fleetd #342: base tab-cleanup teardown on the pane's real placement, not a delegate's static config

HerdrPeerLauncher.stop() used to gate spaces.locatePane() on usesTabPlacement(),
which reads the delegate's OWN configured profiles. When CompositePeerLauncher's
single-daemon stop() shortcut hands a pane to a delegate that never spawned it
(spawnedBy empty after a daemon restart, herdrDaemonCount()==1), that delegate's
placement config says nothing true about how the pane was actually placed, and a
dedicated tab could be skipped and leaked.

Resolve the tab unconditionally instead — WorkspaceControl#locatePane already
tolerates a missing pane by returning null — and let the existing single-occupant
check (tabPaneCount()==1) be the only thing that decides whether to close it, same
as it already protects a shared tab regardless of declared placement.

Adds a mixed-placement CompositePeerLauncherTest (every existing stop-fallback test
configured both adapters as tab placement, so the mis-routing never showed) and
updates FleetAppTest#stopWorkerInPanePlacementClosesOnlyThePane, whose old
assertion (no pane.get on pane placement) documented exactly the skip this fix
removes.

fleetd/src/main/java/dev/ltms/fleet/config/ConfigRef.java

@@ -154,10 +154,13 @@ import java.util.function.Supplier;
  * {@link ConfigRefProfileCoverageTest} shape (one level up, over {@code FleetConfig} itself rather
  * than {@code FleetConfig.Profile}) proves this file's four classes exhaust the record's components
  * — see {@code ConfigRefTopLevelCoverageTest}. That test proves the record's <em>shape</em> is fully
- * triaged; it does NOT prove a {@code SPLIT_KEYS}/{@code COLD_KEYS} member has any reporting code
- * behind it at all — {@code ConfigRefTopLevelReportingCoverageTest} is what fleetd #333 added for
- * that, after measuring that a {@code SPLIT_KEYS} entry with its reporting branch deleted passes
- * both this file's own "kept in step" assert and {@code ConfigRefTopLevelCoverageTest} unchanged.
+ * triaged; it does NOT prove a {@code SPLIT_KEYS}/{@code COLD_KEYS}/{@code DEFERRED_KEYS} member has
+ * any reporting code behind it at all — {@code ConfigRefTopLevelReportingCoverageTest} is what
+ * fleetd #333 added for that, after measuring that a {@code SPLIT_KEYS} entry with its reporting
+ * branch deleted passes both this file's own "kept in step" assert and
+ * {@code ConfigRefTopLevelCoverageTest} unchanged. fleetd #337 extended it to {@code DEFERRED_KEYS}
+ * after measuring the same one-way gap there directly: dropping {@code guard}'s branch out of
+ * {@link #changedDeferredKeys} while {@code "guard"} stayed in the set left the whole suite green.
  *
  * <p><strong>A cold change refuses the whole reload.</strong> Not the hot half applied and the cold
  * half warned about: that would leave the running daemon in a state matching no file on disk, which
@@ -193,6 +196,25 @@ public final class ConfigRef implements Supplier<FleetConfig> {
      */
     static final Set<String> SPLIT_KEYS = Set.of("health", "coordinator", "fleet");
 
+    /**
+     * Top-level keys {@link #changedDeferredKeys} compares — see the class doc's Deferred bullet.
+     * Promoted here from a test-side copy in {@code ConfigRefTopLevelCoverageTest} by fleetd #337,
+     * the same reason {@link #COLD_KEYS} and {@link #SPLIT_KEYS} live here rather than in a test: a
+     * second, hand-maintained copy of this set is exactly the kind of thing that silently drifts
+     * from the method it is supposed to describe. {@code spawnReadyTimeoutMs} and
+     * {@code spawnReadyPollMs} are compared together in one branch and reported under the combined
+     * label {@code "spawnReady*"}; {@code profiles} is compared twice over (added/removed names,
+     * then an existing profile's launch settings) — see {@link #changedDeferredKeys}.
+     *
+     * <p>Package-private (not {@code private}) so {@code ConfigRefTopLevelCoverageTest} and
+     * {@code ConfigRefTopLevelReportingCoverageTest} can both read it, the same way they already
+     * read {@link #COLD_KEYS} and {@link #SPLIT_KEYS}.
+     */
+    static final Set<String> DEFERRED_KEYS = Set.of(
+            "guard", "worktreeRoot", "worktreeGroup", "primary", "configReload",
+            "leadHeartbeat", "lifecycle", "spawnReadyTimeoutMs", "spawnReadyPollMs",
+            "quarantineCooldownSeconds", "profiles");
+
     private final Path path;
     private final AtomicReference<FleetConfig> current;
 
@@ -350,8 +372,16 @@ public final class ConfigRef implements Supplier<FleetConfig> {
         return changed;
     }
 
-    /** Changed keys that were accepted but whose effect waits for a restart. */
-    private static List<String> changedDeferredKeys(FleetConfig old, FleetConfig fresh) {
+    /**
+     * Changed keys that were accepted but whose effect waits for a restart.
+     *
+     * <p>Package-private (not {@code private}) so {@code ConfigRefTopLevelReportingCoverageTest}
+     * can call it directly with a reflection-built {@code FleetConfig} pair, the same reason
+     * {@link #changedColdKeys} and {@link #changedSplitKeys} already are (fleetd #333, extended to
+     * this method by fleetd #337 — membership in {@link #DEFERRED_KEYS} proved nothing about this
+     * method on its own until then; see that test's class doc).
+     */
+    static List<String> changedDeferredKeys(FleetConfig old, FleetConfig fresh) {
         List<String> changed = new ArrayList<>();
         if (!Objects.equals(old.lifecycle(), fresh.lifecycle())) {
             changed.add("lifecycle");

fleetd/src/main/java/dev/ltms/fleet/inject/CompletionResolver.java

@@ -613,9 +613,26 @@ public final class CompletionResolver implements TurnListener {
         return null;
     }
 
-    /** True when the error pattern begins the matched pane line, rather than appearing in prose. */
+    /**
+     * True when the error pattern begins the matched pane line, rather than appearing in prose.
+     *
+     * <p>Leading terminal chrome is skipped first — box-drawing characters, bullets, gutter bars and
+     * spaces. #339 introduced this check with a bare {@code lookingAt}, and that rejected a genuine
+     * error line rendered as {@code "| 503 Service Unavailable: ..."}: the send still failed, but the
+     * credential outage was never recorded. That is the false negative #339's own invariant 3 called
+     * worse than the false positive it set out to fix — measured with a throwaway probe on the
+     * raw-scrape path, which is exactly the path whose comment says to expect leading chrome.
+     *
+     * <p>Skipping only a leading run of non-letter, non-digit characters keeps the fix's intent. A
+     * member's prose ({@code "I checked the retry path. An API Error: makes it back off."}) still
+     * does not match, because there the pattern sits after words, not after chrome.
+     */
     private static boolean startsWithBackendError(String line, Pattern pattern) {
-        return pattern.matcher(line).lookingAt();
+        int i = 0;
+        while (i < line.length() && !Character.isLetterOrDigit(line.charAt(i))) {
+            i++;
+        }
+        return pattern.matcher(line.substring(i)).lookingAt();
     }
 
     /**

fleetd/src/main/java/dev/ltms/fleet/inject/Injector.java

@@ -145,8 +145,58 @@ public final class Injector {
         return router != null ? router.agentsFor(target) : agents;
     }
 
+    /** The result of trying to remove an undelivered message from the injector. */
+    public enum Cancellation {
+        CANCELLED,
+        DELIVERED,
+        NOT_DELIVERED
+    }
+
+    /**
+     * An identity handle for one queued delivery. It is the only value accepted by
+     * {@link #cancel(Delivery)}, so a caller cannot cancel a different message with the same target
+     * or text.
+     */
+    public static final class Delivery {
+        private final Pending pending;
+
+        private Delivery(Pending pending) {
+            this.pending = pending;
+        }
+
+        public CompletableFuture<Void> completion() {
+            return pending.delivered;
+        }
+    }
+
     /** A pending message and the future that completes when it has been delivered. */
-    private record Pending(String text, TurnToken token, CompletableFuture<Void> delivered) {
+    private static final class Pending {
+        enum State { QUEUED, DELIVERED, NOT_DELIVERED, CANCELLED }
+
+        final String target;
+        final String text;
+        final TurnToken token;
+        final CompletableFuture<Void> delivered;
+        volatile State state = State.QUEUED; // written under the owning Target monitor
+
+        Pending(String target, String text, TurnToken token, CompletableFuture<Void> delivered) {
+            this.target = target;
+            this.text = text;
+            this.token = token;
+            this.delivered = delivered;
+        }
+
+        String text() {
+            return text;
+        }
+
+        TurnToken token() {
+            return token;
+        }
+
+        CompletableFuture<Void> delivered() {
+            return delivered;
+        }
     }
 
