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

AGENTS.md

@@ -2,7 +2,7 @@
 
 ## Code map
 - `facts.py`: pure game-state parsing and arithmetic.
-- `brain.py`: decision policy; returns one `Decision` per observation.
+- `brain.py`: decision policy and session-owned `PolicyContext`; proposes one action per observation.
 - `run.py`: observe–decide–act loop, captures, and session attribution.
 - `sts2.py`: local game HTTP client. `jev.py`: TypeSafe model client and gates.
 - `migrate.py`: dataset migration and integrity checks.
@@ -25,6 +25,7 @@ Prefer integration/end-to-end checks; keep only essential regression tests and u
 
 ## Invariants
 - Execute one game action, then read a fresh observation. Card indices can change after each action.
+- Keep policy memory session-owned. Proposals do not record execution; reconcile accepted requests with fresh observations.
 - Compute arithmetic and legality in code, not in Jev. Confidence does not prove correctness.
 - Keep deterministic fallbacks usable with `client=None`. Test model paths with stubs.
 - Preserve session/step attribution. Run outcomes are not per-decision correctness labels.

CONTEXT.md

@@ -28,6 +28,16 @@ A proposed next action. A decision does not establish that the game received or
 **Action result**:
 The game's response to an attempted action. Acceptance alone does not establish that the expected change occurred.
 
+**Pending game action**:
+An accepted action request that has not yet been reconciled with a fresh observation.
+_Avoid_: Completed action
+
+**Accepted toggle**:
+An accepted request to change a card's selection status. It does not prove that the card is selected.
+
+**Screen flow**:
+Related game screens that form one interaction, such as opening and skipping a card reward.
+
 **Run deck**:
 The persistent collection of cards owned during a run, distinct from temporary combat cards and combat piles.
 _Avoid_: Treating all visible combat cards as the persistent deck.

brain.py

@@ -472,8 +472,6 @@ def card_reward_decision(obs: dict, client: JevClient | None, deck: dict | None)
         and skip_answer.yes
         and gate(skip_answer)
     ):
-        global _rewards_skipped_card
-        _rewards_skipped_card = True
         return Decision("skip_card_reward", {},
                         f"jev: skip all (noul={skip_answer.noul:.2f}); keep the deck lean",
                         "jev", skip_answer.noul)
@@ -485,7 +483,6 @@ def card_reward_decision(obs: dict, client: JevClient | None, deck: dict | None)
     # floor the deck is better off lean. This is the safety net that stops the
     # deck bloating when Jev is indifferent.
     if CARD_SKIP_POLICY == "combined" and can_skip and best_noul < CARD_PICK_THRESHOLD:
-        _rewards_skipped_card = True
         return Decision("skip_card_reward", {},
                         f"combined: jev said don't skip but best={best_noul:.2f} "
                         f"< {CARD_PICK_THRESHOLD}; keep the deck lean",
@@ -493,7 +490,6 @@ def card_reward_decision(obs: dict, client: JevClient | None, deck: dict | None)
 
     if best_key is None:
         if can_skip:
-            _rewards_skipped_card = True
             return Decision("skip_card_reward", {},
                             "no usable ratings; keeping the deck lean", "fallback")
         fallback_card = best_by_rarity(cards)
@@ -691,11 +687,6 @@ def map_decision(obs: dict, client: JevClient | None, deck: dict | None) -> Deci
                     f"jev chose {node.get('type')}", "jev", pick.confidence)
 
 
-_last_card_select_sig: str | None = None
-_card_select_picked = False
-_card_select_confirmed = False
-_card_select_chosen: list[int] = []
-
 # Deterministic fallbacks for a grid selection screen, used only when Jev does
 # not clear the threshold. Lower rank wins.
 #
@@ -770,11 +761,7 @@ def card_select_fallback(cards: list[dict], screen: str,
 
 def _card_select_pick(index: int, reason: str, source: str,
                       confidence: float | None = None) -> Decision:
-    """Emit a select_card and remember that this index has been toggled on."""
+    """Propose a toggle. Only an accepted request enters policy memory."""
-    global _card_select_picked
-    _card_select_picked = True
-    if index not in _card_select_chosen:
-        _card_select_chosen.append(index)
     return Decision("select_card", {"index": index}, reason, source, confidence)
 
 
@@ -794,23 +781,20 @@ def card_select_need(prompt: str) -> int:
     return int(match.group(1)) if match else 1
 
 
-def card_select_decision(obs: dict, client: JevClient | None, deck: dict | None) -> Decision:
+def card_select_decision(obs: dict, client: JevClient | None, deck: dict | None,
+                         context: PolicyContext) -> Decision:
     """
     Grid selection overlay: upgrade, transform, remove, choose-a-card.
 
     Two traps, both observed live:
       * `select_card` TOGGLES on grid screens. Calling it twice on one index
         deselects and freezes the screen.
-      * A preview left over from a desynced state makes `confirm_selection`
+      * A preview can make `confirm_selection` report ok without a visible
-        report ok while changing nothing. If the same screen reappears
+        change. The context waits for the screen to close instead of repeating
-        unchanged, reset with cancel_selection instead of confirming again.
+        or cancelling an action whose outcome is still unknown.
     """
-    global _last_card_select_sig, _card_select_picked, _card_select_confirmed
-
     cs = obs.get("card_select") or {}
-    sig = json.dumps(cs, sort_keys=True)
+    chosen = context.accepted_card_indices
-    repeated = sig == _last_card_select_sig
-    _last_card_select_sig = sig
 
     prompt = str(cs.get("prompt") or "Choose a card.")
     screen = str(cs.get("screen_type") or "")
@@ -821,37 +805,31 @@ def card_select_decision(obs: dict, client: JevClient | None, deck: dict | None)
     # selection with NO preview at all. And do NOT confirm early on a
     # multi-select: "Choose 5 cards to Remove" keeps can_confirm FALSE until all
     # five are chosen.
-    if len(_card_select_chosen) >= need:
+    if cs.get("can_confirm") and (len(chosen) >= need or cs.get("preview_showing")):
-        if cs.get("can_confirm"):
+        return Decision("confirm_selection", {}, "confirm the selection", "code")
-            if _card_select_confirmed:
+    if len(chosen) >= need:
-                return Decision("cancel_selection", {},
-                                "confirm did not apply; reset the screen", "code")
-            _card_select_confirmed = True
-            return Decision("confirm_selection", {},
-                            f"confirm {len(_card_select_chosen)}/{need} selected", "code")
         # Enough chosen but the game has not enabled confirm yet. Selecting more
         # would overshoot, so wait.
         return Decision("__wait__", {},
-                        f"{len(_card_select_chosen)}/{need} chosen; waiting for confirm",
+                        f"{len(chosen)}/{need} accepted toggles; waiting for confirm",
                         "code")
 
     cards = [c for c in (cs.get("cards") or []) if isinstance(c, dict)]
     if not cards:
         return Decision("cancel_selection", {}, "no cards to select", "code")
 
