fix(policy): reconcile session-owned action memory with game observations

utils/audit_prototype.py

@@ -53,11 +53,6 @@ def observation(cards: list, *, enemies=None, energy=1, hp=80, block=0) -> dict:
     }
 
 
-def reset_policy() -> None:
-    # Each stored observation is independent. These are not ordered trajectories.
-    brain._reset_screen_guards("audit-reset")
-
-
 def synthetic_probes() -> dict:
     report = {}
     obs = observation([card(0)], enemies=[enemy("E0", 6, 0), enemy("E1", 6, 0)])
@@ -94,20 +89,19 @@ def synthetic_probes() -> dict:
             result = {"propagated": True}
     report["combat_model_error"] = {**result, "diagnostic": output.getvalue().strip()}
 
-    reset_policy()
     obs = {
         "state_type": "card_select",
         "card_select": {"screen_type": "upgrade", "prompt": "Choose a card to Upgrade.",
                         "cards": [card(5)], "can_confirm": False,
                         "can_cancel": True, "preview_showing": False},
     }
-    first = brain.decide(obs, None)
-    second = brain.decide(obs, None)
+    context = brain.PolicyContext()
+    first = brain.decide(obs, None, context=context)
+    second = brain.decide(obs, None, context=context)
     report["selection_without_execution"] = {
         "first": first.action, "second": second.action, "reason": second.reason,
         "actual_game_actions": 0,
     }
-    reset_policy()
 
     try:
         answer = jev.JevClient._parse({"answers": {"test": {"type": "noul"}}}, 0)["test"]
@@ -138,7 +132,6 @@ def corpus_audit() -> dict:
             raise ValueError("State path leaves dataset directory")
         obs = json.loads(gzip.decompress(path.read_bytes()))
         types[obs.get("state_type")] += 1
-        reset_policy()
         try:
             decision = brain.decide(obs, None)
             actions[decision.action if decision else "no_decision"] += 1
@@ -162,7 +155,6 @@ def corpus_audit() -> dict:
                 status_in_deck += 1
         except Exception as exc:
             exceptions.append({"path": row["path"], "error": f"{type(exc).__name__}: {exc}"})
-    reset_policy()
     return {
         "state_types": dict(types), "combat_phases": dict(phases), "actions": dict(actions),
         "exceptions": exceptions, "forced_block_when_covered": covered,
@@ -171,7 +163,6 @@ def corpus_audit() -> dict:
 
 
 def main() -> int:
-    reset_policy()
     report = {"synthetic": synthetic_probes(), "corpus": corpus_audit()}
     print(json.dumps(report, indent=2, sort_keys=True))
     # Successful audit execution is not a clean bill of health. Read the report.