     /** Per-worker delivery state, guarded by its own monitor (single writer per worker). */
@@ -177,15 +227,47 @@ public final class Injector {
      * <p>Uses an atomic map update so a concurrent {@link #drop} cannot slip between "find the
      * target" and "queue the message" and orphan it in a target it just removed.
      */
-    public CompletableFuture<Void> enqueue(String target, String text, TurnToken token) {
+    public Delivery enqueue(String target, String text, TurnToken token) {
         CompletableFuture<Void> delivered = new CompletableFuture<>();
-        Pending p = new Pending(text, token, delivered);
+        Pending p = new Pending(target, text, token, delivered);
         targets.compute(target, (_, existing) -> {
             Target t = (existing != null) ? existing : new Target();
             t.add(p); // synchronized on the Target monitor — atomic with a concurrent drop
             return t;
         });
-        return delivered;
+        return new Delivery(p);
+    }
+
+    /**
+     * Cancel this exact queued delivery. The target monitor serializes this operation with
+     * {@link #onStatus}: if delivery wins that race, this returns {@link Cancellation#DELIVERED}
+     * rather than claiming the message remained queued.
+     */
+    public Cancellation cancel(Delivery delivery) {
+        Pending p = delivery.pending;
+        Target t = targets.get(p.target);
+        if (t == null) {
+            return cancellationOf(p);
+        }
+        synchronized (t) {
+            if (p.state != Pending.State.QUEUED || !t.queue.remove(p)) {
+                return cancellationOf(p);
+            }
+            p.state = Pending.State.CANCELLED;
+            if (isQuiescent(t)) {
+                targets.remove(p.target, t);
+            }
+            return Cancellation.CANCELLED;
+        }
+    }
+
+    private static Cancellation cancellationOf(Pending p) {
+        return p.state == Pending.State.DELIVERED ? Cancellation.DELIVERED : Cancellation.NOT_DELIVERED;
+    }
+
+    private static boolean isQuiescent(Target t) {
+        return t.queue.isEmpty() && !t.awaitingPickup && !t.awaitingCompletion
+                && !t.postTurnPending && !t.awaitingPostTurnPickup && !t.postTurnObserved;
     }
 
     /**
@@ -274,6 +356,7 @@ public final class Injector {
                             try {
                                 agentsFor(target).send(target, p.text());
                                 t.queue.poll();
+                                p.state = Pending.State.DELIVERED;
                                 t.awaitingPickup = true;
                                 t.awaitingCompletion = true;
                                 t.turnObserved = false;
@@ -283,6 +366,7 @@ public final class Injector {
                                 // Delivery failed at herdr; drop the poisoned message and surface it
                                 // rather than blocking the queue behind it.
                                 t.queue.poll();
+                                p.state = Pending.State.NOT_DELIVERED;
                                 sent = p;
                                 sendError = e;
                             }
@@ -293,6 +377,9 @@ public final class Injector {
                             // fail every queued message and release the target (CB-114) instead of
                             // polling it indefinitely with the caller's future never completing.
                             notReady = new ArrayList<>(t.queue);
+                            for (Pending pending : notReady) {
+                                pending.state = Pending.State.NOT_DELIVERED;
+                            }
                             log.warn("readiness grace for {} expired after {} polls ({}s): target never "
                                     + "became deliverable, so failing {} queued message(s) that never "
                                     + "reached its pane",
@@ -340,8 +427,7 @@ public final class Injector {
 
             // Reclaim the entry once the worker is fully quiescent (nothing queued, no pickup or
             // completion awaited), so the map cannot grow without bound across short-lived workers.
-            if (t.queue.isEmpty() && !t.awaitingPickup && !t.awaitingCompletion
-                    && !t.postTurnPending && !t.awaitingPostTurnPickup && !t.postTurnObserved) {
+            if (isQuiescent(t)) {
                 targets.remove(target, t);
             }
         }
@@ -432,6 +518,9 @@ public final class Injector {
         boolean hadDeliveredTurn;
         synchronized (t) {
             pending = new ArrayList<>(t.queue);
+            for (Pending p : pending) {
+                p.state = Pending.State.NOT_DELIVERED;
+            }
             t.queue.clear();
             hadDeliveredTurn = t.awaitingCompletion;
             t.awaitingCompletion = false;

fleetd/src/main/java/dev/ltms/fleet/member/HerdrPeerLauncher.java

@@ -937,14 +937,20 @@ public abstract class HerdrPeerLauncher implements PeerLauncher {
      */
     @Override
     public void stop(String idOrPane) {
-        // Teardown knows only the pane, not which profile spawned it. Attempt tab cleanup when any
-        // profile uses tab placement (so the bridge may have created a dedicated peer tab); the
-        // single-occupant check below is what actually protects the user's shared tabs.
+        // Teardown knows only the pane, not which profile spawned it — and, when this call is
+        // routed here through CompositePeerLauncher's single-daemon stop() shortcut (fleetd #342),
+        // not even which adapter's config actually governed the spawn: the shortcut can hand the
+        // pane to a delegate that never spawned it, whose own profiles say nothing about how THIS
+        // pane was placed. So the decision to look for a tab to clean up is made from the pane's
+        // actual state, not from this delegate's static profile config: resolve the tab
+        // unconditionally and let {@link WorkspaceControl#locatePane} tolerate "not found" (it
+        // returns null rather than throwing); the single-occupant check below is what actually
+        // protects the user's shared tabs, exactly as it always has.
         String paneId = paneByAgentId.remove(idOrPane);
         if (paneId == null) {
             paneId = idOrPane; // raw-pane fallback (reap, gate timeout, pane-addressed callers)
         }
-        WorkspaceControl.PaneLocation loc = usesTabPlacement() ? spaces.locatePane(paneId) : null;
+        WorkspaceControl.PaneLocation loc = spaces.locatePane(paneId);
         try {
             agents.close(paneId);
         } catch (HerdrException e) {
@@ -985,11 +991,6 @@ public abstract class HerdrPeerLauncher implements PeerLauncher {
         }
     }
 
-    /** Whether any configured profile places peers in their own tab (so tabs may need cleanup). */
-    private boolean usesTabPlacement() {
-        return profiles.values().stream().anyMatch(FleetConfig.Profile::tabPlacement);
-    }
-
     /** True when a herdr error means the target is already gone (safe to treat as done). */
     private static boolean isAlreadyGone(HerdrException e) {
         return e.code() != null && e.code().endsWith("_not_found");
@@ -1665,30 +1666,59 @@ public abstract class HerdrPeerLauncher implements PeerLauncher {
 
     /**
      * Guards the {@code effectiveAllowed == null} branch of {@link #logCredentialGap} — the
-     * genuinely-unprotected report (deny-by-default, and the allow-list non-zsh fallback) — to one
-     * WARN per launcher instance, not one per spawn.
+     * genuinely-unprotected report (deny-by-default, and the allow-list non-zsh fallback) — AND
+     * the {@code effectiveAllowed != null} / {@code keptByDerivedList} branch, the allow-list case
+     * where a name is in the gap but the derived allow-list keeps it anyway. Both branches log the
+     * same severity (WARN) about the same fact — a name genuinely reaching a member pane
+     * unprotected — so they share this one guard, keyed per NAME rather than per launcher instance:
+     * each credential-shaped name that is ever reported unprotected gets exactly one WARN, however
+     * many spawns see it and whichever of the two branches first reports it.
+     *
+     * <p>fleetd #341: {@code memberCredentials} is a live, re-read-per-spawn supplier, so the
+     * policy — and so the gap's actual member names — can change between two spawns on the same
+     * launcher. Before this fix the guard was a single {@code AtomicBoolean} tripped by either
+     * branch: spawn 1 could warn about name A and trip the flag, and a later spawn's gap containing
+     * a different name B would never be reported, even though B is just as unprotected as A was.
+     * {@code AtomicBoolean} could not express "once per distinct name" at all — only "once, ever,
+     * for whichever name got there first" — so this is a {@code Set<String>} guard instead, the
+     * same shape {@link OpenCodeLauncher#modelCheckSkippedWarned} already uses for its own
+     * once-per-distinct-thing WARN. {@link #add}'s return value (true only the first time a name is
+     * added) is what turns "log the whole gap" into "log only the names never warned about before".
+     *
+     * <p>Bounded by construction: every name added here first passed {@link
+     * #CREDENTIAL_SHAPED_NAME}'s filter over {@link #hostEnvNames}, i.e. it is an actual
+     * environment variable name from the daemon's own process — a small, OS-bounded set (the host
+     * environment has, in practice, tens to a few hundred entries), not an attacker- or
+     * request-controlled input. So this set cannot grow past "however many distinct credential-
+     * shaped names this host's environment has ever held across this launcher's lifetime," which is
+     * effectively fixed for the life of one daemon process — no separate cap is needed.
      *
      * <p>CB-633 follow-up (#192): kept SEPARATE from {@link #allowListGapLogged} on purpose.
      * {@code memberCredentials} is a live, re-read-per-spawn supplier, so the policy can change
      * between two spawns on the same launcher. A single shared flag would let a harmless allow-list
      * INFO on spawn 1 permanently suppress the real deny-by-default WARN a later spawn deserves —
      * the report that matters most getting hidden by the report that doesn't. Two flags mean each
-     * report kind fires exactly once, independent of what the other kind already logged.
+     * report kind (WARN vs. INFO) fires independently of what the other kind already logged; within
+     * the WARN kind itself, the set above further separates by name, for the same reason.
      */
-    private final AtomicBoolean unprotectedGapLogged = new AtomicBoolean();
+    private final Set<String> unprotectedGapNamesWarned = ConcurrentHashMap.newKeySet();
 