-    # `select_card` TOGGLES, so never re-select an index we already toggled on.
+    # `select_card` TOGGLES. Do not repeat an accepted request for this grid.
-    remaining = [c for c in cards if c.get("index") not in _card_select_chosen]
+    remaining = [c for c in cards if c.get("index") not in chosen]
     if not remaining:
         if cs.get("can_confirm"):
-            _card_select_confirmed = True
             return Decision("confirm_selection", {}, "all selectable cards chosen", "code")
         return Decision("__wait__", {},
-                        f"{len(_card_select_chosen)}/{need} chosen; nothing new to select",
+                        f"{len(chosen)}/{need} accepted toggles; nothing new to select",
                         "code")
 
     # One ABSOLUTE Noul per candidate, argmax in code. A single Choice over a
     # 13+ card deck diluted badly and the fallback then always took index 0.
-    # Candidates are `remaining` -- never an index we already toggled on.
+    # Candidates exclude indices with accepted toggle requests.
     keys = [f"card{c['index']}" for c in remaining]
     kind = screen_kind(screen, prompt)
     questions: dict[str, dict] = {}
@@ -1035,9 +1013,6 @@ def rest_site_decision(obs: dict) -> Decision:
                     "first rest option", "fallback")
 
 
-_last_shop_sig: str | None = None
-_last_shop_purchased = False
-
 # Minimum absolute Noul before buying anything in a shop. Absolute judgements
 # can legitimately be low for every candidate, so this is a floor, not a rank.
 SHOP_BUY_THRESHOLD = 0.60
@@ -1083,8 +1058,6 @@ def shop_decision(obs: dict, client: JevClient | None, deck: dict | None) -> Dec
     The state carries `price`, `is_stocked` and `can_afford` per item, so no
     affordability arithmetic needs to reach the model.
     """
-    global _last_shop_sig, _last_shop_purchased
-
     # `fake_merchant` nests its inventory one level deeper: fake_merchant.shop.
     # Reading only obs["shop"] made every fake-merchant shop look empty, so the
     # bot always left immediately without buying.
@@ -1105,16 +1078,6 @@ def shop_decision(obs: dict, client: JevClient | None, deck: dict | None) -> Dec
         # worked and moved the game to the map. Never wait on it indefinitely.
         return Decision("proceed", {}, "nothing affordable; leave", "code")
 
-    # Only stop buying when we purchased from this EXACT shop state and it did
-    # not change. Keying on the signature alone made a fresh shop look stalled
-    # because module-level state leaked in from an earlier screen.
-    sig = json.dumps(node, sort_keys=True)
-    if sig != _last_shop_sig:
-        _last_shop_sig = sig
-        _last_shop_purchased = False
-    elif _last_shop_purchased:
-        return Decision("proceed", {}, "shop unchanged after a purchase; leave", "code")
-
     if client is None:
         return Decision("proceed", {}, "no jev; skip shop", "fallback")
 
@@ -1178,7 +1141,6 @@ def shop_decision(obs: dict, client: JevClient | None, deck: dict | None) -> Dec
                         f"best item only noul={best_noul:.2f}; leave", "jev", best_noul)
 
     item = item_keys[best_key]
-    _last_shop_purchased = True
     return Decision("shop_purchase", {"index": item["index"]},
                     f"jev bought {shop_item_text(item)[0]} (noul={best_noul:.2f})",
                     "jev", best_noul)
@@ -1279,11 +1241,7 @@ def _weakest_potion_slot(obs: dict) -> int | None:
     return weakest.get("slot", 0)
 
 
-_last_rewards_sig: str | None = None
+def rewards_decision(obs: dict, context: PolicyContext) -> Decision:
-_rewards_skipped_card = False
-
-
-def rewards_decision(obs: dict) -> Decision:
     """
     Reward screen.
 
@@ -1299,7 +1257,7 @@ def rewards_decision(obs: dict) -> Decision:
     node = obs.get("rewards") or {}
     items = [i for i in (node.get("items") or []) if isinstance(i, dict)]
 
-    if _rewards_skipped_card:
+    if context.rewards_skipped_card:
         items = [i for i in items if i.get("type") != "card"]
 
     if not items:
@@ -1321,30 +1279,18 @@ def rewards_decision(obs: dict) -> Decision:
                     f"claim reward {idx} (right-to-left)", "code")
 
 
-_last_bundle_sig: str | None = None
-
-
 def bundle_select_decision(obs: dict, client: JevClient | None,
                            deck: dict | None) -> Decision:
     """
     Bundle choice: pick one of several 3-card bundles.
 
-    Same trap as `card_select`: `select_bundle` errors with "A bundle preview
+    `select_bundle` errors when a preview is already open. Confirm the preview
-    is already open - confirm or cancel it first" once a preview is showing.
+    instead. The context waits for evidence after either accepted request;
-    Confirm instead, and reset with cancel if a repeated state proves the
+    an unchanged read does not prove that confirmation failed.
-    confirm did not apply.
     """
-    global _last_bundle_sig
-
     bs = obs.get("bundle_select") or {}
-    sig = json.dumps(bs, sort_keys=True)
-    repeated = sig == _last_bundle_sig
-    _last_bundle_sig = sig
 
     if bs.get("preview_showing") and bs.get("can_confirm"):
-        if repeated:
-            return Decision("cancel_bundle_selection", {},
-                            "bundle preview did not apply; reset", "code")
         return Decision("confirm_bundle_selection", {}, "confirm the bundle", "code")
 
     bundles = [b for b in (bs.get("bundles") or []) if isinstance(b, dict)]
@@ -1507,10 +1453,7 @@ def hand_select_decision(obs: dict, client: JevClient | None,
                     f"forced selection; gave up {target.get('name')}", "fallback")
 
 
-_crystal_clicked: set[tuple[int, int]] = set()
+def crystal_sphere_decision(obs: dict, context: PolicyContext) -> Decision:
-
-
-def crystal_sphere_decision(obs: dict) -> Decision:
     """
     Crystal Sphere minigame.
 
@@ -1528,7 +1471,7 @@ def crystal_sphere_decision(obs: dict) -> Decision:
     # Never re-click a cell: that wastes a divination and can loop.
     fresh = [
         c for c in clickable
-        if (c.get("x"), c.get("y")) not in _crystal_clicked
+        if (c.get("x"), c.get("y")) not in context.accepted_crystal_cells
     ]
 
     if not fresh:
@@ -1543,15 +1486,15 @@ def crystal_sphere_decision(obs: dict) -> Decision:
     cx, cy = (width - 1) / 2, (height - 1) / 2
     cell = min(fresh, key=lambda c: abs(c.get("x", 0) - cx) + abs(c.get("y", 0) - cy))
 
