fix(graph): build peer retirement guards before spawning their tasks

A task aborted before its first poll never runs its body, so a guard
constructed inside the future did not exist yet and the pre-provisioned
teammate identity stayed messageable. Built outside and moved in, the
guard drops with the unpolled future on both the frontier and map items.

Also rewords the two timeout: 0 validator warnings: the graph-level one
no longer claims max_loop_iterations always applies (it may be 0 too),
and the node-level one names the enclosing bounds that still end the
node instead of saying it runs until it returns.

src/graph/executor.rs

@@ -247,20 +247,21 @@ impl GraphExecutor {
                 #[cfg(test)]
                 let observer = frontier_observer.clone();
 
+                // Retires the teammate identity however this task ends:
+                // dropped explicitly right after step() on the normal path,
+                // by unwinding on the abort return, task abort, or panic.
+                // Built before the spawn so a task aborted before its first
+                // poll still drops it with the future.
+                let retire = registry_for_task
+                    .zip(peer_id)
+                    .map(|(registry, id)| PeerRetireGuard {
+                        registry,
+                        id,
+                        #[cfg(test)]
+                        observer,
+                    });
                 let task = tokio::spawn(async move {
-                    // Retires the teammate identity however this task ends:
-                    // dropped explicitly right after step() on the normal
-                    // path, by unwinding on the abort return, task abort, or
-                    // panic.
-                    let retire =
-                        registry_for_task
-                            .zip(peer_id)
-                            .map(|(registry, id)| PeerRetireGuard {
-                                registry,
-                                id,
-                                #[cfg(test)]
-                                observer,
-                            });
+                    let retire = retire;
                     let _permit = sem_clone
                         .acquire()
                         .await
@@ -1722,6 +1723,48 @@ nodes:
         assert!(!registry.is_finished(&id_b), "the sibling is untouched");
     }
 
+    /// A task aborted before its first poll never runs its body, so a guard
+    /// built inside the future would never exist. Built outside and moved in,
+    /// it drops with the unpolled future and still retires the identity.
+    #[tokio::test]
+    async fn peer_retire_guard_fires_when_task_is_aborted_before_first_poll() {
+        let (registry, assignments) = provision_frontier_peers(vec![
+            ("a".to_string(), "worker".to_string()),
+            ("b".to_string(), "worker".to_string()),
+        ]);
+        let registry = registry.expect("two flagged nodes should get a registry");
+        let id_a = assignments["a"].0.clone();
+        let id_b = assignments["b"].0.clone();
+
+        let outside = PeerRetireGuard {
+            registry: Arc::clone(&registry),
+            id: id_a.clone(),
+            observer: None,
+        };
+        let inside_registry = Arc::clone(&registry);
+        let inside_id = id_b.clone();
+        let task = tokio::spawn(async move {
+            let _outside = outside;
+            let _inside = PeerRetireGuard {
+                registry: inside_registry,
+                id: inside_id,
+                observer: None,
+            };
+            std::future::pending::<()>().await;
+        });
+        // No await between spawn and abort: on the current-thread test
+        // runtime the task has not been polled yet.
+        task.abort();
+        let err = task.await.expect_err("the task was aborted");
+
+        assert!(err.is_cancelled(), "{err}");
+        assert!(registry.is_finished(&id_a), "the moved-in guard retires");
+        assert!(
+            !registry.is_finished(&id_b),
+            "a guard built inside never existed"
+        );
+    }
+
     /// A session already aborted when the run starts trips the synchronous
     /// pre-check: no bridge task, no super-step, the start script never runs.
     #[tokio::test]

src/graph/map.rs

@@ -92,18 +92,20 @@ async fn run_map(
         let sem = semaphore.clone();
         let abort = step_ctx.abort_signal.clone();
 