     /**
      * Guards the {@code effectiveAllowed != null} branch of {@link #logCredentialGap} — the
      * allow-list-scrub-covered report — to one INFO per launcher instance. See {@link
-     * #unprotectedGapLogged}'s javadoc for why this is a separate flag rather than a shared one.
+     * #unprotectedGapNamesWarned}'s javadoc for why this is a separate flag rather than a shared
+     * one; unlike that guard it stays a per-instance {@code AtomicBoolean}, not a per-name set —
+     * fleetd #341 fixed the WARN-vs-WARN suppression, not this INFO's own one-shot shape, which was
+     * not reported as broken and is out of that ticket's scope.
      */
     private final AtomicBoolean allowListGapLogged = new AtomicBoolean();
 
     /**
      * fleetd #185 stage 2: guards {@link #warnUnknownMemberEnvironment} to one WARN per launcher
-     * instance, not one per spawn — the same one-per-instance shape as {@link #unprotectedGapLogged}
-     * and {@link #allowListGapLogged}, kept as its own flag for the same reason those two are split:
-     * this mode is orthogonal to which of the other two branches would otherwise have fired.
+     * instance, not one per spawn — the same one-shot shape {@link #unprotectedGapNamesWarned} and
+     * {@link #allowListGapLogged} guard their own branches with, kept as its own flag for the same
+     * reason those two are split: this mode is orthogonal to which of the other two branches would
+     * otherwise have fired.
      */
     private final AtomicBoolean unknownMemberEnvironmentWarned = new AtomicBoolean();
 
@@ -1816,14 +1846,21 @@ public abstract class HerdrPeerLauncher implements PeerLauncher {
         List<String> blankedByScrub = gap.stream()
                 .filter(name -> !MemberEnvAllowList.keeps(effectiveAllowed, name))
                 .toList();
-        if (!keptByDerivedList.isEmpty() && unprotectedGapLogged.compareAndSet(false, true)) {
+        // fleetd #341: filter to names this guard has never warned about before — not just
+        // "isEmpty" on the whole branch — so a name this spawn's gap shares with an EARLIER
+        // spawn's (already-warned) gap does not re-print, while a name unique to THIS gap still
+        // does, whichever of the two WARN branches reported it first.
+        List<String> newlyUnprotected = keptByDerivedList.stream()
+                .filter(unprotectedGapNamesWarned::add)
+                .toList();
+        if (!newlyUnprotected.isEmpty()) {
             log.warn("memberCredentials gap: {} credential-shaped env var name(s) are on neither "
                             + "known: nor allow: — the derived allow-list keeps them anyway (a profile's "
                             + "gitTokenEnv/gitHostEnv/tokenEnv/env: names one, or this spawn injects it), "
                             + "so every member pane inherits them UNBLOCKED — {}. Add each to "
                             + "memberCredentials.known (or .allow if a member legitimately needs it), or "
                             + "remove it from whatever profile setting derives it in.",
-                    keptByDerivedList.size(), keptByDerivedList);
+                    newlyUnprotected.size(), newlyUnprotected);
         }
         if (!blankedByScrub.isEmpty() && allowListGapLogged.compareAndSet(false, true)) {
             log.info("memberCredentials gap: {} credential-shaped env var name(s) are on neither "
@@ -1836,12 +1873,18 @@ public abstract class HerdrPeerLauncher implements PeerLauncher {
 
     /** The deny-by-default (and allow-list non-zsh fallback) WARN — unchanged byte-for-byte by #192. */
     private void warnGapUnprotected(List<String> gap) {
-        if (unprotectedGapLogged.compareAndSet(false, true)) {
+        // fleetd #341: same "only the names never warned before" filter as the sibling branch in
+        // logCredentialGap above — see unprotectedGapNamesWarned's javadoc. Both branches share
+        // this one guard because both report the exact same fact (a name reaching a member pane
+        // unprotected) at the exact same severity; keying it by name is what lets a later spawn's
+        // DIFFERENT name still get its own WARN after an earlier spawn's already fired.
+        List<String> newlyUnprotected = gap.stream().filter(unprotectedGapNamesWarned::add).toList();
+        if (!newlyUnprotected.isEmpty()) {
             log.warn("memberCredentials gap: {} credential-shaped env var name(s) are on neither "
                             + "known: nor allow: — every member pane inherits them UNBLOCKED — {}. "
                             + "Add each to memberCredentials.known (blocked by default) or .allow "
                             + "(if a member legitimately needs it).",
-                    gap.size(), gap);
+                    newlyUnprotected.size(), newlyUnprotected);
         }
     }
 

fleetd/src/main/java/dev/ltms/fleet/msg/MessageService.java

@@ -863,16 +863,22 @@ public final class MessageService {
                 if (onAccepted != null) {
                     onAccepted.run();
                 }
-                CompletableFuture<Void> delivered = injector.enqueue(target, content, token);
+                Injector.Delivery delivery = injector.enqueue(target, content, token);
                 try {
                     Rendezvous.Resolution r = reply.get(remainingMillis(deadlineNanos), TimeUnit.MILLISECONDS);
                     return recorded(new Reply(outcomeOf(r.kind()), r.text(), r.turnId()));
                 } catch (TimeoutException e) {
-                    boolean wasDelivered = delivered.isDone() && !delivered.isCompletedExceptionally();
+                    boolean wasDelivered = delivery.completion().isDone()
+                            && !delivery.completion().isCompletedExceptionally();
+                    if (!wasDelivered) {
+                        // The target monitor makes cancellation atomic with onStatus picking this
+                        // Pending up. If pickup won, report TIMED_OUT_WORKING because the text landed.
+                        wasDelivered = injector.cancel(delivery) == Injector.Cancellation.DELIVERED;
+                    }
                     log.debug("send to {} timed out (delivered={})", target, wasDelivered);
                     if (!wasDelivered) {
-                        // CB-640: still sitting in the injector's queue, waiting for the member to
-                        // go idle — record the fact for fleet health (see queuedDeliveries).
+                        // CB-640: record that delivery did not happen for fleet health (see
+                        // queuedDeliveries). The exact Pending was cancelled, so it cannot arrive later.
                         queuedDeliveries.put(target, Boolean.TRUE);
                     }
                     return recorded(new Reply(

fleetd/src/test/java/dev/ltms/fleet/config/ConfigRefTopLevelCoverageTest.java

@@ -47,21 +47,21 @@ class ConfigRefTopLevelCoverageTest {
     private static final RecordComponent[] COMPONENTS = FleetConfig.class.getRecordComponents();
 
     /**
-     * Top-level components whose change {@code ConfigRef.changedDeferredKeys} reads and reports on
-     * — verified by reading that method as of fleetd #330, not derived from this test.
-     * {@code spawnReadyTimeoutMs}/{@code spawnReadyPollMs} are compared together and reported under
-     * one combined label ({@code "spawnReady*"}); {@code profiles} is compared twice over — once for
-     * added/removed profile names, once for an existing profile's launch settings — and that second
-     * comparison excludes {@code weight}/{@code maxLoad}/{@code credentialId} as hot sub-fields,
-     * which is what {@link ConfigRefProfileCoverageTest} exists to keep honest at the sub-field
-     * level. {@code profiles} itself still belongs here, not in the hot-exclusion set below: most of
-     * a profile's fields are NOT read live, so citing "read live off the config supplier" for the
-     * whole top-level key would be false.
+     * Top-level components whose change {@code ConfigRef.changedDeferredKeys} reads and reports on.
+     * Fleetd #337 promoted this out of a hand-maintained copy here into {@link ConfigRef#DEFERRED_KEYS}
+     * itself, the same reason {@link ConfigRef#COLD_KEYS} and {@link ConfigRef#SPLIT_KEYS} are
+     * production constants rather than test-side copies: two lists that are supposed to describe the
+     * same method are exactly the shape that silently drifts apart. {@code spawnReadyTimeoutMs}/
+     * {@code spawnReadyPollMs} are compared together and reported under one combined label
+     * ({@code "spawnReady*"}); {@code profiles} is compared twice over — once for added/removed
+     * profile names, once for an existing profile's launch settings — and that second comparison
+     * excludes {@code weight}/{@code maxLoad}/{@code credentialId} as hot sub-fields, which is what
+     * {@link ConfigRefProfileCoverageTest} exists to keep honest at the sub-field level. {@code
+     * profiles} itself still belongs here, not in the hot-exclusion set below: most of a profile's
+     * fields are NOT read live, so citing "read live off the config supplier" for the whole
+     * top-level key would be false.
      */
-    private static final Set<String> DEFERRED_TOP_LEVEL_KEYS = Set.of(
-            "guard", "worktreeRoot", "worktreeGroup", "primary", "configReload",
-            "leadHeartbeat", "lifecycle", "spawnReadyTimeoutMs", "spawnReadyPollMs",
-            "quarantineCooldownSeconds", "profiles");
+    private static final Set<String> DEFERRED_TOP_LEVEL_KEYS = ConfigRef.DEFERRED_KEYS;
 