-    _crystal_clicked.add((cell.get("x"), cell.get("y")))
     return Decision("crystal_sphere_click_cell",
                     {"x": cell.get("x"), "y": cell.get("y")},
                     f"reveal ({cell.get('x')},{cell.get('y')})", "code")
 
 
 def simple_decision(obs: dict, client: JevClient | None = None,
-                    deck: dict | None = None) -> Decision | None:
+                    deck: dict | None = None, *, context: PolicyContext | None = None) -> Decision | None:
     """Mechanical screens. Most need no model -- they are pure procedure."""
+    context = context if context is not None else PolicyContext()
     st = obs.get("state_type")
 
     if st == "menu":
@@ -1576,22 +1519,18 @@ def simple_decision(obs: dict, client: JevClient | None = None,
         #
         # Embarking immediately after selecting is also FLAKY -- measured, three
         # consecutive "select a character first" rejections -- because the
-        # selection has not registered yet. So ALTERNATE: select, embark,
+        # selection has not registered yet. Select after a rejected embark,
-        # select, embark. A rejected embark is always followed by a fresh
+        # but wait after an accepted embark. Proposals alone do not advance
-        # select, which makes the sequence self-correcting whatever the timing.
+        # the sequence.
         if screen == "character_select":
-            global _charselect_phase
-
             options = obs.get("options") or []
             names = [o if isinstance(o, str) else o.get("name") for o in options]
             pick = next((c for c in ("IRONCLAD", "SILENT") if c in names), None)
 
-            if _charselect_phase == 0 and pick is not None:
+            if not context.character_selected and pick is not None:
-                _charselect_phase = 1
                 return Decision("menu_select", {"option": pick},
                                 f"select {pick}", "code")
 
-            _charselect_phase = 0
             return Decision("menu_select", {"option": "embark"},
                             "embark (a rejected embark is followed by a re-select)",
                             "code")
@@ -1601,7 +1540,7 @@ def simple_decision(obs: dict, client: JevClient | None = None,
         return Decision("menu_select", {"option": "main_menu"}, "run ended", "code")
 
     if st == "rewards":
-        return rewards_decision(obs)
+        return rewards_decision(obs, context)
 
     if st == "card_reward":
         return card_reward_decision(obs, client, deck)
@@ -1628,13 +1567,13 @@ def simple_decision(obs: dict, client: JevClient | None = None,
         return hand_select_decision(obs, client, deck)
 
     if st == "card_select":
-        return card_select_decision(obs, client, deck)
+        return card_select_decision(obs, client, deck, context)
 
     if st == "bundle_select":
         return bundle_select_decision(obs, client, deck)
 
     if st == "crystal_sphere":
-        return crystal_sphere_decision(obs)
+        return crystal_sphere_decision(obs, context)
 
     # Transitions and unhandled overlays are not dead ends. Wait and look again;
     # run.py's unchanged-state guard bounds this so a real dead end still stops.
@@ -1644,13 +1583,6 @@ def simple_decision(obs: dict, client: JevClient | None = None,
     return None
 
 
-_last_state_type: str | None = None
-
-
-_last_charselect_sig: str | None = None
-_charselect_seen = False
-_charselect_phase = 0
-
 # rewards and card_reward are two views of one flow: claiming a card reward
 # opens the card screen, and skipping returns to the rewards screen. Treat them
 # as ONE screen group so per-flow state is not cleared on every hop.
@@ -1670,51 +1602,133 @@ def _screen_group(state_type: str | None, menu_screen: str | None = None) -> str
     return SCREEN_GROUPS.get(state_type, state_type)
 
 
-def _reset_screen_guards(state_type: str | None, menu_screen: str | None = None) -> None:
+@dataclass
+class PendingAction:
+    """An accepted request, not proof that the game completed it."""
+
+    decision: Decision
+    screen: dict
+
+
+@dataclass
+class PolicyContext:
+    """Session-owned memory. Proposals never record execution.
+
+    Grid toggles and character selection lack observable selection fields in
+    the mod. Their accepted requests are tracked explicitly, not called facts.
+    Other guarded actions wait for relevant screen evidence before continuing.
     """
-    Clear per-screen module state when the screen changes.
 
-    These guards detect "the same screen reappeared unchanged", which is only
+    screen_group: str | None = None
-    meaningful within a single screen. Left alone they leak across screens and
+    accepted_card_indices: set[int] = field(default_factory=set)
-    runs, so a fresh shop or card grid gets mistaken for a stalled one. Found
+    accepted_card_grid: list[dict] | None = None
-    by test_brain.py: a fresh fake-merchant shop was reported as
+    accepted_crystal_cells: set[tuple[int, int]] = field(default_factory=set)
-    "unchanged after a purchase" because a signature from an earlier case was
+    rewards_skipped_card: bool = False
-    still set.
+    character_selected: bool = False
+    pending: PendingAction | None = None
 
