diff --git a/.project/objectives/p2-67-claude-player-api.md b/.project/objectives/p2-67-claude-player-api.md
index 16f466fb..bc5bb292 100644
--- a/.project/objectives/p2-67-claude-player-api.md
+++ b/.project/objectives/p2-67-claude-player-api.md
@@ -1233,3 +1233,92 @@ closed and the policy is now exercising those branches.
 - Combat overflow at `mc-turn/src/processor.rs:2425` did NOT reproduce
   in this 250-turn run. The harness PvP path apparently exercises a
   different code path than the bench. Tracked under p2-67-followup.
+
+## 2026-05-12 — Real-game analysis follow-up: Bug 3 (attacks don't kill)
+
+### RC
+
+Long-game recap (`.local/demo-runs/2026-05-12-claude-vs-easy-ai-250-turn/recap.md`)
+showed `Units killed (UnitDestroyed events): 0` over 237 turns despite 5
+attack actions and 951 fortify actions. Two compounding bugs:
+
+- **Bug A — `mc-turn`:** `resolve_single_pvp_attack`'s `Killed` /
+  `Destroyed` / `!attacker_survived` branches (`processor.rs:2074-2126`)
+  called `swap_remove` on the dead unit but never pushed a
+  `TurnEvent::UnitKilled`. The discovery-loop emit at
+  `processor.rs:2702` only fires for proximity-discovered combats —
+  queued `AttackRequest`s (the GDScript click path + `apply_action`
+  dispatch path) were silent.
+- **Bug B — `mc-player-api`:** `translate_processor_events`
+  (`dispatch.rs:411`) explicitly enumerated `TurnEvent::UnitKilled` in
+  the drop-set, so even when the chronicle DID carry a kill, the wire
+  layer threw it away. `Event::UnitDestroyed` had a struct definition
+  but no producer.
+
+### Fix
+
+- Added `unit_id: u32` field to `TurnEvent::UnitKilled` (`#[serde(default)]`
+  for legacy fixture compat).
+- New `UnitKilledEvent` struct in `mc-turn::combat_event` + new
+  `units_killed: Vec<UnitKilledEvent>` field on `PendingCaptureEvents`.
+  `drain_into` translates each entry to a `TurnEvent::UnitKilled` push.
+- Each `swap_remove` call site in `resolve_single_pvp_attack` (3 sites:
+  `Killed`, `Destroyed`, `!attacker_survived`) now stages a
+  `UnitKilledEvent` before mutating the unit vec.
+- Discovery-loop emit at `processor.rs:2702` populates `unit_id`.
+- `dispatch.rs::translate_processor_events` now translates
+  `TurnEvent::UnitKilled { unit_id, .. }` → `Event::UnitDestroyed { unit_id: format!("u_{unit_id}"), killer_unit_id: None }`. `killer_unit_id` correlation deferred (chronicle carries clan-id, not attacker unit id).
+- `api-gdext::replay::dict_for_event` includes the new `unit_id` field
+  in the GDScript-visible dictionary. Also wired
+  `CityBuildingCompleted` translation that was previously latent.
+
+### Evidence
+
+- `crates/mc-turn/tests/queued_pvp_unit_killed.rs` (new) — minimal
+  2-player fixture, single queued `AttackRequest` against a 1-HP
+  defender, asserts exactly one `TurnEvent::UnitKilled { defender =
+  ClanId(1), unit_id = 200, attacker = ClanId(0) }` lands in
+  `result.events_emitted`. Passes.
+- `crates/mc-player-api/tests/queued_attack_unit_destroyed.rs` (new)
+  — same fixture via `apply_action(Attack)` + `apply_action(EndTurn)`,
+  asserts exactly one `Event::UnitDestroyed { unit_id: "u_200", .. }`
+  comes back from dispatch. Passes.
+- `long_game_transcript` re-run (250-turn ceiling, terminated at turn
+  236/237 for "10 consecutive EndTurn-only" stuck-guard):
+  - **before**: `Units killed (UnitDestroyed events): 0`
+  - **after**: `Units killed (UnitDestroyed events): 1998`
+
+### Files touched
+
+- `src/simulator/crates/mc-replay/src/event.rs` — added `unit_id` to
+  `TurnEvent::UnitKilled`.
+- `src/simulator/crates/mc-replay/tests/award_computation.rs` — fixture
+  updates.
+- `src/simulator/crates/mc-replay/src/awards.rs` — unchanged (uses
+  `attacker` field only; `..` rest covers the new field).
+- `src/simulator/crates/mc-turn/src/combat_event.rs` — new
+  `UnitKilledEvent` struct.
+- `src/simulator/crates/mc-turn/src/game_state.rs` — extended
+  `PendingCaptureEvents` (new vec + drain + is_empty).
+- `src/simulator/crates/mc-turn/src/processor.rs` — staged event in 3
+  swap-remove sites; populated `unit_id` in discovery-loop emit;
+  imported `UnitKilledEvent`.
+- `src/simulator/crates/mc-player-api/src/dispatch.rs` — removed
+  `UnitKilled` from drop-set; added translation arm.
+- `src/simulator/api-gdext/src/replay.rs` — `unit_id` field in
+  `UnitKilled` dict; new `CityBuildingCompleted` arm.
+
+### Residual / deferred
+
+- `killer_unit_id` on `Event::UnitDestroyed` stays `None` — the
+  chronicle variant carries only the killing clan-id, not the specific
+  attacker unit id. Wiring that requires a new field on
+  `TurnEvent::UnitKilled` (`killer_unit_id: u32`) plus threading it
+  through `resolve_single_pvp_attack` and the discovery loop. Tracked
+  for a future follow-up; not blocking dispatch-event correctness.
+- Destroy-posture civilian kills now emit BOTH
+  `TurnEvent::CivilianDestroyed` AND `TurnEvent::UnitKilled`. The wire
+  layer drops `CivilianDestroyed` and surfaces `UnitDestroyed` via
+  `UnitKilled`, so net wire behaviour is correct. AI memory /
+  chronicle subscribers that consume `CivilianDestroyed` separately
+  are unaffected.