Add gates 1-2 audit and repair workflow

tests/lost_cities_jax/test_gates.py

@@ -0,0 +1,19 @@
+from __future__ import annotations
+
+from lost_cities_jax.gates import load_gates_config
+
+
+def test_gates_config_loads_repair_settings():
+    cfg = load_gates_config("configs/jax_ppo/gates-1-2.yaml")
+    assert cfg.gate2_pass_threshold == 0.55
+    assert len(cfg.exploiters) == 3
+    assert cfg.repair.max_cycles == 3
+    assert cfg.repair.league_template == "configs/jax_ppo/league-v1.yaml"
+    assert cfg.repair.evaluation_games == 2000
+
+
+def test_gates_smoke_config_uses_smoke_repair_template():
+    cfg = load_gates_config("configs/jax_ppo/gates-1-2-smoke.yaml")
+    assert cfg.repair.max_cycles == 1
+    assert cfg.repair.league_template == "configs/jax_ppo/league-smoke.yaml"
+    assert cfg.repair.guard_max_steps_rate == 1.0

tests/lost_cities_jax/test_ppo_stack.py

@@ -10,10 +10,12 @@ import pytest
 from lost_cities_jax import legal_action_mask, reset
 from lost_cities_jax.engine import current_hand_sorted, decode_action, reset_from_order
 from lost_cities_jax.opponents import (
+    HeuristicExpertConfig,
     discard_only_action,
     heuristic_balanced_action,
     heuristic_cautious_action,
     heuristic_expert_action,
+    make_heuristic_expert_policy,
 )
 from lost_cities_jax.ppo import (
     JaxPPOConfig,
@@ -178,6 +180,76 @@ def test_expert_avoids_discarding_immediately_useful_opponent_card():
         assert hand[hand_slot] != dangerous
 
 
+def test_expert_cap2_blocks_third_open_color():
+    state = _manual_state(
+        p0_hand=[
+            _hs_card(2),
+            _rank_card(2, 7),
+            _rank_card(2, 8),
+            _rank_card(2, 9),
+            _rank_card(2, 10),
+            _rank_card(3, 2),
+            _rank_card(3, 3),
+            _rank_card(4, 2),
+        ],
+        p0_board=[_rank_card(0, 2), _rank_card(1, 2)],
+        to_move=0,
+    )
+    policy = make_heuristic_expert_policy(HeuristicExpertConfig(max_open_colors=2))
+    action = int(policy(state, jnp.int32(0), jax.random.PRNGKey(1)))
+    hand_slot, place_type, _ = [int(x) for x in decode_action(jnp.asarray(action))]
+    hand = [int(x) for x in current_hand_sorted(state, 0)]
+    opened_new_color = (
+        place_type == PLAY and state.col_len[0, hand[hand_slot] // CARDS_PER_COLOR] == 0
+    )
+    assert not bool(opened_new_color)
+
+
+def test_expert_cap3_allows_third_but_blocks_fourth_open_color():
+    cap3 = make_heuristic_expert_policy(HeuristicExpertConfig(max_open_colors=3))
+    third_state = _manual_state(
+        p0_hand=[
+            _hs_card(2),
+            _rank_card(2, 7),
+            _rank_card(2, 8),
+            _rank_card(2, 9),
+            _rank_card(2, 10),
+            _rank_card(3, 2),
+            _rank_card(3, 3),
+            _rank_card(4, 2),
+        ],
+        p0_board=[_rank_card(0, 2), _rank_card(1, 2)],
+        to_move=0,
+    )
+    action = int(cap3(third_state, jnp.int32(0), jax.random.PRNGKey(1)))
+    hand_slot, place_type, _ = [int(x) for x in decode_action(jnp.asarray(action))]
+    hand = [int(x) for x in current_hand_sorted(third_state, 0)]
+    assert place_type == PLAY
+    assert third_state.col_len[0, hand[hand_slot] // CARDS_PER_COLOR] == 0
+
+    fourth_state = _manual_state(
+        p0_hand=[
+            _hs_card(3),
+            _rank_card(3, 7),
+            _rank_card(3, 8),
+            _rank_card(3, 9),
+            _rank_card(3, 10),
+            _rank_card(4, 2),
+            _rank_card(4, 3),
+            _rank_card(4, 4),
+        ],
+        p0_board=[_rank_card(0, 2), _rank_card(1, 2), _rank_card(2, 2)],
+        to_move=0,
+    )
+    action = int(cap3(fourth_state, jnp.int32(0), jax.random.PRNGKey(1)))
+    hand_slot, place_type, _ = [int(x) for x in decode_action(jnp.asarray(action))]
+    hand = [int(x) for x in current_hand_sorted(fourth_state, 0)]
+    opened_new_color = (
+        place_type == PLAY and fourth_state.col_len[0, hand[hand_slot] // CARDS_PER_COLOR] == 0
+    )
+    assert not bool(opened_new_color)
+
+
 @pytest.mark.parametrize(
     "policy_name", ["discard_only", "heuristic_balanced", "heuristic_cautious"]
 )