-    `unknown` and `overlay` are TRANSITIONS, not screens. Resetting on them
+    def observe(self, obs: dict) -> None:
-    wipes the state mid-flow -- during embark the state flickers through
-    `unknown`, which used to clear the character-select guard and restart the
-    select/embark cycle.
-    """
-    global _last_state_type, _last_card_select_sig, _last_bundle_sig
-    global _last_shop_sig, _last_shop_purchased, _rewards_skipped_card
-    global _charselect_seen, _charselect_phase
-    global _card_select_picked, _card_select_confirmed, _card_select_chosen
-    global _crystal_clicked
-
-    if state_type in ("unknown", "overlay"):
-        return
-
-    group = _screen_group(state_type, menu_screen)
-    if group == _last_state_type:
-        return
-    _last_state_type = group
-    _last_card_select_sig = None
-    _last_bundle_sig = None
-    _last_shop_sig = None
-    _last_shop_purchased = False
-    _rewards_skipped_card = False
-    _charselect_seen = False
-    _charselect_phase = 0
-    _card_select_picked = False
-    _card_select_confirmed = False
-    _card_select_chosen.clear()
-    _crystal_clicked.clear()
-
-
-def decide(obs: dict, client: JevClient | None, deck: dict | None = None) -> Decision | None:
-    _reset_screen_guards(obs.get("state_type"), obs.get("menu_screen"))
         st = obs.get("state_type")
+        if st in ("unknown", "overlay"):
+            return  # A transient overlay is not evidence of completion.
+        group = _screen_group(st, obs.get("menu_screen"))
+        if group != self.screen_group:
+            self.screen_group = group
+            self.accepted_card_indices.clear()
+            self.accepted_card_grid = None
+            self.accepted_crystal_cells.clear()
+            self.rewards_skipped_card = False
+            self.character_selected = False
+            self.pending = None
+            return
+        pending = self.pending
+        if pending is None:
+            return
+        action, params = pending.decision.action, pending.decision.params
+        screen = obs.get(st) or {}
+        if action == "select_card" and pending.screen.get("screen_type") != "choose":
+            # The API exposes no selected indices. Reserve accepted toggles
+            # after a fresh read, even when their effect is not visible yet.
+            if screen.get("cards") != pending.screen.get("cards"):
+                return  # Cannot map a toggle safely onto a changed grid.
+            self.accepted_card_indices.add(params["index"])
+            self.accepted_card_grid = pending.screen.get("cards")
+        elif action == "menu_select" and params.get("option") != "embark":
+            self.character_selected = True
+        elif action == "skip_card_reward":
+            if st != "rewards":
+                return
+            self.rewards_skipped_card = True
+        elif action in ("confirm_selection", "cancel_selection",
+                        "confirm_bundle_selection", "cancel_bundle_selection",
+                        "combat_confirm_selection"):
+            return  # Wait for the screen to close; do not cancel a slow confirm.
+        elif action == "crystal_sphere_click_cell":
+            cell = (params["x"], params["y"])
+            clickable = {(c.get("x"), c.get("y"))
+                         for c in screen.get("clickable_cells", [])}
+            if cell in clickable and not screen.get("can_proceed"):
+                return
+            self.accepted_crystal_cells.add(cell)
+        elif action == "shop_purchase":
+            # Gold alone can change for unrelated reasons. Require this item
+            # to change/disappear, or a screen transition (e.g. card removal).
+            before = pending.screen.get("shop", pending.screen)
+            after = screen.get("shop", screen)
+            index = params["index"]
+            old = next((i for i in before.get("items", []) if i.get("index") == index), None)
+            new = next((i for i in after.get("items", []) if i.get("index") == index), None)
+            def inventory_item(item):
+                return {k: v for k, v in item.items() if k not in ("can_afford", "price")} if item else None
+            if inventory_item(old) == inventory_item(new):
+                return
+        elif action == "combat_select_card":
+            if len(screen.get("selected_cards") or []) <= len(pending.screen.get("selected_cards") or []):
+                return
+        elif action == "select_bundle":
+            if not screen.get("preview_showing") or not screen.get("can_confirm"):
+                return
+        else:
+            return  # Direct choices and embark require a screen transition.
+        self.pending = None
+
+    def record_result(self, obs: dict, decision: Decision, *, accepted: bool) -> None:
+        """Called only after a real action result. Transport errors stop the runner."""
+        if not accepted:
+            self.pending = None
+            if decision.action == "menu_select" and decision.params.get("option") == "embark":
+                self.character_selected = False
+            return
+        st = obs.get("state_type")
+        guarded = decision.action in {
+            "select_card", "confirm_selection", "cancel_selection",
+            "select_bundle", "confirm_bundle_selection", "cancel_bundle_selection",
+            "shop_purchase", "skip_card_reward", "crystal_sphere_click_cell",
+            "combat_select_card", "combat_confirm_selection",
+        } or (st == "menu" and obs.get("menu_screen") == "character_select")
+        if guarded:
+            # Detach from mutable fixtures/callers. This is a small screen node,
+            # not another copy of the entire combat observation.
+            screen = json.loads(json.dumps(obs.get(st) or {}))
+            self.pending = PendingAction(decision, screen)
+
+
+def decide(obs: dict, client: JevClient | None, deck: dict | None = None,
+           *, context: PolicyContext | None = None) -> Decision | None:
+    """Reconcile a fresh observation, then propose one action.
+
+    Omit context for independent snapshot analysis. Runners must retain one
+    context and report action results; repeated proposals alone advance nothing.
+    """
+    context = context if context is not None else PolicyContext()
+    context.observe(obs)
+    st = obs.get("state_type")
+    if context.pending is not None:
+        return Decision("__wait__", {},
+                        f"awaiting screen evidence after {context.pending.decision.action}", "code")
+    if (st == "card_select" and context.accepted_card_grid is not None
+            and (obs.get("card_select") or {}).get("cards") != context.accepted_card_grid):
+        return Decision("__wait__", {}, "card grid changed; accepted indices cannot be mapped safely", "code")
     if st in ("monster", "elite", "boss"):
         return combat_decision(F.combat_facts(obs), client)
-    return simple_decision(obs, client, deck)
+    return simple_decision(obs, client, deck, context=context)

docs/research/12-correctness-pass.md

@@ -117,7 +117,8 @@ The audit remains diagnostic, not a pass/fail suite. It also reports known issue
 
 ## Next: policy state and recording
 
-Still open:
+Update: [the policy-state pass](13-policy-state.md) addresses item 1 below.
+The list records the remaining work at the end of this correctness pass:
 
 1. Selection/shop/minigame memory can change when an action is proposed, before execution succeeds.
 2. Captures cover only combat and reuse filenames across sessions.

docs/research/13-policy-state.md

@@ -0,0 +1,113 @@
+# Session-owned policy state
+
+## Scope
+
+This pass replaces per-screen globals with explicit session memory. It follows
+[the correctness pass](12-correctness-pass.md). It does not redesign recording,
+identify runs, or establish persistent-deck provenance.
+
+`brain.py` remains one file. Moving policy code and changing its state behavior
+at the same time would make failures harder to diagnose. A later extraction
+can preserve the interface and the runner-level tests.
+
+## Lifecycle
+
+1. `run.main()` creates one `brain.PolicyContext` for the session.
+2. `brain.decide(..., context=context)` reconciles the fresh observation.
+3. The policy returns a `Decision`, which is only a proposal.
+4. The runner executes the action unless this is a dry-run or a suppressed duplicate.
+5. The runner calls `context.record_result(..., accepted=result.ok)`.
+6. The next observation supplies evidence before further guarded actions.
+
+A rejected action does not reserve a card index, consume a crystal cell, or
+mark a reward as skipped. A rejected embark clears the accepted character
+selection, so the next proposal selects the character again.
+
+Transport errors stop the runner. They can leave the game's action outcome
+unknown. The runner does not retry those POSTs automatically.
+
+Calling `decide` without a context performs independent snapshot analysis.
+Sequential callers must retain a context and report real action results.
+Direct screen handlers are policy helpers, not a replacement for this lifecycle.
+
+## Evidence and limits
+
+An accepted response does not prove completion. `PendingAction` represents an
+accepted request awaiting reconciliation, not a successful game effect.
+
+| Action | Condition for further progress |
+| --- | --- |
+| Grid toggle | A fresh read of the same grid reserves the accepted index. Selection itself remains unverified. |
+| Direct choose-card screen | A screen-flow transition. |
+| Hand selection | The observed selected-card count increases, or the screen changes. Selected-list indices use a separate index space. |
+| Grid, hand, or bundle confirmation/cancellation | A screen-flow transition. No automatic repeated confirm or speculative cancel. |
+| Bundle selection | An observable preview with confirmation enabled, or a screen transition. |
+| Purchase | The targeted inventory item changes or disappears, or the screen changes. Price and affordability changes alone do not release the guard. |
+| Crystal click | The cell is no longer clickable, progression unlocks, or the screen changes. |
+| Character selection | A fresh read releases the accepted request. The mod does not expose the selected character. |
+| Embark | A screen transition. Rejection instead allows re-selection. |
+| Card-reward skip | The reward list reappears before the policy marks the reward as skipped. |
+
+### Missing selection information
+
+The vendored mod's `BuildCardSelectState` exposes grid cards, preview state,
+and confirmation availability. It does not expose all selected card indices.
+`ExecuteSelectCard` reports that it emitted a toggle signal, not that the
+selection is complete.
+
+The policy therefore stores **accepted toggles**, not confirmed selections.
+This permits multi-select screens whose observation remains unchanged between
+selections. It excludes accepted indices from later proposals in the same flow.
+If the grid changes while accepted indices remain in use, the policy waits
+rather than mapping old indices onto the changed grid.
+
+This is a limited inference, not a full game-state model. A lost accepted toggle
+can stall the flow. Restarting the bot on a partially selected grid also loses
+the acceptance ledger. Reliable recovery needs selected-card identities from
+the mod, or an explicit operator reset. This pass adds neither.
+
+### Screen boundaries and waiting
+
+- `rewards` and `card_reward` share one flow, so skipping survives the return to rewards.
+- Menu screen names distinguish character selection from the main menu.
+- `unknown` and `overlay` do not reset memory or prove completion.
+- Other observed flow changes clear screen memory and pending requests.
+- An unchanged screen does not prove an action failed. The policy waits instead of cancelling a slow confirmation.
+- The existing stuck limit bounds unchanged observations. A separate pending-action deadline uses the same `--stuck-seconds` limit, even when unrelated fields change.
+- A bounded session can still end at its step limit with an unresolved action. Exit zero for that mode does not claim action completion.
+
+Same-type screen replacements are not uniquely identified. There is no run or
+screen-instance identifier in the context yet. Recovery is conservative: an
+ambiguous flow can stop rather than risk another purchase or toggle.
+
+## Validation
+
+No live game connection, paid model calls, new dependencies, or new test files.
+
+- Existing scripts: 68 facts, 110 policy, and 84 runner assertions passed; 262 total.
+- Runner flows cover rejected and delayed selections, confirmation, purchases,
+  bundles, crystal clicks, character selection, reward skipping, session
+  isolation, dry-run, and pending-action timeout under unrelated state changes.
+- Old tests that advanced policy memory by merely proposing actions were removed.
+  The combined policy/runner test code is smaller than before this pass.
+- Thirteen temporary whole-process scenarios passed with real HTTP clients and
+  loopback fixture servers. These include the eight prior correctness scenarios
+  and five selection, dry-run, purchase, bundle, and crystal flows.
+- The offline audit replayed 346 stored observations without exceptions.
+  Repeated proposals with one retained context no longer consume a selection.
+- Dataset integrity: 37 runs, 1,052 decisions, and 346 observations passed.
+- Shell syntax checks passed.
+
+Temporary probes are not part of the repository. The permanent state regressions
+use the actual runner and policy together, with isolated game/model boundaries.
+
+## Next boundaries
+
+The state interface now gives a stable boundary for extracting combat and
+selection policy without changing execution behavior. Keep a small dispatcher;
+no plugin framework or class hierarchy is needed.
+
+The next state work is run identity and deck provenance. Recording should then
+link observations, proposals, action attempts, results, and reconciliation.
+Session finalization, capture-name collisions, and actual policy-gate metadata
+remain separate work.

