feat(discovery): parse SELF: report + score self_correct in make_turn_trace

proteus/game/runtime/_session_core.py

@@ -30,6 +30,7 @@ from proteus.game.metrics.persona import (
 )
 from proteus.game.metrics.rollout import optimal_rollout
 from proteus.game.runtime.trace import SessionTrace, TurnTrace
+from proteus.game.agents.parsing import extract_self_belief
 
 _ACTIONS = ["up", "down", "left", "right", "stay"]
 _DIRECTIONS = {"up", "down", "left", "right"}
@@ -245,6 +246,17 @@ def make_turn_trace(
             blocked=blocked,
         )
         reward_regret = ref_reward - model_reward
+    # Find-your-body discovery: parse the optional SELF: report from raw_text
+    # and score it against the scenario's (hidden) true body index. No-op for
+    # non-discovery scenarios (discovery_candidates() == 0).
+    n_candidates = scenario.discovery_candidates()
+    self_belief = extract_self_belief(raw_text, n_candidates) if n_candidates else None
+    true_index = scenario.discovery_true_index()
+    self_correct = (
+        (self_belief == true_index)
+        if (self_belief is not None and true_index is not None)
+        else None
+    )
     return TurnTrace(
         turn_idx=turn_idx,
         observation=observation,
@@ -278,6 +290,8 @@ def make_turn_trace(
         model_reward=model_reward,
         reward_regret=reward_regret,
         pressure=turn_pressure,
+        self_belief=self_belief,
+        self_correct=self_correct,
     )
 
 

tests/runtime/test_errand_discovery.py

@@ -33,3 +33,34 @@ def test_template_has_no_discovery():
     scen = get_scenario("template")()
     assert scen.discovery_candidates() == 0
     assert scen.discovery_true_index() is None
+
+
+from proteus.game.runtime import _session_core as core
+
+
+def test_make_turn_trace_parses_correct_self_report():
+    scen, game = _errand()
+    true = scen.true_body_index
+    tt = core.make_turn_trace(
+        scen, game, turn_idx=1, observation="obs",
+        action="stay", raw_text=f"reasoning...\nSELF: {true}\nACTION: stay",
+    )
+    assert tt.self_belief == true
+    assert tt.self_correct is True
+
+
+def test_make_turn_trace_marks_wrong_self_report():
+    scen, game = _errand()
+    wrong = (scen.true_body_index + 1) % 3
+    tt = core.make_turn_trace(
+        scen, game, turn_idx=1, observation="obs",
+        action="stay", raw_text=f"SELF: {wrong}\nACTION: stay",
+    )
+    assert tt.self_belief == wrong and tt.self_correct is False
+
+
+def test_make_turn_trace_no_self_report_is_none():
+    scen, game = _errand()
+    tt = core.make_turn_trace(scen, game, turn_idx=1, observation="o",
+                              action="stay", raw_text="ACTION: stay")
+    assert tt.self_belief is None and tt.self_correct is None