     /**
      * The escape hatch: top-level components with no reload bookkeeping at all, because every read

fleetd/src/test/java/dev/ltms/fleet/config/ConfigRefTopLevelReportingCoverageTest.java

@@ -16,61 +16,68 @@ import static org.junit.jupiter.api.Assertions.assertEquals;
 
 /**
  * fleetd #333, finding F2: {@link ConfigRefTopLevelCoverageTest} proves every {@link FleetConfig}
- * top-level component sits in exactly one of {@link ConfigRef#COLD_KEYS}, {@code
- * DEFERRED_TOP_LEVEL_KEYS}, {@link ConfigRef#SPLIT_KEYS} or the hot-excluded set. It does
- * <strong>not</strong> prove that a key's membership in {@code COLD_KEYS} or {@code SPLIT_KEYS}
- * corresponds to any actual comparison in {@link ConfigRef}: a key can sit in either set with no
- * branch in {@code changedColdKeys}/{@code changedSplitKeys} checking it, and both
- * {@link ConfigRefTopLevelCoverageTest} and the "kept in step" {@code assert} inside each of those
- * methods stay green, because neither one reads the method body — the coverage test only reads
- * set membership, and the assert only checks that reported entries are a SUBSET of the set, never
- * that every set member produced a reported entry.
+ * top-level component sits in exactly one of {@link ConfigRef#COLD_KEYS}, {@link
+ * ConfigRef#DEFERRED_KEYS}, {@link ConfigRef#SPLIT_KEYS} or the hot-excluded set. It does
+ * <strong>not</strong> prove that a key's membership in one of the first three sets corresponds to
+ * any actual comparison in {@link ConfigRef}: a key can sit in a set with no branch in {@code
+ * changedColdKeys}/{@code changedSplitKeys}/{@code changedDeferredKeys} checking it, and both
+ * {@link ConfigRefTopLevelCoverageTest} and the "kept in step" {@code assert} inside the first two
+ * of those methods stay green, because neither one reads the method body — the coverage test only
+ * reads set membership, and the assert only checks that reported entries are a SUBSET of the set,
+ * never that every set member produced a reported entry. {@code changedDeferredKeys} does not even
+ * have a "kept in step" assert of its own.
  *
  * <p>Measured directly, live, while fixing fleetd #333: dropping the {@code coordinator} branch out
  * of {@code ConfigRef.changedSplitKeys} while leaving {@code "coordinator"} in
  * {@link ConfigRef#SPLIT_KEYS} left {@link ConfigRefTopLevelCoverageTest} and the in-method assert
  * both green — only a hand-written behavioural case in {@link ConfigRefTest} caught it, because it
  * happened to name that exact key. This is the {@link ConfigRefProfileCoverageTest} mechanism one
- * level up, generalised over every {@code COLD_KEYS}/{@code SPLIT_KEYS} member rather than one
- * hand-picked field: enumerate {@link FleetConfig}'s own record components by reflection, build "a
- * config where only {@code <key>} differs" for each cold/split key, and call the real
- * {@link ConfigRef#changedColdKeys}/{@link ConfigRef#changedSplitKeys} methods (made
- * package-private for exactly this, the same reason {@link ConfigRef#sameLaunchSettings} already
- * is) to prove each one is actually reported — not assumed from a set literal.
+ * level up, generalised over every {@code COLD_KEYS}/{@code SPLIT_KEYS}/{@code DEFERRED_KEYS}
+ * member rather than one hand-picked field: enumerate {@link FleetConfig}'s own record components by
+ * reflection, build "a config where only {@code <key>} differs" for each key, and call the real
+ * {@link ConfigRef#changedColdKeys}/{@link ConfigRef#changedSplitKeys}/
+ * {@link ConfigRef#changedDeferredKeys} methods (all package-private for exactly this, the same
+ * reason {@link ConfigRef#sameLaunchSettings} already is) to prove each one is actually reported —
+ * not assumed from a set literal.
  *
- * <h2>What this deliberately does NOT cover</h2>
- * {@code DEFERRED_TOP_LEVEL_KEYS} is not exercised here. That bucket carries the identical
- * one-way risk in principle — a key added to it with no matching branch in
- * {@code changedDeferredKeys} would pass {@link ConfigRefTopLevelCoverageTest} exactly the way
- * {@code coordinator} passed it above — but this test stays narrow to {@code COLD_KEYS} and
- * {@code SPLIT_KEYS} for two reasons. First, that is where fleetd #333 actually found and measured
- * the gap (F1 was a live instance of it). Second, most of {@code DEFERRED_TOP_LEVEL_KEYS} already
- * carries an individual behavioural test in {@link ConfigRefTest} naming it by key — {@code
- * worktreeGroup}, {@code primary}, {@code configReload}, {@code profiles}' launch settings /
- * weight-maxLoad / exhaustedPattern / errorPattern / ideProjectDir — which is the same protection
- * this class gives {@code COLD_KEYS}/{@code SPLIT_KEYS}, just written by hand per key instead of
- * generated by reflection over the whole set. {@code guard}, {@code lifecycle},
- * {@code leadHeartbeat}, {@code spawnReadyTimeoutMs}/{@code spawnReadyPollMs} and
- * {@code quarantineCooldownSeconds} do NOT have a dedicated behavioural test naming them, so the
- * one-way gap fleetd's own memory notes ("pre-existing on COLD_KEYS and on the test's own
- * DEFERRED_TOP_LEVEL_KEYS") is real and not fully closed by this class — extending this mechanism's
- * {@code BASE}/{@code ALT} map to cover every top-level component and adding a third
- * {@code everyDeferredKeyIsActuallyReportedByChangedDeferredKeys} test is the natural next step, left
- * for whoever next finds a deferred key with the same shape as this ticket's {@code fleet.leaders}.
+ * <h2>fleetd #337 — DEFERRED_KEYS was the gap left open here</h2>
+ * This class originally covered {@code COLD_KEYS} and {@code SPLIT_KEYS} only — {@code
+ * DEFERRED_TOP_LEVEL_KEYS} (now {@link ConfigRef#DEFERRED_KEYS}) carried the identical one-way risk
+ * in principle, unexercised. fleetd #337 measured the real consequence rather than assuming it from
+ * the shape of the gap: dropping {@code guard}'s comparison out of {@code changedDeferredKeys} while
+ * {@code "guard"} stayed in the set left all 1355 tests green — the same failure mode {@code
+ * coordinator} demonstrated for {@code SPLIT_KEYS} in fleetd #333, now confirmed for {@code
+ * DEFERRED_KEYS} too. Re-deriving the full list by mutation (drop each key's branch in turn, run the
+ * suite, restore) found six of the eleven {@code DEFERRED_KEYS} members with no behavioural test in
+ * {@link ConfigRefTest} naming them: {@code guard}, {@code leadHeartbeat}, {@code worktreeRoot},
+ * {@code spawnReadyTimeoutMs}, {@code spawnReadyPollMs} and {@code quarantineCooldownSeconds}. That
+ * list corrects fleetd #333's own guess at it in two ways the mutation proved and a reading did not:
+ * {@code lifecycle} is NOT on it — {@code ConfigRefTest.aDeferredChangeIsAppliedAndReported} already
+ * names it, and dropping its branch fails that test — and {@code worktreeRoot} IS on it, which #333
+ * never named at all. The other five {@code DEFERRED_KEYS} members ({@code lifecycle}, {@code
+ * worktreeGroup}, {@code primary}, {@code configReload}, {@code profiles}) already had a hand-written
+ * case each. {@link #everyDeferredKeyIsActuallyReportedByChangedDeferredKeys} below now covers all
+ * eleven the reflective way, so the six with no hand-written test are no longer silently unpinned —
+ * every {@code DEFERRED_KEYS} component turned out to be a scalar or a simple record, so, unlike
+ * {@code fleet.leaders} in fleetd #333, none needed an exclusion set: {@link #BASE}/{@link #ALT} give
+ * every top-level component (not only {@code COLD_KEYS}/{@code SPLIT_KEYS}) a real, distinct value.
  */
 class ConfigRefTopLevelReportingCoverageTest {
 
     private static final RecordComponent[] COMPONENTS = FleetConfig.class.getRecordComponents();
 
     /**
-     * One valid value per top-level {@link FleetConfig} component — "the a value". Components not
-     * exercised by either test below ({@code profiles}, {@code guard}, {@code worktreeRoot}, …) are
-     * left {@code null}/empty; {@link FleetConfig}'s compact constructor only normalizes
-     * {@code profiles}, so every other field accepts {@code null} unmutated.
+     * One valid value per top-level {@link FleetConfig} component — "the a value". fleetd #337 gave
+     * every {@code DEFERRED_KEYS} component a real value here too (previously left {@code null} on
+     * both sides, which meant {@code mutate(key)} produced no actual difference for any of them);
+     * only {@code placement}, {@code memberCredentials} and {@code memberLoginShell} — the
+     * hot-excluded set, never compared by any {@code changed*Keys} method — stay {@code null}.
+     * {@link FleetConfig}'s compact constructor only normalizes {@code profiles}, so every other
+     * field accepts whatever is put here unmutated.
      */
     private static final Map<String, Object> BASE = baseValues();
 