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
+        # Acceptance does not prove completion. The policy context handles
-        # animating. If the action was rejected, fall through and pick a
+        # pending selections and purchases. Rejections allow another proposal
-        # different one instead of waiting out the stuck counter.
+        # 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
+        # allowed here, but cannot bypass a pending policy action above.
-        # shops work. The bound matters because some actions legitimately need
+        # 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",

test_brain.py

@@ -114,20 +114,6 @@ def combat(state_type="monster", hand=None, enemies=None, **pkw) -> dict:
     }
 
 
-_reset_counter = 0
-
-
-def force_reset() -> None:
-    """Force a screen-group change so every per-screen guard is cleared.
-
-    `_reset_screen_guards` only resets on a GROUP CHANGE, so calling it with
-    the same group twice is a deliberate no-op. Tests need a guaranteed reset.
-    """
-    global _reset_counter
-    _reset_counter += 1
-    brain._reset_screen_guards(f"__test{_reset_counter}__")
-
-
 print("=== 1. every state_type produces a LEGAL action ===")
 
 cases: list[tuple[str, dict]] = [
@@ -391,7 +377,6 @@ check("fake_merchant reports nothing affordable",
 
 # A skipped card reward is NOT consumed by the game. Without the guard the bot
 # loops claim_reward -> card screen -> skip -> claim_reward forever.
-force_reset()
 rewards_with_card = {
     "state_type": "rewards",
     "rewards": {"items": [{"index": 0, "type": "gold", "description": "18 Gold",
@@ -405,31 +390,6 @@ rewards_with_card = {
 d = brain.decide(rewards_with_card, None)
 check("rewards claims the last item first", d.params.get("index"), 1)
 
-brain._rewards_skipped_card = True
-d = brain.decide(rewards_with_card, None)
-check("a skipped card reward is not re-claimed", d.action, "claim_reward")
-check("...and the gold is taken instead", d.params.get("index"), 0)
-
-brain._rewards_skipped_card = True
-d = brain.decide(
-    {"state_type": "rewards",
-     "rewards": {"items": [{"index": 0, "type": "card",
-                             "description": "Add a card to your deck."}],
-                 "can_proceed": True},
-     "run": {"act": 1, "floor": 2}, "player": player()},
-    None,
-)
-check("rewards with only a skipped card proceeds", d.action, "proceed")
-
-# rewards -> card_reward must stay in ONE screen group, or the skip flag is
-# cleared on every hop and the loop returns.
-brain._reset_screen_guards("rewards")
-brain._rewards_skipped_card = True
-brain._reset_screen_guards("card_reward")
-check("skip flag survives rewards <-> card_reward", brain._rewards_skipped_card, True)
-brain._reset_screen_guards("monster")
-check("skip flag clears when leaving the flow", brain._rewards_skipped_card, False)
-
 class StubClient:
     """
     Deterministic stand-in for JevClient so the model paths can be tested
@@ -487,7 +447,6 @@ valid_indices = {OFFSET, OFFSET + 1, OFFSET + 2}
 stub = StubClient()
 
 # card_select (upgrade)
-force_reset()
 d = brain.decide(
     {"state_type": "card_select",
      "card_select": {"screen_type": "upgrade", "prompt": "Choose a card to Upgrade.",
@@ -504,7 +463,6 @@ check("...and picks by identity, not position",
 # card_reward with the stub. `skip_all` is answered NO so the take path is
 # exercised; otherwise a stub that says yes to everything always skips.
 take_stub = StubClient(noul_override={"skip_all": 0.10})
-force_reset()
 d = brain.decide(
     {"state_type": "card_reward",
      "card_reward": {"cards": [offset_card("Common A", 0),
@@ -517,7 +475,6 @@ check("card_reward emits the card's own index", d.params.get("card_index") in va
 
 # And with `skip_all` answered YES, Jev's skip is honoured.
 skip_stub = StubClient(noul_override={"skip_all": 0.95})
-force_reset()
 d = brain.decide(
     {"state_type": "card_reward",
      "card_reward": {"cards": [offset_card("Common A", 0)], "can_skip": True},
@@ -529,7 +486,6 @@ check("...and the reason cites jev", "jev:" in d.reason, True)
 
 # An UNCERTAIN skip answer must not skip -- take the best card instead.
 unsure_stub = StubClient(noul_override={"skip_all": 0.55})
-force_reset()
 d = brain.decide(
     {"state_type": "card_reward",
      "card_reward": {"cards": [offset_card("Common A", 0)], "can_skip": True},
@@ -539,7 +495,6 @@ d = brain.decide(
 check("an uncertain skip answer takes a card", d.action, "select_card_reward")
 
 # `can_skip: false` must take a card even if Jev would skip.
-force_reset()
 d = brain.decide(
     {"state_type": "card_reward",
      "card_reward": {"cards": [offset_card("Common A", 0)], "can_skip": False},
@@ -549,7 +504,6 @@ d = brain.decide(
 check("cannot skip -> takes a card anyway", d.action, "select_card_reward")
 
