Add JAX PPO ladder v2 expert pass

tests/lost_cities_jax/test_ppo_stack.py

@@ -13,6 +13,7 @@ from lost_cities_jax.opponents import (
     discard_only_action,
     heuristic_balanced_action,
     heuristic_cautious_action,
+    heuristic_expert_action,
 )
 from lost_cities_jax.ppo import (
     JaxPPOConfig,
@@ -65,7 +66,12 @@ def tiny_config(tmp_path) -> JaxPPOConfig:
 def test_static_opponents_return_legal_actions():
     state = reset(jax.random.PRNGKey(0))
     mask = np.asarray(legal_action_mask(state), dtype=bool)
-    for fn in [discard_only_action, heuristic_balanced_action, heuristic_cautious_action]:
+    for fn in [
+        discard_only_action,
+        heuristic_balanced_action,
+        heuristic_cautious_action,
+        heuristic_expert_action,
+    ]:
         action = int(fn(state, jnp.int32(0), jax.random.PRNGKey(1)))
         assert mask[action]
 
@@ -112,6 +118,66 @@ def test_cautious_uses_own_board_when_playing_as_p1():
     assert hand[hand_slot] == playable
 
 
+def test_expert_opens_strong_ev_hand():
+    strong_cards = [
+        _hs_card(0, 0),
+        _rank_card(0, 7),
+        _rank_card(0, 8),
+        _rank_card(0, 9),
+        _rank_card(0, 10),
+        _rank_card(1, 2),
+        _rank_card(2, 2),
+        _rank_card(3, 2),
+    ]
+    state = _manual_state(p0_hand=strong_cards, to_move=0)
+    action = int(heuristic_expert_action(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)]
+    assert place_type == PLAY
+    assert hand[hand_slot] // CARDS_PER_COLOR == 0
+
+
+def test_expert_rejects_weak_new_open():
+    weak_cards = [
+        _rank_card(0, 2),
+        _rank_card(0, 3),
+        _rank_card(0, 4),
+        _rank_card(1, 2),
+        _rank_card(2, 2),
+        _rank_card(3, 2),
+        _rank_card(4, 2),
+        _rank_card(4, 3),
+    ]
+    state = _manual_state(p0_hand=weak_cards, to_move=0)
+    action = int(heuristic_expert_action(state, jnp.int32(0), jax.random.PRNGKey(1)))
+    _, place_type, _ = [int(x) for x in decode_action(jnp.asarray(action))]
+    assert place_type != PLAY
+
+
+def test_expert_avoids_discarding_immediately_useful_opponent_card():
+    dangerous = _rank_card(0, 8)
+    safe = _rank_card(4, 2)
+    state = _manual_state(
+        p0_hand=[
+            dangerous,
+            safe,
+            _rank_card(1, 2),
+            _rank_card(1, 3),
+            _rank_card(2, 2),
+            _rank_card(2, 3),
+            _rank_card(3, 2),
+            _rank_card(3, 3),
+        ],
+        p1_board=[_rank_card(0, 7)],
+        to_move=0,
+    )
+    action = int(heuristic_expert_action(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)]
+    if place_type != PLAY:
+        assert hand[hand_slot] != dangerous
+
+
 @pytest.mark.parametrize(
     "policy_name", ["discard_only", "heuristic_balanced", "heuristic_cautious"]
 )
@@ -128,6 +194,23 @@ def test_static_policy_duplicate_mirror_score_diff_is_zero(policy_name):
     assert result["wins"] == result["losses"]
 
 
+def test_expert_duplicate_mirror_regression_stats():
+    result = evaluate_static_mirror(
+        "heuristic_expert",
+        games=1000,
+        duplicate=True,
+        shuffle_bank_seed=20260704,
+        batch_games=1000,
+    )
+    assert result["games"] == 2000
+    assert abs(result["mean_score_diff"]) <= 1.0e-6
+    assert result["wins"] == result["losses"]
+    assert result["max_steps_rate"] < 0.05
+    assert 2.0 <= result["opened_colors_per_game"] <= 3.5
+    assert 0.30 <= result["play_action_rate"] <= 0.60
+    assert result["mean_game_length"] <= 70.0
+
+
 def test_gate3_checkpoint_duplicate_self_mirror_score_diff_is_zero():
     checkpoint = Path(GATE3_CHECKPOINT)
     if not checkpoint.exists():
@@ -189,6 +272,10 @@ def _rank_card(color: int, rank: int) -> int:
     return color * CARDS_PER_COLOR + rank + 1
 
 
+def _hs_card(color: int, slot: int = 0) -> int:
+    return color * CARDS_PER_COLOR + slot
+
+
 def _manual_state(
     *,
     p0_hand: list[int] | None = None,