-    /** The same shape, each value distinct from {@link #BASE} — "the b value" — for COLD_KEYS/SPLIT_KEYS only. */
+    /** The same shape, each value distinct from {@link #BASE} — "the b value". */
     private static final Map<String, Object> ALT = altValues();
 
     private static Map<String, Object> baseValues() {
@@ -79,26 +86,26 @@ class ConfigRefTopLevelReportingCoverageTest {
         v.put("herdrSocket", "~/.config/herdr/a.sock");
         v.put("memberHerdrSocket", "~/.config/herdr/member-a.sock");
         v.put("profiles", Map.of());
-        v.put("guard", null);
-        v.put("worktreeRoot", null);
-        v.put("lifecycle", null);
-        v.put("spawnReadyTimeoutMs", null);
-        v.put("spawnReadyPollMs", null);
+        v.put("guard", new FleetConfig.Guard(List.of("host-a")));
+        v.put("worktreeRoot", "/wt/a");
+        v.put("lifecycle", new FleetConfig.Lifecycle(300, 5, 30, false));
+        v.put("spawnReadyTimeoutMs", 5000);
+        v.put("spawnReadyPollMs", 100);
         v.put("broker", new FleetConfig.Broker("amqp://a", null, 1));
-        v.put("primary", null);
+        v.put("primary", new FleetConfig.Primary("term-a", 1, 1000));
         v.put("fleet", new FleetConfig.Fleet(
                 Map.of("opus", new FleetConfig.Leader("sonnet", "lead: opus-a", 1, null, 10,
                         "claude", null, null, null)),
                 Map.of(), Map.of(), Map.of(), Map.of(), "{role}: {profile} #{n}"));
-        v.put("leadHeartbeat", null);
+        v.put("leadHeartbeat", new FleetConfig.LeadHeartbeat(300, 60_000L, 3));
         v.put("health", new FleetConfig.Health(true, 30, 600, null, null));
         v.put("placement", null);
         v.put("auth", new FleetConfig.Auth("loopback-trust", null));
-        v.put("configReload", null);
-        v.put("quarantineCooldownSeconds", null);
+        v.put("configReload", new FleetConfig.ConfigReload(true, 10));
+        v.put("quarantineCooldownSeconds", 1800);
         v.put("memberCredentials", null);
         v.put("coordinator", new FleetConfig.Coordinator("amqp://coord-a", null, "self-a", 1));
-        v.put("worktreeGroup", null);
+        v.put("worktreeGroup", "group-a");
         v.put("memberLoginShell", null);
         assertNamesMatchComponents(v);
         return v;
@@ -109,14 +116,18 @@ class ConfigRefTopLevelReportingCoverageTest {
         v.put("bind", new FleetConfig.Bind("127.0.0.2", 8766));
         v.put("herdrSocket", "~/.config/herdr/b.sock");
         v.put("memberHerdrSocket", "~/.config/herdr/member-b.sock");
-        v.put("profiles", Map.of());
-        v.put("guard", null);
-        v.put("worktreeRoot", null);
-        v.put("lifecycle", null);
-        v.put("spawnReadyTimeoutMs", null);
-        v.put("spawnReadyPollMs", null);
+        // A single added profile — enough to trip the "added/removed" comparison in
+        // ConfigRef.changedDeferredKeys, which is all this mechanism needs to prove "profiles" has
+        // a branch behind it; the launch-settings comparison already has its own hand-written cases
+        // in ConfigRefTest (changingAProfilesLaunchSettingsIsReportedAsDeferred and siblings).
+        v.put("profiles", Map.of("sonnet", minimalProfile("sonnet")));
+        v.put("guard", new FleetConfig.Guard(List.of("host-b")));
+        v.put("worktreeRoot", "/wt/b");
+        v.put("lifecycle", new FleetConfig.Lifecycle(600, 10, 60, true));
+        v.put("spawnReadyTimeoutMs", 10_000);
+        v.put("spawnReadyPollMs", 200);
         v.put("broker", new FleetConfig.Broker("amqp://b", null, 2));
-        v.put("primary", null);
+        v.put("primary", new FleetConfig.Primary("term-b", 2, 2000));
         // Differs from BASE.fleet only in fleet.leaders (a different tab for "opus") — the frozen
         // sub-field ConfigRef.changedSplitKeys actually compares. A Fleet that instead differed only
         // in tabLabel would correctly NOT be reported (see
@@ -126,20 +137,27 @@ class ConfigRefTopLevelReportingCoverageTest {
                 Map.of("opus", new FleetConfig.Leader("sonnet", "lead: opus-b", 1, null, 10,
                         "claude", null, null, null)),
                 Map.of(), Map.of(), Map.of(), Map.of(), "{role}: {profile} #{n}"));
-        v.put("leadHeartbeat", null);
+        v.put("leadHeartbeat", new FleetConfig.LeadHeartbeat(600, 120_000L, 5));
         v.put("health", new FleetConfig.Health(false, 90, 900, null, null));
         v.put("placement", null);
         v.put("auth", new FleetConfig.Auth("token", "TOKEN_ENV"));
-        v.put("configReload", null);
-        v.put("quarantineCooldownSeconds", null);
+        v.put("configReload", new FleetConfig.ConfigReload(false, 20));
+        v.put("quarantineCooldownSeconds", 3600);
         v.put("memberCredentials", null);
         v.put("coordinator", new FleetConfig.Coordinator("amqp://coord-b", null, "self-b", 2));
-        v.put("worktreeGroup", null);
+        v.put("worktreeGroup", "group-b");
         v.put("memberLoginShell", null);
         assertNamesMatchComponents(v);
         return v;
     }
 
+    /** A minimal, otherwise-null {@link FleetConfig.Profile} — just enough to name one in a map. */
+    private static FleetConfig.Profile minimalProfile(String name) {
+        return new FleetConfig.Profile(name, null, null, null, null, null, null, null, null, null,
+                null, null, null, null, null, null, null, null, null, null, null, null, null, null,
+                null, null);
+    }
+
     private static void assertNamesMatchComponents(Map<String, Object> values) {
         Set<String> names = new TreeSet<>();
         for (RecordComponent rc : COMPONENTS) {
@@ -216,4 +234,51 @@ class ConfigRefTopLevelReportingCoverageTest {
                         + "entry from changedSplitKeys — a set entry with no comparison behind it, "
                         + "exactly the fleetd #333 F2 shape: " + uncovered);
     }
+
+    /**
+     * For most {@link ConfigRef#DEFERRED_KEYS} members, {@code changedDeferredKeys} reports the key
+     * name verbatim — the default this map assumes. Two entries don't: {@code spawnReadyTimeoutMs}
+     * and {@code spawnReadyPollMs} are compared together in one branch and reported under the
+     * combined label {@code "spawnReady*"} (see {@link ConfigRef#changedDeferredKeys}). {@code
+     * profiles} keeps the default: mutating it here only exercises the added/removed comparison
+     * (see {@link #altValues}), which reports {@code "profiles (added/removed: …)"} — starts with
+     * {@code "profiles"}, same as the default would expect.
+     */
+    private static final Map<String, String> DEFERRED_REPORT_PREFIX = Map.of(
+            "spawnReadyTimeoutMs", "spawnReady*",
+            "spawnReadyPollMs", "spawnReady*");
+
+    /**
+     * fleetd #337: the same mechanism applied to {@link ConfigRef#DEFERRED_KEYS}, closing the gap
+     * this class's own javadoc left open since fleetd #333. Mutate each deferred key in isolation
+     * and prove {@code changedDeferredKeys} actually names it (message starting with the key's
+     * expected report prefix — see {@link #DEFERRED_REPORT_PREFIX}), not just that {@code
+     * DEFERRED_KEYS} claims it does. This is the exact check that fails for {@code guard} the way
+     * {@code coordinator} failed {@link #everySplitKeyIsActuallyReportedByChangedSplitKeys} in
+     * fleetd #333 — verified live: dropping {@code guard}'s branch from {@code changedDeferredKeys}
+     * while {@code "guard"} stayed in {@code DEFERRED_KEYS} left the whole 1355-test suite green,
+     * and this test is what now catches it (it fails naming {@code guard} with that mutation in
+     * place).
+     */
+    @Test
+    void everyDeferredKeyIsActuallyReportedByChangedDeferredKeys() throws ReflectiveOperationException {
+        FleetConfig base = configOf(BASE);
+        List<String> uncovered = new ArrayList<>();
+        for (String key : new TreeSet<>(ConfigRef.DEFERRED_KEYS)) {
+            FleetConfig mutated = mutate(key);
+            List<String> deferred = ConfigRef.changedDeferredKeys(base, mutated);
+            String prefix = DEFERRED_REPORT_PREFIX.getOrDefault(key, key);
+            if (deferred.stream().noneMatch(s -> s.startsWith(prefix))) {
+                uncovered.add(key);
+            }
+        }
+        System.out.printf(
+                "ConfigRef.changedDeferredKeys reporting coverage — %d DEFERRED_KEYS, %d verified%n",
+                ConfigRef.DEFERRED_KEYS.size(), ConfigRef.DEFERRED_KEYS.size() - uncovered.size());
+        assertEquals(List.of(), uncovered,
+                "these keys are in ConfigRef.DEFERRED_KEYS but mutating them alone produces no "
+                        + "matching entry from changedDeferredKeys — a set entry with no comparison "
+                        + "behind it, exactly the fleetd #333 F2 shape, confirmed here for "
+                        + "DEFERRED_KEYS by fleetd #337: " + uncovered);
+    }
 }

fleetd/src/test/java/dev/ltms/fleet/inject/CompletionResolverTest.java

@@ -970,6 +970,40 @@ class CompletionResolverTest {
         assertTrue(notified.get(0).contains("503 Service Unavailable"), notified.get(0));
     }
 