 # hand_select (in-combat selectable subset)
-force_reset()
 d = brain.decide(
     {"state_type": "hand_select",
      "hand_select": {"mode": "simple_select", "prompt": "Choose cards to exhaust.",
@@ -562,7 +516,6 @@ check("hand_select emits the selectable card's own index",
       d.params.get("card_index"), OFFSET + 1)
 
 # rewards
-force_reset()
 d = brain.decide(
     {"state_type": "rewards",
      "rewards": {"items": [{"index": OFFSET, "type": "gold", "description": "1 Gold"},
@@ -574,7 +527,6 @@ d = brain.decide(
 check("rewards emits the item's own index", d.params.get("index"), OFFSET + 1)
 
 # shop
-force_reset()
 d = brain.decide(
     {"state_type": "shop",
      "shop": {"items": [{"index": OFFSET, "category": "relic", "price": 10,
@@ -590,7 +542,6 @@ d = brain.decide(
 check("shop emits the item's own index", d.params.get("index") in valid_indices, True)
 
 # map
-force_reset()
 d = brain.decide(
     {"state_type": "map",
      "map": {"next_options": [{"index": OFFSET, "col": 0, "row": 1, "type": "Monster"},
@@ -602,7 +553,6 @@ d = brain.decide(
 check("map emits the node's own index", d.params.get("index") in valid_indices, True)
 
 # bundle_select
-force_reset()
 d = brain.decide(
     {"state_type": "bundle_select",
      "bundle_select": {"screen_type": "bundle", "prompt": "Choose a bundle.",
@@ -615,7 +565,6 @@ d = brain.decide(
 check("bundle_select emits the bundle's own index", d.params.get("index") in valid_indices, True)
 
 # relic_select
-force_reset()
 d = brain.decide(
     {"state_type": "relic_select",
      "relic_select": {"prompt": "Choose a relic.",
@@ -631,7 +580,6 @@ check("relic_select emits the relic's own index", d.params.get("index") in valid
 
 # A state whose list positions are ALL zero-ish while the real indices are high:
 # any position-based code returns 0, which is not a valid index here.
-force_reset()
 d = brain.decide(
     {"state_type": "card_select",
      "card_select": {"screen_type": "remove", "prompt": "Choose a card to Remove.",
@@ -644,66 +592,6 @@ check("a position-based pick of 0 would be invalid here",
       d.params.get("index") != 0, True)
 
 print()
-print("=== 5. character select needs a select-then-embark sequence ===")
-
-# The state carries NO "selected" indicator -- the mod hardcodes `message` to
-# "Select a character." whatever is chosen. So the handler selects once and
-# embarks when the screen comes back unchanged. Without this the bot selects
-# the already-selected character forever and stalls.
-char_select_state = {
-    "state_type": "menu", "menu_screen": "character_select",
-    "message": "Select a character.",
-    "options": [{"name": "IRONCLAD", "enabled": True},
-                {"name": "SILENT", "enabled": True},
-                {"name": "confirm", "enabled": True},
-                {"name": "embark", "enabled": True},
-                {"name": "back", "enabled": True}],
-}
-
-force_reset()
-d1 = brain.decide(char_select_state, None)
-check("character_select first selects a character", d1.params.get("option"), "IRONCLAD")
-
-# Same screen unchanged -> a character is already selected -> embark.
-d2 = brain.decide(char_select_state, None)
-check("character_select then embarks", d2.params.get("option"), "embark")
-
-# The sequence ALTERNATES, which is what makes it self-correcting: a rejected
-# embark ("select a character first") is always followed by a fresh select.
-d2b = brain.decide(char_select_state, None)
-check("a third call re-selects", d2b.params.get("option"), "IRONCLAD")
-d2c = brain.decide(char_select_state, None)
-check("and then embarks again", d2c.params.get("option"), "embark")
-
-# A changing signature must not derail the alternation.
-changed = dict(char_select_state)
-changed["characters"] = [{"name": "The Ironclad", "id": "IRONCLAD", "locked": False}]
-d2d = brain.decide(changed, None)
-check("a changing signature does not derail the cycle",
-      d2d.params.get("option"), "IRONCLAD")
-
-# Both actions are legal for this screen.
-check("select is legal", legal(char_select_state, d1), True)
-check("embark is legal", legal(char_select_state, d2), True)
-
-# Leaving the screen must clear the guard, or the next run embarks immediately
-# without selecting anything.
-brain._reset_screen_guards("menu", "main")
-d3 = brain.decide(char_select_state, None)
-check("guard resets when leaving character select", d3.params.get("option"), "IRONCLAD")
-
-# The alternation must also reset, or a fresh run embarks without selecting.
-brain._reset_screen_guards("menu", "main")
-d3b = brain.decide(char_select_state, None)
-check("alternation resets too", d3b.params.get("option"), "IRONCLAD")
-
-# menu_screen must be part of the screen key: main menu and character select are
-# both state_type "menu" but have different valid actions.
-brain._reset_screen_guards("menu", "main")
-check("main menu and character select are different screens",
-      brain._screen_group("menu", "main") != brain._screen_group("menu", "character_select"),
-      True)
-
 # combat_select_card TOGGLES: re-selecting an already-chosen card deselects it,
 # so the state never changes and the loop stalls. `selected_cards` is the
 # authoritative list of what is already chosen.
@@ -754,18 +642,15 @@ check("upgrade_select confirms instead of selecting", d.action, "combat_confirm_
 # No enemies but combat still in progress (Phrog Parasite spawning the next
 # wave). Measured live: the bot emitted play_card(card_index=1) with no target
 # and the game rejected it with "Card requires a target".
-force_reset()
 obs_no_enemies = combat(enemies=[])
 d = brain.decide(obs_no_enemies, None)
 check("no enemies -> wait, do not play", d.action, "__wait__")
 
 # A target-requiring card must never be emitted without a target.
-force_reset()
 obs_one_enemy = combat(hand=[card("Strike", 0)], enemies=[enemy()])
 d = brain.decide(obs_one_enemy, None)
 check("a single enemy is auto-targeted", d.params.get("target"), "NIBBIT_0")
 
-force_reset()
 d = brain.decide(obs_no_enemies, StubClient())
 check("even with jev, no enemies means wait", d.action, "__wait__")
 
@@ -777,28 +662,6 @@ check("raw class name normalises to upgrade",
 check("friendly names still work",
       (brain.screen_kind("remove"), brain.screen_kind("transform")), ("remove", "transform"))
 
