fix(ai): stop proposing strategically vacuous loop-shortcut Shortens

crates/engine/src/ai_support/mod.rs

@@ -6,6 +6,7 @@ mod evoke;
 pub mod filter;
 mod payment_continuation;
 mod prospective_mana;
+mod shortcut_efficacy;
 mod swarm;
 mod targeted_exchange;
 
@@ -1352,25 +1353,48 @@ fn has_activatable_sacrifice_for_mana(state: &GameState) -> bool {
         && mana_actions_include_meaningful_sacrifice(state, &activatable_object_mana_actions(state))
 }
 
+/// The mana activations [`has_activatable_sacrifice_for_mana`] counts, as
+/// items rather than as a yes/no.
+///
+/// SINGLE AUTHORITY for the predicate. `smart_shortcut_response`'s stage 2 has
+/// to classify exactly the actions stage 1 counted, and stage 1 counts these —
+/// which the flat `legal_actions` list structurally cannot contain (they live
+/// in `legal_actions_by_object` only). Re-testing the same shape at that call
+/// site would be a parallel copy free to drift, so the call site consumes this
+/// iterator and `mana_actions_include_meaningful_sacrifice` is defined as its
+/// emptiness.
+fn meaningful_sacrifice_mana_actions<'a>(
+    state: &'a GameState,
+    object_mana_actions: &'a [GameAction],
+) -> impl Iterator<Item = &'a GameAction> {
+    object_mana_actions.iter().filter(move |action| {
+        matches!(
+            action,
+            GameAction::ActivateAbility {
+                source_id,
+                ability_index,
+            } if activate_ability_is_meaningful_priority(state, *source_id, *ability_index)
+        )
+    })
+}
+
 /// Slice-taking core of [`has_activatable_sacrifice_for_mana`]: given a
 /// precomputed activatable mana-action sweep, true iff any action is a
 /// meaningful (sacrifice-for-mana) mana activation. Extracted so
 /// `auto_pass_recommended` can compute the sweep ONCE and share it between the
 /// G1 beneficial-mana-tap hold (rung 5) and this rung-9 sac check, avoiding the
 /// PR #5229 double-evaluation of the mana-action sweep.
+///
+/// `Iterator::any(p)` and `.filter(p).next().is_some()` agree on every input
+/// (both are the first-match short-circuit over the same predicate), so this
+/// re-expression leaves the loop-firewall and auto-pass answers unchanged.
 fn mana_actions_include_meaningful_sacrifice(
     state: &GameState,
     object_mana_actions: &[GameAction],
 ) -> bool {
-    object_mana_actions.iter().any(|action| {
-        matches!(
-            action,
-            GameAction::ActivateAbility {
-                source_id,
-                ability_index,
-            } if activate_ability_is_meaningful_priority(state, *source_id, *ability_index)
-        )
-    })
+    meaningful_sacrifice_mana_actions(state, object_mana_actions)
+        .next()
+        .is_some()
 }
 
 /// True when `actions` contains a priority action that materially changes the
@@ -1387,24 +1411,22 @@ pub fn has_meaningful_priority_action(state: &GameState, actions: &[GameAction])
         || has_activatable_sacrifice_for_mana(state)
 }
 