+    /**
+     * fleetd #339 follow-up: a genuine backend error rendered behind terminal chrome must still
+     * record the credential outage. #339 added a start-of-line check to stop a member's own prose
+     * being counted as an outage, and a bare {@code lookingAt} also rejected this — the send failed
+     * but the sink never fired. #339's invariant 3 named that direction as the worse one: a real
+     * outage going unrecorded leaves the fleet spawning into a dead credential.
+     *
+     * <p>The raw-scrape path is where this matters, because its own comment says to expect leading
+     * TUI chrome there.
+     */
+    @Test
+    void aRealErrorBehindTerminalChromeStillNotifiesTheSink() {
+        String block = """
+                ╭──────────────────────────────────────╮
+                │ 503 Service Unavailable: upstream credential rejected
+                """;
+        FakeHerdr herdr = new FakeHerdr().readText(block);
+        Rendezvous rendezvous = new Rendezvous();
+        BackendErrorPatternLookup patterns = target -> Pattern.compile("(?i)503 Service Unavailable");
+        java.util.List<String> notified = new java.util.ArrayList<>();
+        BackendErrorSink sink = (target, matchedLine, reason) -> notified.add(target + ": " + matchedLine);
+        CompletionResolver resolver = new CompletionResolver(new AgentControl(herdr), rendezvous,
+                ExhaustedPatternLookup.none(), ExhaustionSink.none(), patterns, sink);
+
+        var waiter = rendezvous.open("term_a");
+        resolver.resolve("term_a", new CompletionResolver.InFlight(waiter, null));
+
+        assertEquals(Rendezvous.Kind.FAILED, waiter.getNow(null).kind(),
+                "a real backend error must still fail the send");
+        assertEquals(1, notified.size(),
+                "a real error line behind box chrome is still a real outage — it must reach the sink, "
+                        + "or the fleet keeps spawning into a dead credential");
+    }
+
     // --- fleetd#201 Unit 1: classification inside the fleetd#164 MIN_TURN_NANOS floor -------------
 
     @Test

fleetd/src/test/java/dev/ltms/fleet/inject/InjectorTest.java

@@ -48,7 +48,7 @@ class InjectorTest {
 
     @Test
     void deliversWhenIdle() {
-        CompletableFuture<Void> f = injector.enqueue(T, "hello", TestTurnTokens.inert(T));
+        CompletableFuture<Void> f = injector.enqueue(T, "hello", TestTurnTokens.inert(T)).completion();
         assertFalse(f.isDone(), "not delivered until an injectable status arrives");
         injector.onStatus(T, AgentStatus.IDLE);
         assertTrue(f.isDone());
@@ -170,6 +170,32 @@ class InjectorTest {
         assertEquals(List.of("a", "b", "c"), sent());
     }
 
+    @Test
+    void cancellingTheMiddleDeliveryKeepsTheFollowingDeliveryReachable() {
+        Injector.Delivery first = injector.enqueue(T, "same text", TestTurnTokens.inert(T));
+        Injector.Delivery cancelled = injector.enqueue(T, "same text", TestTurnTokens.inert(T));
+        injector.enqueue(T, "after cancelled", TestTurnTokens.inert(T));
+
+        assertEquals(Injector.Cancellation.CANCELLED, injector.cancel(cancelled));
+        injector.onStatus(T, AgentStatus.IDLE);
+        injector.onStatus(T, AgentStatus.WORKING);
+        injector.onStatus(T, AgentStatus.IDLE);
+
+        assertEquals(List.of("same text", "after cancelled"), sent(),
+                "cancellation must match the exact Delivery and preserve the remaining FIFO queue");
+        assertTrue(first.completion().isDone());
+    }
+
+    @Test
+    void cancellationReportsDeliveredWhenPickupWonTheRace() {
+        Injector.Delivery delivery = injector.enqueue(T, "already sent", TestTurnTokens.inert(T));
+        injector.onStatus(T, AgentStatus.IDLE);
+
+        assertEquals(Injector.Cancellation.DELIVERED, injector.cancel(delivery),
+                "a cancellation after pickup must not claim that the text stayed queued");
+        assertEquals(List.of("already sent"), sent());
+    }
+
     @Test
     void activeWhileQueuedOrInFlightThenQuietAfterTurnCompletes() {
         assertTrue(injector.activeTargets().isEmpty());
@@ -378,7 +404,7 @@ class InjectorTest {
     void sendFailureDropsMessageAndFailsItsFuture() {
         FakeHerdr failing = new FakeHerdr().agentSendFailsWith("send_failed");
         Injector inj = new Injector(new AgentControl(failing));
-        CompletableFuture<Void> f = inj.enqueue(T, "boom", TestTurnTokens.inert(T));
+        CompletableFuture<Void> f = inj.enqueue(T, "boom", TestTurnTokens.inert(T)).completion();
 
         inj.onStatus(T, AgentStatus.IDLE);
         assertTrue(f.isCompletedExceptionally());
@@ -387,7 +413,7 @@ class InjectorTest {
 
     @Test
     void dropFailsPendingWaiters() {
-        CompletableFuture<Void> f = injector.enqueue(T, "orphan", TestTurnTokens.inert(T));
+        CompletableFuture<Void> f = injector.enqueue(T, "orphan", TestTurnTokens.inert(T)).completion();
         injector.drop(T, new HerdrException("worker gone", "pane_not_found", null));
         assertTrue(f.isCompletedExceptionally(), "queued waiters unblock when the worker vanishes");
     }
@@ -396,8 +422,8 @@ class InjectorTest {
     void dropPassesTheRealCauseForQueuedAndDeliveredWork() {
         Captor cap = new Captor();
         Injector inj = new Injector(new AgentControl(herdr), cap);
-        CompletableFuture<Void> delivered = inj.enqueue(T, "delivered", TestTurnTokens.inert(T));
-        CompletableFuture<Void> queued = inj.enqueue(T, "queued", TestTurnTokens.inert(T));
+        CompletableFuture<Void> delivered = inj.enqueue(T, "delivered", TestTurnTokens.inert(T)).completion();
+        CompletableFuture<Void> queued = inj.enqueue(T, "queued", TestTurnTokens.inert(T)).completion();
 
         inj.onStatus(T, AgentStatus.IDLE);    // deliver the first message
         inj.onStatus(T, AgentStatus.WORKING); // its turn is now in flight; one remains queued
@@ -449,7 +475,7 @@ class InjectorTest {
         Captor cap = new Captor();
         List<String> forgotten = new ArrayList<>();
         Injector inj = new Injector(new AgentControl(herdr), cap, _ -> false, forgotten::add);
-        CompletableFuture<Void> f = inj.enqueue(T, "task", TestTurnTokens.inert(T));
+        CompletableFuture<Void> f = inj.enqueue(T, "task", TestTurnTokens.inert(T)).completion();
 
         for (int i = 0; i < READINESS_SAMPLES; i++) inj.onStatus(T, AgentStatus.IDLE);
 