-enchant_state = {
-    "state_type": "card_select",
-    "card_select": {"screen_type": "NDeckEnchantSelectScreen",
-                    "prompt": "Choose a card to Enchant.",
-                    "cards": [offset_card("Strike", 0), offset_card("Perfected Strike", 1)],
-                    "preview_showing": False,
-                    "can_cancel": False, "can_confirm": True},
-    "run": {"act": 1, "floor": 13}, "player": player(),
-}
-
-force_reset()
-d1 = brain.decide(enchant_state, None)
-check("enchant screen selects first", d1.action, "select_card")
-
-d2 = brain.decide(enchant_state, None)
-check("enchant screen then confirms", d2.action, "confirm_selection")
-
-# The full flow must terminate: select, confirm, done.
-force_reset()
-seq = [brain.decide(enchant_state, None).action, brain.decide(enchant_state, None).action]
-check("enchant flow terminates", seq, ["select_card", "confirm_selection"])
-
 # MULTI-select: "Choose 5 cards to Remove" keeps can_confirm FALSE until all
 # five are picked, and card_select has no `selected_cards` field. Measured live:
 # the handler re-selected the same index forever and stalled.
@@ -817,57 +680,8 @@ check("an ADD screen is not an upgrade screen",
 check("a REMOVE prompt beats a generic screen_type",
       brain.screen_kind("select", "Choose 5 cards to Remove."), "remove")
 
-multi_state = {
-    "state_type": "card_select",
-    "card_select": {"screen_type": "select", "prompt": "Choose 5 cards to Remove.",
-                    "cards": [offset_card(f"Card{i}", i) for i in range(8)],
-                    "preview_showing": False,
-                    "can_cancel": False, "can_confirm": False},
-    "run": {"act": 2, "floor": 18}, "player": player(),
-}
-
-force_reset()
-seq = [brain.decide(multi_state, None).action for _ in range(6)]
-check("multi-select picks 5 then waits",
-      seq, ["select_card"] * 5 + ["__wait__"])
-
-# An ADD screen: 2 cards, then the game closes the screen by itself (no confirm).
-force_reset()
-add_state = {
-    "state_type": "card_select",
-    "card_select": {"screen_type": "simple_select",
-                    "prompt": "Choose 2 Common Cards to Add to Your Deck.",
-                    "cards": [offset_card(f"Card{i}", i) for i in range(8)],
-                    "preview_showing": False,
-                    "can_cancel": False, "can_confirm": False},
-    "run": {"act": 1, "floor": 5}, "player": player(),
-}
-add_seq = [brain.decide(add_state, None).action for _ in range(4)]
-check("an ADD screen picks exactly 2 then waits",
-      add_seq, ["select_card", "select_card", "__wait__", "__wait__"])
-
-# Distinct picks need a fresh screen, or the picks continue from above.
-force_reset()
-add_idxs = [brain.decide(add_state, None).params.get("index") for _ in range(2)]
-check("ADD picks are distinct", len(set(add_idxs)), 2)
-
-# Every pick must be a DIFFERENT index, or select_card toggles it back off.
-force_reset()
-idxs = [brain.decide(multi_state, None).params.get("index") for _ in range(5)]
-check("each pick targets a different index", len(set(idxs)), 5)
-
-# Once can_confirm turns true, it confirms.
-force_reset()
-for _ in range(5):
-    brain.decide(multi_state, None)
-confirmable = dict(multi_state)
-confirmable["card_select"] = dict(multi_state["card_select"], can_confirm=True)
-d = brain.decide(confirmable, None)
-check("multi-select confirms once the count is met", d.action, "confirm_selection")
-
 # Crystal Sphere: can_proceed is FALSE until tiles are revealed. The old
 # unconditional crystal_sphere_proceed was rejected and stalled the run.
-force_reset()
 cs_state = {
     "state_type": "crystal_sphere",
     "crystal_sphere": {"grid_width": 11, "grid_height": 11,
@@ -880,13 +694,7 @@ d = brain.decide(cs_state, None)
 check("crystal sphere reveals a cell first", d.action, "crystal_sphere_click_cell")
 check("...starting from the centre", (d.params.get("x"), d.params.get("y")), (5, 5))
 
-# It must not re-click the same cell.
-d2 = brain.decide(cs_state, None)
-check("crystal sphere does not repeat a cell",
-      (d2.params.get("x"), d2.params.get("y")) != (d.params.get("x"), d.params.get("y")), True)
-
 # Once can_proceed unlocks, it proceeds.
-force_reset()
 cs_done = dict(cs_state)
 cs_done["crystal_sphere"] = dict(cs_state["crystal_sphere"], can_proceed=True)
 d3 = brain.decide(cs_done, None)
@@ -899,7 +707,6 @@ check("a singular prompt needs 1", brain.hand_select_need("Choose a card to Exha
 check("a counted prompt needs N", brain.hand_select_need("Choose 2 cards to discard."), 2)
 check("\"any number\" is unbounded", brain.hand_select_need("Choose any number of cards to replace."), None)
 
-force_reset()
 d = brain.decide(
     {"state_type": "hand_select",
      "hand_select": {"mode": "simple_select", "prompt": "Choose a card to Exhaust.",
@@ -911,7 +718,6 @@ d = brain.decide(
 check("a satisfied singular prompt confirms", d.action, "combat_confirm_selection")
 
 # An unsatisfied counted prompt keeps selecting.
-force_reset()
 d = brain.decide(
     {"state_type": "hand_select",
      "hand_select": {"mode": "simple_select", "prompt": "Choose 2 cards to discard.",
@@ -929,7 +735,6 @@ try:
     weak_stub = StubClient(noul=0.50, noul_override={"skip_all": 0.10})
 
     brain.CARD_SKIP_POLICY = "jev"
-    force_reset()
     d = brain.decide(
         {"state_type": "card_reward",
          "card_reward": {"cards": [offset_card("Common A", 0)], "can_skip": True},
@@ -939,7 +744,6 @@ try:
     check("jev policy takes a weak card", d.action, "select_card_reward")
 
     brain.CARD_SKIP_POLICY = "combined"
-    force_reset()
     d = brain.decide(
         {"state_type": "card_reward",
          "card_reward": {"cards": [offset_card("Common A", 0)], "can_skip": True},
@@ -953,7 +757,6 @@ try:
     strong_stub = StubClient(noul=0.80, noul_override={"skip_all": 0.10})
     for policy in ("jev", "combined"):
         brain.CARD_SKIP_POLICY = policy
-        force_reset()
         d = brain.decide(
             {"state_type": "card_reward",
              "card_reward": {"cards": [offset_card("Strong", 0)], "can_skip": True},

test_run.py

@@ -108,13 +108,14 @@ class FakeSts2:
         self.actions.append((a, k))
         if self.action_error:
             raise self.action_error
-        return types.SimpleNamespace(ok=self.action_ok, message="rejected" if not self.action_ok else "")
+        ok = (self.action_ok[min(len(self.actions) - 1, len(self.action_ok) - 1)]
+              if isinstance(self.action_ok, list) else self.action_ok)
+        return types.SimpleNamespace(ok=ok, message="rejected" if not ok else "")
 
 
 def invoke(fake, *flags, client=None, clock=None, decide=None):
     """Run the real loop with isolated files, no delays, and no live services."""
     callbacks = []
-    brain._reset_screen_guards("test-run-reset")
     with tempfile.TemporaryDirectory() as directory, contextlib.ExitStack() as stack:
         capdir = pathlib.Path(directory) / "capture"
         stack.enter_context(patch.object(run, "sts2", fake))
@@ -291,6 +292,104 @@ result = invoke(FakeSts2([menu, combat]), "--max-jev-errors", "2", client=client
 check("successful model call resets the failure budget", client.calls, 4)
 check("two failures after recovery abort", result.rc, 3)
 
