diff --git a/packages/chess/src/net/prediction.ts b/packages/chess/src/net/prediction.ts
index b91ee20..66bc95a 100644
--- a/packages/chess/src/net/prediction.ts
+++ b/packages/chess/src/net/prediction.ts
@@ -130,10 +130,23 @@ export class PredictionManager {
    * previously-legal optimistic moves, and re-validating them here would
    * duplicate server logic. Simpler to let the next user action
    * re-predict against the freshly-synced base.
+   *
+   * We route through `setActivePresets` (not bare `activePresets.replaceAll`)
+   * so preset `onActivate` / `onDeactivate` lifecycle hooks fire on the
+   * client's base engine. That's essential for rules like `piece-hp`
+   * which install per-piece state (Hp facts) from onActivate — without
+   * this the client would know the preset is active but have no Hp
+   * facts to render, because `game.presets` carries only the
+   * activation list, not the resulting facts.
+   *
+   * Idempotence guarantee: preset hooks are expected to be idempotent
+   * (piece-hp.onActivate only inserts Hp when absent). Server and
+   * client run the same hook logic, so both arrive at the same state.
+   * The next `game.state` or `game.delta` reconciles any drift.
    */
   private applyPresets(payload: GamePresetsPayload): void {
     try {
-      this.baseEngine.activePresets.replaceAll(payload.activations);
+      this.baseEngine.setActivePresets(payload.activations);
     } catch {
       // A bad set from the server shouldn't crash the client; the server
       // already validated, so this branch is defensive only. We clear