_5(), 150), 2)
class TestZenithVerdictBudget(unittest.TestCase):
def test_verdict_pass_skips_when_only_reserve_remains(self):
snake = ZenithBattleSnake_ClaudeOpus5_5()
verdicts = snake._duel_verdicts(
safe_moves={"up": {"x": 0, "y": 2}, "right": {"x": 1, "y": 1}},
my_body=[{"x": 0, "y": 1}], enemy={"body": [{"x": 4, "y": 4}]},
my_health=100, food_set=set(), hazard_set=set(), hazard_damage=15,
hazard_count={}, previous_hazard_set=set(), width=5, height=5,
deadline=perf_counter() + 0.020,
)
self.assertEqual(verdicts, {})
if __name__ == "__main__":
unittest.main()