#298: AmqpReplyInbox.release() requeues held deliveries instead of dropping them

release(target) used to cancel the target's consumer and clear held's local
record for it. Cancelling a consumer does not requeue the broker's in-flight
deliveries — they stay unacked on the still-open channel until a real
connection drop. So a held-but-undrained reply became permanently
unreachable: never acked, never nacked, never requeued, invisible to peek.

Fix: cancel the consumer first (so it can no longer receive redeliveries),
then nack-with-requeue every held delivery for that target before dropping
the local record. Nacking before the cancel was tried first but a real
broker demonstrated a race: the still-active consumer immediately received
the requeued message back, racing held.remove and leaving peek non-empty.
Cancelling first avoids that. A failed requeue is logged at WARN and does
not abort release(), matching the best-effort teardown style #293 settled
for HerdrPeerLauncher.stop().

Extends AmqpReplyInboxContractTest.releaseCancelsConsumer... to assert the
held delivery is recoverable via a later own(), not just absent from peek.

fleetd/src/main/java/dev/ltms/fleet/health/FleetHealthMonitor.java

@@ -14,6 +14,7 @@ import java.util.HashSet;
 import java.util.List;
 import java.util.Map;
 import java.util.Objects;
+import java.util.concurrent.ConcurrentHashMap;
 import java.util.concurrent.ScheduledExecutorService;
 import java.util.concurrent.TimeUnit;
 import java.util.function.BiConsumer;
@@ -46,7 +47,16 @@ public final class FleetHealthMonitor {
     private final long workingSuspectAfterNanos;
     private final BiConsumer<String, String> failTarget;
     private final Map<String, HealthPrior> priors = new HashMap<>();
-    private final Map<String, HealthState> states = new HashMap<>();
+    /**
+     * The live classification per member, and the only one of this class's three maps that more
+     * than one scheduler task touches. {@code tick} writes it (and prunes it to the roster);
+     * fleetd #280's delayed {@link #recheckTerminalTarget} reads it from its own separate scheduled
+     * task. Both run on the single-threaded scheduler {@code Fleetd} passes in today, so they are
+     * serialised — but nothing in this class enforces that, and an unsynchronised {@link HashMap}
+     * read racing a resize can spin a CPU forever rather than fail visibly. {@code priors} and
+     * {@code orphanStreaks} stay plain maps because {@code tick} is still their only toucher.
+     */
+    private final Map<String, HealthState> states = new ConcurrentHashMap<>();
     /**
      * CB-643: consecutive ticks on which a target looked like an orphaned delegation. The fact
      * {@link MessageService#hasOrphanedDelegation} reports is a true snapshot, but it can read true

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

@@ -913,9 +913,14 @@ public abstract class HerdrPeerLauncher implements PeerLauncher {
      * {@link #reapOrphanWorkers() orphan-reap} and spawn-gate-timeout paths, plus any caller that
      * passes a pane directly, keep working without an owning id.
      *
-     * <p>Resolves the tab from the pane <em>before</em> closing it. An already-gone pane/tab
-     * (repeated DELETE, crashed peer) is treated as success; any other failure propagates so a
-     * genuinely failed teardown is not reported as done.
+     * <p>Resolves the tab from the pane <em>before</em> closing it. {@code agents.close} (the pane)
+     * is the one step whose failure means the teardown itself may not have happened: an already-gone
+     * pane (repeated DELETE, crashed peer) is treated as success, but any other failure propagates so
+     * a genuinely failed teardown is not reported as done. {@code spaces.closeTab} (fleetd #293) is
+     * different — by the time it runs the pane is already closed, so it is cosmetic workspace tidying
+     * rather than a real teardown failure, and a failure there is logged and never propagates, so it
+     * cannot mask the two cleanups below it ({@link #releaseZdotdir}, and the caller's worktree
+     * removal in {@code SessionManager.release}).
      */
     @Override
     public void stop(String idOrPane) {
@@ -934,7 +939,25 @@ public abstract class HerdrPeerLauncher implements PeerLauncher {
             log.debug("pane.close({}) ignored — already gone: {}", paneId, e.getMessage());
         }
         if (loc != null && loc.tabPaneCount() == 1) {
-            spaces.closeTab(loc.tabId());
+            // fleetd #293: the pane above is already closed by this point, so a failing tab.close is
+            // cosmetic workspace tidying, not a real teardown failure — it must not mask the two
+            // cleanups below it (releaseZdotdir, and the caller's worktree removal). Unlike
+            // agents.close above, this is not narrowed to "already gone": any failure here, whatever
+            // its cause, is one we continue past, so we log it at WARN (not debug) with the tab id a
+            // person can go close by hand.
+            try {
+                spaces.closeTab(loc.tabId());
+            } catch (RuntimeException e) {
+                // Caught as RuntimeException, not HerdrException, to match releaseZdotdir's own
+                // guard five lines below. Today the two are the same set — HerdrCodec wraps every
+                // encode/decode failure and UnixSocketHerdrClient wraps every IOException, so
+                // HerdrException is all closeTab can actually throw. Narrowing to it anyway would
+                // leave this step guarded against the expected failure and bare against any other,
+                // which is the exact asymmetry fleetd #293 exists to remove. No behaviour change
+                // today; it stops a later change inside WorkspaceControl.closeTab reopening it.
+                log.warn("tab.close({}) failed — the pane is already torn down, so continuing; the "
+                        + "tab may need manual cleanup: {}", loc.tabId(), e.getMessage());
+            }
         } else if (loc != null) {
             log.debug("not closing tab {} — it holds {} panes (not a dedicated peer tab)",
                     loc.tabId(), loc.tabPaneCount());