+print()
+print("=== 8. policy memory follows action results and fresh observations ===")
+# Exercise the real policy and runner together. No global reset between sessions.
+grid = {"state_type": "card_select", "card_select": {
+    "screen_type": "select", "prompt": "Choose 2 cards to Remove.",
+    "cards": [{"index": 0, "name": "Strike"}, {"index": 1, "name": "Defend"}],
+    "can_confirm": False, "preview_showing": False}}
+confirmable = dict(grid, card_select=dict(grid["card_select"], can_confirm=True))
+# One rejected toggle, two accepted toggles, a delayed confirm, then transition.
+fake = FakeSts2([menu, grid, grid, grid, confirmable, confirmable,
+                 {"state_type": "overlay"}, confirmable, menu],
+                action_ok=[False, True])
+result = invoke(fake, "--no-jev", "--steps", "8", "--max-duplicate-waits", "0")
+check("selection flow completes", result.rc, 0)
+check("rejected selection is retried; accepted indices are not toggled twice",
+      fake.actions, [(("select_card",), {"index": 0}),
+                     (("select_card",), {"index": 0}),
+                     (("select_card",), {"index": 1}),
+                     (("confirm_selection",), {}),
+                     (("menu_select",), {"option": "singleplayer"})])
+changed_grid = dict(grid, card_select=dict(grid["card_select"],
+                    cards=[{"index": 0, "name": "Defend"}, {"index": 1, "name": "Strike"}]))
+fake = FakeSts2([menu, grid, changed_grid])
+result = invoke(fake, "--no-jev", "--steps", "2", "--max-duplicate-waits", "0")
+check("changed grid does not reuse the old selection indices", len(fake.actions), 1)
+for session in range(2):
+    fake = FakeSts2([menu, grid])
+    result = invoke(fake, "--no-jev", "--steps", "1")
+    check(f"session {session + 1} starts with fresh selection memory",
+          fake.actions, [(("select_card",), {"index": 0})])
+fake = FakeSts2([menu, grid])
+result = invoke(fake, "--no-jev", "--steps", "3", "--dry-run", "--max-duplicate-waits", "0")
+check("dry-run sends no selection actions", fake.actions, [])
+check("dry-run proposals never advance selection memory",
+      [r["params"] for r in result.rows if r["event"] == "decide"], [{"index": 0}] * 3)
+
+shop = {"state_type": "shop", "shop": {"items": [
+    {"index": 0, "category": "card", "card_name": "Inflame",
+     "card_description": "Gain 2 Strength.", "price": 50,
+     "is_stocked": True, "can_afford": True}]}}
+sold = {"state_type": "shop", "shop": {"items": []}}
+affordability_only = dict(shop, shop={"items": [dict(shop["shop"]["items"][0], can_afford=False)]})
+fake = FakeSts2([menu, shop, shop, affordability_only, sold])
+result = invoke(fake, "--steps", "4", "--max-duplicate-waits", "0")
+check("purchase waits through unchanged inventory and unrelated state changes",
+      fake.actions, [(("shop_purchase",), {"index": 0}), (("proceed",), {})])
+
+bundle = {"state_type": "bundle_select", "bundle_select": {
+    "bundles": [{"index": 2, "cards": []}], "preview_showing": False}}
+preview = dict(bundle, bundle_select=dict(bundle["bundle_select"],
+                                         preview_showing=True, can_confirm=True))
+fake = FakeSts2([menu, bundle, bundle, preview, preview, menu])
+result = invoke(fake, "--no-jev", "--steps", "5", "--max-duplicate-waits", "0")
+check("bundle waits for preview and never cancels a delayed confirmation",
+      [a[0][0] for a in fake.actions], ["select_bundle", "confirm_bundle_selection", "menu_select"])
+
+crystal = {"state_type": "crystal_sphere", "crystal_sphere": {
+    "grid_width": 3, "grid_height": 3, "clickable_cells": [{"x": 1, "y": 1}],
+    "can_proceed": False}}
+revealed = dict(crystal, crystal_sphere=dict(crystal["crystal_sphere"], can_proceed=True))
+fake = FakeSts2([menu, crystal, crystal, crystal, revealed], action_ok=[False, True])
+result = invoke(fake, "--no-jev", "--steps", "4", "--max-duplicate-waits", "0")
+check("rejected crystal click stays available; accepted click waits for evidence",
+      [a[0][0] for a in fake.actions],
+      ["crystal_sphere_click_cell", "crystal_sphere_click_cell", "crystal_sphere_proceed"])
+
+characters = {"state_type": "menu", "menu_screen": "character_select", "options": ["IRONCLAD", "embark"]}
+fake = FakeSts2([menu, characters, characters, characters, characters, characters,
+                 {"state_type": "unknown"}, characters, menu],
+                action_ok=[True, False, True])
+result = invoke(fake, "--no-jev", "--steps", "8", "--max-duplicate-waits", "0")
+check("rejected embark re-selects; accepted embark waits through transitions",
+      [a[1].get("option") for a in fake.actions],
+      ["IRONCLAD", "embark", "IRONCLAD", "embark", "singleplayer"])
+
+rewards = {"state_type": "rewards", "rewards": {"items": [{"index": 0, "type": "card"}]}}
+fake = FakeSts2([menu, card_reward, card_reward, rewards, menu, rewards], action_ok=[False, True])
+result = invoke(fake, "--steps", "5", "--max-duplicate-waits", "0")
+check("skip memory follows acceptance and clears outside the reward flow",
+      [a[0][0] for a in fake.actions],
+      ["skip_card_reward", "skip_card_reward", "proceed", "menu_select", "claim_reward"])
+
+hand = {"state_type": "hand_select", "hand_select": {
+    "mode": "simple_select", "prompt": "Choose a card to Exhaust.",
+    "cards": [{"index": 2, "name": "Strike"}], "can_confirm": False}}
+selected_hand = dict(hand, hand_select=dict(hand["hand_select"], can_confirm=True,
+                                           selected_cards=[{"index": 0, "name": "Strike"}]))
+fake = FakeSts2([menu, hand, hand, selected_hand, selected_hand, menu])
+result = invoke(fake, "--no-jev", "--steps", "5", "--max-duplicate-waits", "0")
+check("hand selection waits for selected count, not the separate selected-card index",
+      [a[0][0] for a in fake.actions], ["combat_select_card", "combat_confirm_selection", "menu_select"])
+
+fake = FakeSts2([menu] + [dict(shop, animation_tick=i) for i in range(100)])
+result = invoke(fake, "--steps", "100", "--stuck-seconds", "10", clock=itertools.count())
+check("pending purchase times out despite unrelated observation changes",
+      (result.rc, result.sessions[0]["stop_reason"], len(fake.actions)),
+      (1, "pending_action_timeout", 1))
+
 print()
 print(f"=== {PASS} passed, {FAIL} failed ===")
 sys.exit(1 if FAIL else 0)

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)
+    context = brain.PolicyContext()
-    second = brain.decide(obs, None)
+    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.