+        // Cancellation net behind the chain runner's own retirement: a task
+        // aborted before or during its chain still leaves the item's identity
+        // finished. Built before the spawn so a task aborted before its first
+        // poll drops it with the future.
+        let retire = item_peers
+            .as_ref()
+            .map(|(registry, assignment)| PeerRetireGuard {
+                registry: Arc::clone(registry),
+                id: assignment.0.clone(),
+                #[cfg(test)]
+                observer: None,
+            });
         let task = tokio::spawn(async move {
-            // Cancellation net behind the chain runner's own retirement: a
-            // task aborted before or during its chain still leaves the item's
-            // identity finished.
-            let _retire = item_peers
-                .as_ref()
-                .map(|(registry, assignment)| PeerRetireGuard {
-                    registry: Arc::clone(registry),
-                    id: assignment.0.clone(),
-                    #[cfg(test)]
-                    observer: None,
-                });
+            let _retire = retire;
             let _permit = sem
                 .acquire()
                 .await

src/graph/validator.rs

@@ -628,8 +628,8 @@ impl GraphValidator {
         if graph.settings.timeout == Some(0) {
             result.warning(ValidationError::new(
                 "settings.timeout: 0 disables the graph wall-clock bound; the run \
-                 ends only when the graph completes, is aborted, or hits \
-                 max_loop_iterations",
+                 ends only when the graph completes, is aborted, or another \
+                 limit stops it",
             ));
         }
         for (node_id, node) in &graph.nodes {
@@ -655,7 +655,8 @@ impl GraphValidator {
                 node_id,
                 format!(
                     "timeout: 0 disables the wall-clock bound for {kind} node \
-                     '{node_id}'; it runs until it returns"
+                     '{node_id}'; it runs until it returns or an enclosing bound \
+                     (settings.timeout, an outer agent node's timeout, or abort) ends it"
                 ),
             ));
         }
@@ -3293,7 +3294,7 @@ mod tests {
         assert_eq!(
             w[0].message,
             "settings.timeout: 0 disables the graph wall-clock bound; the run ends only when \
-             the graph completes, is aborted, or hits max_loop_iterations"
+             the graph completes, is aborted, or another limit stops it"
         );
     }
 
@@ -3312,7 +3313,8 @@ mod tests {
         assert_eq!(w[0].node_id.as_deref(), Some("a"));
         assert_eq!(
             w[0].message,
-            "timeout: 0 disables the wall-clock bound for agent node 'a'; it runs until it returns"
+            "timeout: 0 disables the wall-clock bound for agent node 'a'; it runs until it returns \
+             or an enclosing bound (settings.timeout, an outer agent node's timeout, or abort) ends it"
         );
     }
 
@@ -3332,7 +3334,8 @@ mod tests {
         assert_eq!(w[0].node_id.as_deref(), Some("s"));
         assert_eq!(
             w[0].message,
-            "timeout: 0 disables the wall-clock bound for script node 's'; it runs until it returns"
+            "timeout: 0 disables the wall-clock bound for script node 's'; it runs until it returns \
+             or an enclosing bound (settings.timeout, an outer agent node's timeout, or abort) ends it"
         );
     }
 
@@ -3352,7 +3355,8 @@ mod tests {
         assert_eq!(w[0].node_id.as_deref(), Some("l"));
         assert_eq!(
             w[0].message,
-            "timeout: 0 disables the wall-clock bound for llm node 'l'; it runs until it returns"
+            "timeout: 0 disables the wall-clock bound for llm node 'l'; it runs until it returns \
+             or an enclosing bound (settings.timeout, an outer agent node's timeout, or abort) ends it"
         );
     }
 
@@ -3372,7 +3376,8 @@ mod tests {
         assert_eq!(w[0].node_id.as_deref(), Some("r"));
         assert_eq!(
             w[0].message,
-            "timeout: 0 disables the wall-clock bound for rag node 'r'; it runs until it returns"
+            "timeout: 0 disables the wall-clock bound for rag node 'r'; it runs until it returns \
+             or an enclosing bound (settings.timeout, an outer agent node's timeout, or abort) ends it"
         );
     }