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

@@ -208,18 +208,72 @@ public final class AmqpReplyInbox implements ReplyInbox, AutoCloseable {
         }
     }
 
+    /**
+     * Release ownership of {@code target}: cancel its consumer, then nack-with-requeue every
+     * delivery still held for it instead of just dropping the local record.
+     *
+     * <p><strong>Cancelling a consumer does not requeue its in-flight deliveries.</strong> In AMQP,
+     * a delivery that was pushed to a consumer stays unacked, attached to the still-open
+     * {@link #channel}, until that channel or the connection closes — {@code basicCancel} alone does
+     * neither. So before this method existed with a requeue step, it dropped {@link #held}'s entries
+     * for {@code target} while the broker still considered them outstanding: never acked, never
+     * nacked, never requeued, and no longer reachable by {@link #peek} — permanently invisible. This
+     * is unlike {@link #handleRecovery} and {@link #close()}, whose bare {@code held.clear()} is
+     * correct because each has already made the broker requeue (a real connection drop, or
+     * {@code channel.close()} respectively) before clearing local state.
+     *
+     * <p><strong>Order: cancel first, then nack.</strong> A delivery tag stays valid for
+     * {@code basicNack} on this channel regardless of whether its consumer is still attached — only
+     * a channel/connection close invalidates it — so cancelling {@code target}'s consumer first does
+     * not risk the tags. Doing it the other way round does: nacking a delivery with {@code requeue}
+     * while its consumer is still active hands the message straight back to that <em>same</em>
+     * consumer the instant a prefetch slot frees up (confirmed against a real broker — see
+     * {@code AmqpReplyInboxContractTest.releaseCancelsConsumerAndRequeuesHeldDeliveryForRecovery}),
+     * which races this method's own {@code held.remove(target)}: the redelivery can land after the
+     * clear and leave a stale entry behind, so {@link #peek} is no longer reliably empty right after
+     * {@link #release}. Cancelling first closes that consumer, so the requeued message goes back to
+     * the queue for whichever consumer picks it up next (a later {@link #own}), not this one.
+     *
+     * <p><strong>Failure of the requeue is best-effort, not fatal.</strong> {@link #release} runs
+     * during teardown ({@code Fleetd} calls it right after {@code MessageService.abandon}), and a
+     * throw here would abort cleanups the caller depends on — the same argument fleetd #293 settled
+     * for {@code HerdrPeerLauncher.stop()}'s tab-close step. So a failed {@code basicNack} is logged
+     * at WARN, naming the target and delivery tag that leaked, and release proceeds; the delivery
+     * stays unacked on the broker rather than being silently dropped, so it is still recoverable by a
+     * later connection drop even though this release did not manage to requeue it immediately. A
+     * failed {@code basicCancel} still throws, unchanged from before this fix — that failure means
+     * the consumer may still be attached, so best-effort requeue is not attempted underneath it.
+     */
     @Override
     public void release(String target) {
         synchronized (channelLock) {
             String tag = consumerTags.remove(target);
-            held.remove(target);                   // stale delivery tags must not survive release
-            if (tag == null) {
-                return;
+            if (tag != null) {
+                try {
+                    channel.basicCancel(tag);
+                } catch (IOException e) {
+                    throw new IllegalStateException("cannot cancel consumer for " + target, e);
+                }
             }
-            try {
-                channel.basicCancel(tag);
-            } catch (IOException e) {
-                throw new IllegalStateException("cannot cancel consumer for " + target, e);
+            var perTarget = held.remove(target);
+            if (perTarget != null) {
+                synchronized (perTarget) {
+                    for (Held h : perTarget.values()) {
+                        try {
+                            channel.basicNack(h.deliveryTag(), false, true); // requeue, don't drop
+                        } catch (IOException | RuntimeException e) {
+                            // Caught broadly (not just IOException) for the same reason #293 catches
+                            // RuntimeException in HerdrPeerLauncher.stop(): best-effort teardown must
+                            // not be guarded only against the expected failure and bare against any
+                            // other. The message stays unacked on the broker either way — not lost,
+                            // just not proactively requeued — until a connection drop frees it.
+                            log.warn("release({}): could not requeue held delivery (msgId={}, tag={})"
+                                            + " back to the broker — it stays unacked until a connection"
+                                            + " drop frees it: {}",
+                                    target, h.message().msgId(), h.deliveryTag(), e.getMessage());
+                        }
+                    }
+                }
             }
         }
     }