-/// PR-7 Phase 4c (LOW-2): CR 732.2b/c + CR 104.4b — an AI opponent polled on an
-/// OPTIONAL loop shortcut. The offer is raised only for optional loops the polled
-/// player CAN break (a Path A optional drain, `WaitingFor::LoopShortcut`); the win
-/// is a loss condition for every polled opponent (single-faller: see
-/// `interactive_loop_bridge`'s Path A gate, CR 104.2a). SELF-PRESERVATION: if this
-/// player has a meaningful priority action (a way to break the loop), name
-/// `Shorten{0}` — the engine realizes Shorten as a real `WaitingFor::Priority`
-/// window (`game::engine::apply_action`'s `RespondToShortcut(Shorten)` arm) where
-/// it can act — rather than Accept its own loss. No meaningful action ⇒ Accept
-/// (nothing to do). Single authority for all 3 `RespondToShortcut` emission sites
-/// (engine `candidates.rs` + phase-ai `projection.rs`/`search.rs`) so the
-/// self-preservation heuristic can't drift between them. Reuses the exact
-/// `no_living_player_has_meaningful_priority_action` probe recipe
-/// (`game::engine`), scoped to the single polled player.
-pub fn smart_shortcut_response(
+/// The probe [`smart_shortcut_response`] folds over: `state` re-parked at
+/// `polled_player`'s priority with auto-pass cleared and layers flushed, plus
+/// that state's flat priority actions. Mirrors
+/// `game::engine`'s `no_living_player_has_meaningful_priority_action` recipe,
+/// scoped to the single polled player.
+///
+/// Public because the shortcut tests' reach-guards must assert on THIS list and
+/// THIS state. A private copy of the recipe in a test silently starts measuring
+/// a different action set — and a different `has_meaningful_priority_action`
+/// answer, since that predicate's sacrifice-for-mana rung only fires while
+/// `waiting_for` is `Priority`, which is true of the probe state and false of
+/// the `RespondToShortcut` state the caller holds.
+pub fn shortcut_probe(
     state: &GameState,
     polled_player: PlayerId,
-) -> crate::analysis::loop_check::ShortcutResponse {
+) -> (casting::PriorityCastProbe, Vec<GameAction>) {
     let mut probe_state = state.clone();
     probe_state.auto_pass.clear();
     probe_state.priority_player = polled_player;
@@ -1414,11 +1436,167 @@ pub fn smart_shortcut_response(
     layers::flush_layers(&mut probe_state);
     let probe = casting::PriorityCastProbe::from_flushed_state(probe_state, polled_player);
     let actions = flat_priority_actions_with_probe(probe.state(), Some(&probe));
-    if has_meaningful_priority_action(probe.state(), &actions) {
+    (probe, actions)
+}
+
+/// The actions [`smart_shortcut_response`]'s stage 2 classifies, given a probe
+/// state and the flat priority list stage 1 read.
+///
+/// COVERAGE INVARIANT — stage 2 must classify every action stage 1 counted as
+/// meaningful. The flat list does not satisfy that on its own:
+/// [`has_meaningful_priority_action`] is a disjunction, and its second rung
+/// ([`has_activatable_sacrifice_for_mana`]) reads `state`, not `actions`, so it
+/// counts sacrifice-for-mana activations that `legal_actions` structurally omits
+/// (they live in `legal_actions_by_object` only — see issue #544). Folding stage
+/// 2 over the flat list alone therefore lets a seat whose ONLY meaningful action
+/// is such an activation Accept BY OMISSION: an action stage 2 never sees
+/// reaches no arm, so `shortcut_efficacy`'s fail-closed `_ => MayInterfere`
+/// default cannot protect it, and a false Accept is the direction the module doc
+/// says can lose a game.
+///
+/// `activatable_object_mana_actions` is gated on `waiting_for` — the probe state
+/// is re-parked at `Priority`, which is the same gate
+/// `has_activatable_sacrifice_for_mana` passes there, so the two stages read the
+/// SAME sweep. The added items come from
+/// [`meaningful_sacrifice_mana_actions`], the single authority stage 1's rung is
+/// also defined in terms of; re-testing the shape here would be a parallel copy
+/// free to drift.
+///
+/// Deliberately does NOT add the whole mana sweep, only the subset stage 1
+/// counts: a plain land tap is not a meaningful priority action to stage 1, and
+/// widening past that would buy `Shorten`s stage 1 never asked for.
+///
+/// Public for the same reason [`shortcut_probe`] is: the shortcut tests'
+/// coverage row must assert on THIS set, and a private copy of the recipe would
+/// silently start measuring a different one.
+pub fn stage_two_action_set(
+    probe_state: &GameState,
+    flat_actions: &[GameAction],
+) -> Vec<GameAction> {
+    let mana_actions = activatable_object_mana_actions(probe_state);
+    flat_actions
+        .iter()
+        .chain(meaningful_sacrifice_mana_actions(
+            probe_state,
+            &mana_actions,
+        ))
+        .cloned()
+        .collect()
+}
+
+/// PR-7 Phase 4c (LOW-2): CR 732.2b/c + CR 104.4b — an AI opponent polled on an
+/// OPTIONAL or bounded loop shortcut. The win is a loss condition for every
+/// polled opponent (single-faller: see `interactive_loop_bridge`'s Path A gate,
+/// CR 104.2a).
+///
+/// SINGLE AUTHORITY, and what it does NOT cover. MEASURED — production (i.e.
+/// non-`#[cfg(test)]`) code under `crates/*/src/` builds a
+/// `GameAction::RespondToShortcut` value at exactly four places:
+///   * `ai_support::candidates::candidate_actions_broad_with_probe`,
+///   * `phase_ai::projection::resolve_choice`,
+///   * `phase_ai::search::fallback_action` — these three are AI seats and all
+///     route here, so the heuristic cannot drift between them; and
+///   * `game::interaction::materialize_shortcut_reply_response`, which turns a
+///     HUMAN player's submitted `InteractionShortcutReply` into the action. That
+///     one is deliberately NOT covered: running a stated human choice through an
