fix(@projects/@magic-civilization): 🐛 revert dominion_bench ai integration

Co-Authored-By: Lilith Autocommit <noreply@atlilith.com>

.project/designs/p2-80-bullet2-port-sizing.md

@@ -159,3 +159,29 @@ not `mc-turn`'s inline stub.
 **retiring `mc-turn`'s inline `nearest_lair` movement** (dead-stub removal,
 no-debt). Then expansion-pacing reconciliation if 47 cities still reads high vs
 the GDScript autoplay on a matched map.
+
+---
+
+## Step 1 follow-up (2026-06-08) — `dominion_bench` is the wrong probe; reverted
+
+Attempted to make `dominion_bench` drive the real AI (`drive_ai_slot` →
+`run_ai_turn`) so the multi-clan bench would show combat. **It did not work and
+was reverted** (commits exposing `drive_ai_slot`, adding the `mc-player-api`
+dep, and the bench drive-loop are undone — clean, no dead code left).
+
+**Finding:** with `drive_ai_slot` wired in, the bench result was **byte-identical**
+(214 cities, turn 51, 0 PvP) — even after deriving real per-personality
+`ScoringWeights::from_personality`. `drive_ai_slot` has **zero net effect** on
+`dominion_bench`'s hand-built `GameState`: its `run_ai_turn` actions don't apply
+(the minimal bench state lacks the structure the real AI/action-application
+expects). Making `dominion_bench` representative would require rebuilding its
+state to full-harness fidelity — surgery, not a wire-up.
+
+**Resolution — no new probe needed:** `full_game_transcript` /
+`long_game_transcript` (mc-player-api tests) **already are** the working
+real-AI multi-clan bench — they build a realistic `GameState`, drive
+`run_ai_turn` via the production `apply_end_turn` → `drive_ai_slot` path, and
+produce live-grade games (**2531 kills**, conquest, 47-city winner over 250
+turns). Use those as the convergence harness for the cutover; leave
+`dominion_bench` as the fast inline-AI economic/expansion bench it was designed
+to be. The real-AI viability question (step 1) is **already answered** there.

src/simulator/crates/mc-player-api/src/dispatch.rs

@@ -976,12 +976,7 @@ fn translate_processor_events(events: &[mc_replay::TurnEvent]) -> Vec<Event> {
 ///      counted but DO NOT abort the rest of the turn — a single
 ///      unknown unit or illegal move must not collapse the entire AI
 ///      side (matches the GDScript dispatch behaviour).
-/// Drive one AI slot for a turn: project vision, run the slot's controller
-/// (`run_ai_turn` by default), and apply the chosen actions. Public so headless
-/// multi-clan benches (`dominion_bench`) can drive the *real* AI through the
-/// same path the interactive game uses (`apply_end_turn`), instead of relying
-/// on `mc-turn`'s inline fallback movement. Returns the action count.
-pub fn drive_ai_slot(state: &mut GameState, ai_slot: u8) -> u32 {
+fn drive_ai_slot(state: &mut GameState, ai_slot: u8) -> u32 {
     let pi: usize = ai_slot as usize;
     if pi >= state.players.len() {
         return 0;

src/simulator/crates/mc-sim/Cargo.toml

@@ -23,7 +23,6 @@ mc-city      = { path = "../mc-city" }
 mc-culture   = { path = "../mc-culture" }
 mc-economy   = { path = "../mc-economy" }
 mc-ai        = { path = "../mc-ai" }
-mc-player-api = { path = "../mc-player-api" }
 serde.workspace      = true
 serde_json.workspace = true
 rayon           = "1"

src/simulator/crates/mc-sim/src/bin/dominion_bench.rs

@@ -355,20 +355,7 @@ fn run_scenario_with_profiles(num_players: usize, all_profiles: &[ProfileJson])
                 ..Default::default()            }],
             unit_upkeep: vec![],
             strategic_axes: make_axes(profile),
-            // Real per-personality scoring weights (military/expansion/etc.),
-            // the same source the live game uses (`ScoringWeights::from_personality`).
-            // Without these the AI scores every action ~0 and `drive_ai_slot`
-            // produces no moves — the inert `default()` was why the bench showed
-            // zero PvP. Falls back to default if the personality file is absent.
-            scoring_weights: ScoringWeights::from_personality(
-                &profile.id,
-                &Path::new(env!("CARGO_MANIFEST_DIR"))
-                    .ancestors()
-                    .nth(2)
-                    .expect("workspace root")
-                    .join("public/resources/ai/personalities"),
-            )
-            .unwrap_or_default(),
+            scoring_weights: ScoringWeights::default(),
             expansion_points: 0,
             city_buildings: vec![vec![]],
             city_improvements: Default::default(),
@@ -426,14 +413,6 @@ fn run_scenario_with_profiles(num_players: usize, all_profiles: &[ProfileJson])
     let mut victory_turn: Option<u32> = None;
 
     for turn_num in 1..=TOTAL_TURNS {
-        // Drive every slot's REAL AI (`run_ai_turn` via the controller registry)
-        // before the turn resolves — the same `drive_ai_slot` the interactive
-        // game runs inside `apply_end_turn`. Without this the bench relied on
-        // `mc-turn`'s inline `nearest_lair` fallback (PvE only → zero PvP). All
-        // slots are AI here (no human), so we drive all of them.
-        for slot in 0..state.players.len() as u8 {
-            mc_player_api::dispatch::drive_ai_slot(&mut state, slot);
-        }
         let result = processor.step(&mut state);
 
         // Track unit production per player.