@@ -563,7 +589,7 @@ class InjectorTest {
         StatusPoller poller = new StatusPoller(new AgentControl(idle), inj, 10);
         poller.start();
         try {
-            CompletableFuture<Void> delivered = inj.enqueue(T, "via-poller", TestTurnTokens.inert(T));
+            CompletableFuture<Void> delivered = inj.enqueue(T, "via-poller", TestTurnTokens.inert(T)).completion();
             delivered.get(2, TimeUnit.SECONDS); // completes when the poller drives the send
         } finally {
             poller.stop();
@@ -582,7 +608,7 @@ class InjectorTest {
     void deliveredFutureCarriesSendFailure() {
         FakeHerdr failing = new FakeHerdr().agentSendFailsWith("send_failed");
         Injector inj = new Injector(new AgentControl(failing));
-        CompletableFuture<Void> f = inj.enqueue(T, "boom", TestTurnTokens.inert(T));
+        CompletableFuture<Void> f = inj.enqueue(T, "boom", TestTurnTokens.inert(T)).completion();
         inj.onStatus(T, AgentStatus.IDLE);
         ExecutionException ex = assertThrows(ExecutionException.class, f::get);
         assertInstanceOf(HerdrException.class, ex.getCause());

fleetd/src/test/java/dev/ltms/fleet/inject/StatusPollerRoutingTest.java

@@ -41,7 +41,7 @@ class StatusPollerRoutingTest {
         poller.start();
         try {
             CompletableFuture<Void> delivered =
-                    injector.enqueue(LEAD_TARGET, "via-poller", TestTurnTokens.inert(LEAD_TARGET));
+                    injector.enqueue(LEAD_TARGET, "via-poller", TestTurnTokens.inert(LEAD_TARGET)).completion();
             // Must resolve quickly: refining against the WRONG daemon (member) never classifies
             // out of UNKNOWN, so this would time out under the bug.
             delivered.get(2, TimeUnit.SECONDS);
@@ -65,7 +65,7 @@ class StatusPollerRoutingTest {
         poller.start();
         try {
             CompletableFuture<Void> delivered =
-                    injector.enqueue(LEAD_TARGET, "via-poller", TestTurnTokens.inert(LEAD_TARGET));
+                    injector.enqueue(LEAD_TARGET, "via-poller", TestTurnTokens.inert(LEAD_TARGET)).completion();
             assertThrows(TimeoutException.class, () -> delivered.get(500, TimeUnit.MILLISECONDS),
                     "a lead target must never be refined from the member daemon's pane content");
         } finally {

fleetd/src/test/java/dev/ltms/fleet/member/CompositePeerLauncherTest.java

@@ -299,6 +299,41 @@ class CompositePeerLauncherTest {
                 "two adapter kinds sharing one daemon keep the fallback route");
     }
 
+    @Test
+    void stopThroughTheSingleDaemonShortcutStillClosesTheTabWhenTheFallbackDelegateUsesPanePlacement() {
+        // fleetd #342: the single-daemon shortcut (spawnedBy empty, herdrDaemonCount()==1) always
+        // routes stop() through delegates.getFirst() — here the claude adapter, configured for
+        // PANE placement (its own profiles never create a dedicated tab). The pane being torn
+        // down here actually belongs to the opencode adapter's TAB placement, sharing the same
+        // herdr daemon — mixing placements is the point: every existing stop-fallback test in this
+        // class configures BOTH adapters as "tab", so usesTabPlacement() was true either way and
+        // the mis-routing never showed.
+        //
+        // Before the fix, HerdrPeerLauncher#stop gated tab resolution on usesTabPlacement() of the
+        // delegate it happened to be called through, so the wrongly-routed (pane-placement) claude
+        // adapter never even looked for a tab to close, and the now-empty tab leaked with nothing
+        // to reap it. The fix (fleetd #342) resolves the pane's real tab unconditionally, so the
+        // decision follows the pane's actual placement rather than the fallback delegate's static
+        // config.
+        FakeHerdr herdr = new FakeHerdr();
+        FleetConfig.Profile claudePane = new FleetConfig.Profile("claude", "http://gx00.gw:8000",
+                "coder", null, "FLEETD_WORKER_TOKEN", List.of("claude"), "pane", "fleetd-workers",
+                "w #{n}", null, null, null);
+        ClaudeCodeLauncher claude = new ClaudeCodeLauncher(new AgentControl(herdr), new WorkspaceControl(herdr),
+                new SubscriptionGuard(Set.of("gx00.gw")), Map.of("claude", claudePane), "claude", _ -> null);
+        PeerLauncher composite = new CompositePeerLauncher(List.of(claude, opencodeAdapter(herdr)), "claude");
+
+        // "w9:pW" was never spawned through this composite instance, so spawnedBy has no entry for
+        // it (the same in-memory-cache-miss shape a daemon restart leaves behind) — stop() falls
+        // through to the single-daemon shortcut and hands it to delegates.getFirst() (claude).
+        composite.stop("w9:pW");
+
+        assertTrue(herdr.called("pane.close"), "the pane itself is still closed on every routed path");
+        assertTrue(herdr.called("tab.close"),
+                "the pane's real (sole-occupant) tab must be closed even though the fallback routed "
+                        + "through a delegate configured for pane placement");
+    }
+
     @Test
     void listKeepsBothPanesWhenTwoDaemonsShareAPaneId() {
         // herdr pane ids are per-daemon counters, so two daemons really can both hold w1:p1 on

fleetd/src/test/java/dev/ltms/fleet/member/HerdrPeerLauncherAllowListWiringTest.java

@@ -23,6 +23,7 @@ import java.util.List;
 import java.util.Map;
 import java.util.Set;
 import java.util.concurrent.atomic.AtomicBoolean;
+import java.util.concurrent.atomic.AtomicReference;
 import java.util.function.Function;
 import java.util.function.Supplier;
 
@@ -370,6 +371,161 @@ class HerdrPeerLauncherAllowListWiringTest {
                 "expected the pre-existing 'scrub blanks them' INFO unchanged, got: " + messages);
     }
 