+///     AI heuristic would overwrite it. It is also the only site that can emit a
+///     non-zero `at_iteration` — every AI site emits `Shorten { at_iteration: 0 }`.
+///
+/// `candidate_actions_broad_with_probe` additionally routes
+/// `WaitingFor::RespondToPrecastCopyShortcut` here and maps the answer onto
+/// `PrecastCopyShortcutResponse`, so this function answers BOTH accept-or-shorten
+/// windows. That is intended: both windows ask the identical question — is a real
+/// priority window worth taking here — so a seat whose only action cannot touch
+/// the loop should decline both. `shorten_efficacy.rs`'s
+/// `v8_precast_window_takes_the_same_efficacy_answer` measures that window rather
+/// than assuming it.
+///
+/// TWO STAGES.
+///
+/// **Stage 1 — POSSIBILITY, byte-identical to the shipped predicate.** Reuses
+/// the exact `no_living_player_has_meaningful_priority_action` probe recipe
+/// (`game::engine`), scoped to the single polled player. No meaningful priority
+/// action ⇒ Accept, exactly as before. `has_meaningful_priority_action` is
+/// wired into the CR 732.5 NON-COMPULSION rule —
+/// "No player can be forced to perform an action that would end a loop other
+/// than actions called for by objects involved in the loop" — and is untouched
+/// here.
+///
+/// **Stage 2 — EFFICACY (AI POLICY, no CR licence claimed).** CR 732.2b grants
+/// an unconditioned accept-or-shorten option and CR 732.2c requires only that a
+/// shortening player's next choice be
+/// *different*, which a fetchland activation already satisfies; nothing in the
+/// Comprehensive Rules states an efficacy criterion. This stage is therefore
+/// policy, and it declines to burn a real priority window on a response that
+/// cannot change the outcome:
+///   - arm (A): the offer's own predicted result already crowns this seat, so
+///     shortening moves the game away from a win it already holds;
+///   - arm (B): every action stage 1 counted as meaningful is confined to this
+///     seat's own resources (`shortcut_efficacy::any_action_may_interfere`), so
+///     no available choice touches the loop. Scoped to what stage 1 counted, not
+///     to "every action this seat could take": the stage-2 set is a SUPERSET of,
+///     and never smaller than, what stage 1 counted — the flat list enters
+///     whole, so it also carries actions the stage-1 fold
+///     (`flat_actions_have_meaningful_priority`) does not count as meaningful,
+///     `PassPriority` and mana-ability activations among them, and extra items
+///     can only push toward `MayInterfere`. See the coverage invariant at the
+///     call below; the wider phrasing would claim coverage of shapes neither
+///     stage enumerates.
+///
+/// Otherwise the seat still Shortens and gets its window
+/// (`game::engine::apply_action`'s `RespondToShortcut(Shorten)` arm).
+///
+/// READ-ORDER: the proposal is read off the ORIGINAL `state`, before
+/// [`shortcut_probe`] re-parks its clone at `Priority` — the probe state carries
+/// no offer at all, so reading the crown from it would make arm (A) dead code.
+pub fn smart_shortcut_response(
+    state: &GameState,
+    polled_player: PlayerId,
+) -> crate::analysis::loop_check::ShortcutResponse {
+    // Both accept-or-shorten windows are named, so neither is answered by
+    // accident. The wildcard is not removable — `WaitingFor` has 128 variants and
+    // enumerating 126 `=> None` arms here would be noise, not a guard — and this
+    // two-named-arms + `_` shape is the module's existing idiom for the same
+    // question (`game::precast_copy_shortcut::normalize_untrusted_restore`,
+    // `::rekey_after_trusted_restore`).
+    let crowned_winner = match &state.waiting_for {
+        WaitingFor::RespondToShortcut { proposal, .. } => proposal.predicted_winner,
+        // STRUCTURAL, not an oversight: `RespondToPrecastCopyShortcut` carries no
+        // proposal summary and therefore no `predicted_winner` field, so the
+        // pre-cast route has no crown to read and arm (A) is inapplicable rather
+        // than skipped. Stage 1 and arm (B) do apply, and both run below.
+        WaitingFor::RespondToPrecastCopyShortcut { .. } => None,
+        _ => None,
+    };
+
+    let (probe, actions) = shortcut_probe(state, polled_player);
+    if !has_meaningful_priority_action(probe.state(), &actions) {
+        // CR 732.2c: nothing to do — agree to take the shortcut.
+        return crate::analysis::loop_check::ShortcutResponse::Accept;
+    }
+
+    // Arm (A). Keyed on `predicted_winner`, never `proposer`: CR 732.2a lets a
+    // player propose a shortcut whose outcome crowns someone else, and the
+    // proposer is excluded from the APNAP response queue anyway.
+    if crowned_winner == Some(polled_player) {
+        return crate::analysis::loop_check::ShortcutResponse::Accept;
+    }
+
+    // Arm (B). No bounded/unbounded exemption: an `IterationCount::Fixed(n)`
+    // offer ends the game just as surely as an `UntilLethal` one, so both
+    // classes take the identical rule.
+    //
+    // The fold reads [`stage_two_action_set`], never the bare flat `actions` —
+    // see that function for the coverage invariant and what folding over the
+    // flat list alone would silently Accept.
+    if shortcut_efficacy::any_action_may_interfere(
+        probe.state(),
+        polled_player,
+        &stage_two_action_set(probe.state(), &actions),
+    ) {
         // CR 732.2b: name an earlier stopping point — take my window instead of losing.
         crate::analysis::loop_check::ShortcutResponse::Shorten { at_iteration: 0 }
     } else {
-        // CR 732.2c: nothing to do — agree to take the shortcut.
         crate::analysis::loop_check::ShortcutResponse::Accept
     }
 }