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

run.py

@@ -365,6 +365,9 @@ def main() -> int:
     duplicate_waits = 0
     jev_errors = 0
     saw_a_run = False
+    policy_context = brain.PolicyContext()
+    pending_action = None
+    pending_since = time.monotonic()
 
     for step in range(1, args.steps + 1):
         try:
@@ -428,9 +431,9 @@ def main() -> int:
             print(json.dumps(obs, indent=2)[:900])
             return finish(1, "stuck")
 
-        # Only wait when our own action actually landed and the game is still
-        # animating. If the action was rejected, fall through and pick a
-        # different one instead of waiting out the stuck counter.
+        # Acceptance does not prove completion. The policy context handles
+        # pending selections and purchases. Rejections allow another proposal
+        # instead of waiting out the stuck counter.
         #
         # NOTE: this check now happens AFTER deciding, and only suppresses a
         # REPEATED action. Waiting on "state unchanged" alone blocked
@@ -458,7 +461,7 @@ def main() -> int:
             # must not inherit the previous step's answers in the log.
             client.last = None
         try:
-            decision = brain.decide(obs, client, deck_snapshot)
+            decision = brain.decide(obs, client, deck_snapshot, context=policy_context)
             # A procedural action or animation wait does not establish model
             # recovery. Only a successful model response resets the budget.
             if client is not None and client.last is not None:
@@ -478,7 +481,7 @@ def main() -> int:
             # the same failing request -- measured, a DNS blip re-raised out of
             # the "fallback" and killed the session.
             try:
-                decision = brain.decide(obs, None, deck_snapshot)
+                decision = brain.decide(obs, None, deck_snapshot, context=policy_context)
             except Exception as inner:  # noqa: BLE001
                 trace({"step": step, "event": "fallback_error", "error": str(inner)[:160]})
                 print(f"[{step:03d}] fallback also failed: {inner}")
@@ -499,6 +502,13 @@ def main() -> int:
             print(json.dumps(obs, indent=2)[:800])
             return finish(1, "no_decision")
 
+        # Bound unresolved actions even if unrelated observation fields change.
+        if policy_context.pending is not pending_action:
+            pending_action = policy_context.pending
+            pending_since = time.monotonic()
+        if pending_action is not None and time.monotonic() - pending_since > args.stuck_seconds:
+            return finish(1, "pending_action_timeout")
+
         # Between turns there is nothing to do but look again.
         if decision.action == "__wait__":
             waits += 1
@@ -508,8 +518,8 @@ def main() -> int:
 
         # Suppress DUPLICATE actions on an unchanged state, but only for a
         # bounded number of reads. Proposing a DIFFERENT action is always
-        # allowed, which is what makes select-then-embark and multi-purchase
-        # shops work. The bound matters because some actions legitimately need
+        # allowed here, but cannot bypass a pending policy action above.
+        # The bound matters because some actions legitimately need
         # repeating (multi-line Ancient dialogue) and some transitions are just
         # slow -- waiting forever on those stalls the run.
         action_key = (decision.action, json.dumps(decision.params, sort_keys=True))
@@ -543,6 +553,7 @@ def main() -> int:
                    "action": decision.action, "error": str(exc)[:200]})
             return finish(1, "action_error")
 
+        policy_context.record_result(obs, decision, accepted=result.ok)
         if not result.ok:
             print(f"[{step:03d}] action rejected: {result.message}")
             trace({"step": step, "event": "action_rejected",