+    /**
+     * fleetd #341: {@code unprotectedGapLogged} guarded TWO WARN branches that name DIFFERENT env
+     * var names — the allow-list branch ({@code keptByDerivedList}, below) and the deny-by-default
+     * / non-zsh-fallback branch ({@link HerdrPeerLauncher#warnGapUnprotected}). {@code
+     * memberCredentials} is a live, re-read-per-spawn supplier, so the policy can change between
+     * two spawns on the same launcher instance — a config reload needs no restart. Spawn 1 runs
+     * under {@code deny-by-default} with a gap of {@code SPAWN_ONE_UNCOVERED_TOKEN}, which trips
+     * the (before this fix) SHARED one-shot flag. The policy is then reloaded to {@code
+     * allow-list}; spawn 2's gap is {@code FLEETD_WORKER_TOKEN} instead — the test profile's own
+     * {@code tokenEnv}, which the derived allow-list keeps even though it is on neither {@code
+     * known:} nor {@code allow:}, so it is genuinely unprotected and deserves its own WARN. Before
+     * this fix that WARN never fires, because the shared flag was already {@code true} — the
+     * operator is never told {@code FLEETD_WORKER_TOKEN} reaches every member pane unblocked. Real
+     * path: two real {@link HerdrPeerLauncher#spawn} calls on ONE launcher instance, with mutable
+     * {@code memberCredentials}/host-env suppliers standing in for a live config reload between
+     * spawns.
+     */
+    @Test
+    void aDifferentUnprotectedGapOnALaterSpawnIsNotSuppressedByAnEarlierSpawnsWarn() {
+        FakeHerdr herdr = new FakeHerdr();
+        AtomicReference<FleetConfig.MemberCredentials> credsState = new AtomicReference<>(
+                new FleetConfig.MemberCredentials(null, List.of(), List.of(), null)); // deny-by-default
+        AtomicReference<Set<String>> hostEnvState =
+                new AtomicReference<>(Set.of("SPAWN_ONE_UNCOVERED_TOKEN"));
+        WiringLauncher launcher = new WiringLauncher(herdr, credsState::get, "/bin/zsh", hostEnvState::get);
+
+        Logger logger = (Logger) LoggerFactory.getLogger(HerdrPeerLauncher.class);
+        Level original = logger.getLevel();
+        logger.setLevel(Level.WARN);
+        ListAppender<ILoggingEvent> appender = new ListAppender<>();
+        appender.start();
+        logger.addAppender(appender);
+        try {
+            // Spawn 1: deny-by-default, gap = {SPAWN_ONE_UNCOVERED_TOKEN} — the effectiveAllowed ==
+            // null branch, via warnGapUnprotected.
+            launcher.spawn(new SpawnRequest("test", null, null, null, null, MemberRole.DEV));
+
+            // Live policy reload to allow-list, with a DIFFERENT gap name.
+            credsState.set(new FleetConfig.MemberCredentials(
+                    FleetConfig.MemberCredentials.POLICY_ALLOW_LIST, List.of(), List.of(), null));
+            hostEnvState.set(Set.of("FLEETD_WORKER_TOKEN"));
+            // Spawn 2: allow-list, gap = {FLEETD_WORKER_TOKEN} — kept by the derived allow-list
+            // (the profile's own tokenEnv), so it is the effectiveAllowed != null / keptByDerivedList
+            // branch, at the SAME log line HerdrPeerLauncher:1819 guards with the shared flag.
+            launcher.spawn(new SpawnRequest("test", null, null, null, null, MemberRole.DEV));
+        } finally {
+            logger.detachAppender(appender);
+            logger.setLevel(original);
+        }
+
+        List<String> messages = appender.list.stream().map(ILoggingEvent::getFormattedMessage).toList();
+        assertTrue(messages.stream().anyMatch(
+                        m -> m.contains("UNBLOCKED") && m.contains("SPAWN_ONE_UNCOVERED_TOKEN")),
+                "spawn 1's deny-by-default gap must still warn — got: " + messages);
+        assertTrue(messages.stream().anyMatch(
+                        m -> m.contains("UNBLOCKED") && m.contains("FLEETD_WORKER_TOKEN")),
+                "spawn 2's gap names a DIFFERENT env var than spawn 1 (FLEETD_WORKER_TOKEN, not "
+                        + "SPAWN_ONE_UNCOVERED_TOKEN) — it must still be warned about even though a "
+                        + "flag already fired once for spawn 1's unrelated name. Before fleetd #341's "
+                        + "fix this WARN never fires because unprotectedGapLogged was already true. "
+                        + "Got: " + messages);
+    }
+
+    /**
+     * fleetd #341 follow-up: the OTHER half of the guard's contract. The set exists to report every
+     * distinct name, but it must still report each one only ONCE — the noise control is the reason
+     * a guard is here at all, and the ticket named it as invariant 1. Two spawns, same policy, same
+     * gap name: exactly one WARN mentioning it.
+     *
+     * <p>Measured before this test existed: replacing {@code .filter(unprotectedGapNamesWarned::add)}
+     * with a filter that adds and always returns {@code true} — so every name is logged on every
+     * spawn — left all 1358 tests green. The fix was correct and nothing held it there. That is the
+     * "a test on the seam does not prove the caller" shape: the {@code Set} behaves, and nothing
+     * proved this class used it as a guard rather than as a record.
+     */
+    @Test
+    void theSameUnprotectedNameIsWarnedAboutOnlyOnceAcrossSpawns() {
+        FakeHerdr herdr = new FakeHerdr();
+        AtomicReference<FleetConfig.MemberCredentials> credsState = new AtomicReference<>(
+                new FleetConfig.MemberCredentials(null, List.of(), List.of(), null)); // deny-by-default
+        AtomicReference<Set<String>> hostEnvState =
+                new AtomicReference<>(Set.of("REPEATED_UNCOVERED_TOKEN"));
+        WiringLauncher launcher = new WiringLauncher(herdr, credsState::get, "/bin/zsh", hostEnvState::get);
+
+        Logger logger = (Logger) LoggerFactory.getLogger(HerdrPeerLauncher.class);
+        Level original = logger.getLevel();
+        logger.setLevel(Level.WARN);
+        ListAppender<ILoggingEvent> appender = new ListAppender<>();
+        appender.start();
+        logger.addAppender(appender);
+        try {
+            launcher.spawn(new SpawnRequest("test", null, null, null, null, MemberRole.DEV));
+            // Same policy, same gap, second spawn. Nothing new to tell the operator.
+            launcher.spawn(new SpawnRequest("test", null, null, null, null, MemberRole.DEV));
+        } finally {
+            logger.detachAppender(appender);
+            logger.setLevel(original);
+        }
+
+        List<String> messages = appender.list.stream().map(ILoggingEvent::getFormattedMessage).toList();
+        long mentioning = messages.stream()
+                .filter(m -> m.contains("REPEATED_UNCOVERED_TOKEN"))
+                .count();
+        assertEquals(1, mentioning,
+                "one unchanged unprotected name across two spawns must produce exactly one WARN — "
+                        + "the set is a guard, not just a record. Got: " + messages);
+    }
+
+    /**
+     * fleetd #341 follow-up: the reverse policy order. The original defect was found going
+     * deny-by-default then allow-list, and a guard that is fixed in one direction is not
+     * necessarily fixed in the other — "ask which states still OPEN the gate". Here spawn 1 runs
+     * under {@code allow-list} (the {@code keptByDerivedList} branch) and spawn 2 under
+     * {@code deny-by-default} ({@link HerdrPeerLauncher#warnGapUnprotected}), with a different name
+     * each time. Both must be reported.
+     */
+    @Test
+    void anAllowListWarnDoesNotSuppressALaterDenyByDefaultWarnForADifferentName() {
+        FakeHerdr herdr = new FakeHerdr();
+        AtomicReference<FleetConfig.MemberCredentials> credsState = new AtomicReference<>(
+                new FleetConfig.MemberCredentials(
+                        FleetConfig.MemberCredentials.POLICY_ALLOW_LIST, List.of(), List.of(), null));
+        AtomicReference<Set<String>> hostEnvState =
+                new AtomicReference<>(Set.of("FLEETD_WORKER_TOKEN"));
+        WiringLauncher launcher = new WiringLauncher(herdr, credsState::get, "/bin/zsh", hostEnvState::get);
+
+        Logger logger = (Logger) LoggerFactory.getLogger(HerdrPeerLauncher.class);
+        Level original = logger.getLevel();
+        logger.setLevel(Level.WARN);
+        ListAppender<ILoggingEvent> appender = new ListAppender<>();
+        appender.start();
+        logger.addAppender(appender);
+        try {
+            // Spawn 1: allow-list, gap kept by the derived list — the keptByDerivedList WARN.
+            launcher.spawn(new SpawnRequest("test", null, null, null, null, MemberRole.DEV));
+
+            // Live reload the OTHER way: back to deny-by-default, with a different name.
+            credsState.set(new FleetConfig.MemberCredentials(null, List.of(), List.of(), null));
+            hostEnvState.set(Set.of("LATER_UNCOVERED_TOKEN"));
+            launcher.spawn(new SpawnRequest("test", null, null, null, null, MemberRole.DEV));
+        } finally {
+            logger.detachAppender(appender);
+            logger.setLevel(original);
+        }
+
+        List<String> messages = appender.list.stream().map(ILoggingEvent::getFormattedMessage).toList();
+        assertTrue(messages.stream().anyMatch(
+                        m -> m.contains("UNBLOCKED") && m.contains("FLEETD_WORKER_TOKEN")),
+                "spawn 1's allow-list gap must warn — got: " + messages);
+        assertTrue(messages.stream().anyMatch(
+                        m -> m.contains("UNBLOCKED") && m.contains("LATER_UNCOVERED_TOKEN")),
+                "spawn 2's deny-by-default gap names a different variable and must still be warned "
+                        + "about, even though an allow-list WARN already fired. Got: " + messages);
+    }
+
     /**
      * fleetd #185 stage 2: with {@code memberHerdrSocket:} configured, member panes run under a
      * different OS user — {@link HerdrPeerLauncher#hostEnvNames} describes fleetd's own process, not

fleetd/src/test/java/dev/ltms/fleet/msg/MessageServiceTest.java

@@ -258,7 +258,7 @@ class MessageServiceTest {
         injector.onStatus(T, AgentStatus.IDLE);    // first delivery
         injector.onStatus(T, AgentStatus.WORKING); // first turn in flight
 
-        CompletableFuture<Void> queued = injector.enqueue(T, "second task", TestTurnTokens.inert(T));
+        CompletableFuture<Void> queued = injector.enqueue(T, "second task", TestTurnTokens.inert(T)).completion();
         CompletableFuture<Rendezvous.Resolution> waiter = rendezvous.currentWaiter(T);
         injector.drop(T, new HerdrException("agent target sol not found", "agent_not_found", null));
 
@@ -1768,7 +1768,10 @@ class MessageServiceTest {
         assertEquals(MessageService.Outcome.TIMED_OUT_QUEUED, r.outcome());
 
         assertTrue(messages.hasQueuedDelivery(T),
-                "a TIMED_OUT_QUEUED send leaves the message still queued in the injector");
+                "a TIMED_OUT_QUEUED send still records the undelivered delivery for fleet health");
+        injector.onStatus(T, AgentStatus.IDLE);
+        assertTrue(herdr.calls.stream().noneMatch(c -> c.method().equals("agent.prompt")),
+                "a TIMED_OUT_QUEUED send must be cancelled, not delivered when the worker later goes idle");
     }
 
     @Test

fleetd/src/test/java/dev/ltms/fleet/rest/FleetAppTest.java

@@ -711,13 +711,21 @@ class FleetAppTest {
 
     @Test
     void stopWorkerInPanePlacementClosesOnlyThePane() throws Exception {
-        FakeHerdr herdr = new FakeHerdr();
+        // fleetd #342: tab-cleanup resolution is no longer skipped based on a profile's declared
+        // placement — a stop() routed through the wrong delegate (e.g. CompositePeerLauncher's
+        // single-daemon shortcut after a daemon restart) could carry a placement config that says
+        // nothing true about how THIS pane was actually placed. So the pane's real tab is now
+        // always resolved, and the single-occupant check below is what protects a pane-placement
+        // peer's shared tab, exactly as it always protected a tab-placement one. Model that
+        // realistically: the peer's pane was split into an existing tab that already held another
+        // occupant, so the tab must never be closed.
+        FakeHerdr herdr = new FakeHerdr().withWorkerTabPaneCount(2);
         int port = start(herdr, "http://gx00.gw:8000", Set.of("gx00.gw"), "pane");
 
         assertEquals(204, req(port, "DELETE", "/members/w9:pW").statusCode());
         assertTrue(herdr.called("pane.close"));
-        assertFalse(herdr.called("tab.close"), "pane placement owns no tab to close");
-        assertFalse(herdr.called("pane.get"), "no tab resolution in pane placement");
+        assertTrue(herdr.called("pane.get"), "tab resolution now always runs, regardless of placement");
+        assertFalse(herdr.called("tab.close"), "pane placement's shared tab must never be closed");
     }
 
     @Test