fleetd/src/test/java/dev/ltms/fleet/herdr/FakeHerdr.java

@@ -7,6 +7,7 @@ import java.util.ArrayList;
 import java.util.LinkedHashMap;
 import java.util.List;
 import java.util.Map;
+import java.util.concurrent.ConcurrentHashMap;
 import java.util.concurrent.CopyOnWriteArrayList;
 
 /**
@@ -41,6 +42,9 @@ public final class FakeHerdr implements HerdrClient {
     private int agentPaneBusyFor = 0;
     private int workerTabPaneCount = 1;
     private String paneCloseErrorCode = null;
+    private final Map<String, String> paneCloseErrorCodeFor = new ConcurrentHashMap<>();
+    private String tabCloseErrorCode = null;
+    private final Map<String, String> tabCloseErrorCodeFor = new ConcurrentHashMap<>();
     private String agentSendErrorCode = null;
     private boolean noPanes = false;
     private volatile String agentStatus = "idle"; // steady-state agent.get status
@@ -86,12 +90,44 @@ public final class FakeHerdr implements HerdrClient {
         return this;
     }
 
-    /** Make {@code pane.close} fail with this herdr error code. */
+    /** Make {@code pane.close} fail with this herdr error code, for every pane. */
     public FakeHerdr paneCloseFailsWith(String code) {
         this.paneCloseErrorCode = code;
         return this;
     }
 
+    /**
+     * Make {@code pane.close} fail with this herdr error code, but only for the given {@code
+     * pane_id} — every other pane's {@code pane.close} still succeeds. Unlike {@link
+     * #paneCloseFailsWith}, which fails every call regardless of which pane it targets, this lets a
+     * test reap/release several sessions at once and make exactly one of them fail to stop, so the
+     * others' teardown can be asserted to proceed normally (fleetd #290).
+     */
+    public FakeHerdr paneCloseFailsForPane(String paneId, String code) {
+        this.paneCloseErrorCodeFor.put(paneId, code);
+        return this;
+    }
+
+    /** Make {@code tab.close} fail with this herdr error code, for every tab. */
+    public FakeHerdr tabCloseFailsWith(String code) {
+        this.tabCloseErrorCode = code;
+        return this;
+    }
+
+    /**
+     * Make {@code tab.close} fail with this herdr error code, but only for the given {@code
+     * tab_id} — every other tab's {@code tab.close} still succeeds. The {@code tab.close}
+     * counterpart to {@link #paneCloseFailsForPane} (fleetd #290): lets a test make exactly one
+     * session's tab teardown fail while proving the rest of {@code stop()} — {@code
+     * releaseZdotdir}, and the caller's worktree removal — still runs (fleetd #293). Named "ForTab"
+     * rather than "ForPane" (unlike its sibling) because {@code tab.close} keys on {@code tab_id},
+     * not a pane id.
+     */
+    public FakeHerdr tabCloseFailsForTab(String tabId, String code) {
+        this.tabCloseErrorCodeFor.put(tabId, code);
+        return this;
+    }
+
     /**
      * Make {@code pane.list} report no panes at all — models a second herdr daemon (CB-185) that
      * simply does not host the pane a {@link PaneLocator} is searching for.
@@ -338,7 +374,17 @@ public final class FakeHerdr implements HerdrClient {
                             .formatted(workerTabPaneCount,
                                     seeded.isEmpty() ? "" : "," + String.join(",", seeded)));
                 }
-                case "tab.close" -> mapper.readTree("{\"type\":\"ok\"}");
+                case "tab.close" -> {
+                    Object tabIdParam = params instanceof Map<?, ?> m ? m.get("tab_id") : null;
+                    String perTabCode = tabIdParam == null ? null
+                            : tabCloseErrorCodeFor.get(String.valueOf(tabIdParam));
+                    String code = perTabCode != null ? perTabCode : tabCloseErrorCode;
+                    if (code != null) {
+                        throw new HerdrException("herdr error [" + code + "]: tab.close failed",
+                                code, null);
+                    }
+                    yield mapper.readTree("{\"type\":\"ok\"}");
+                }
                 case "pane.get" -> mapper.readTree("""
                         {"type":"pane_info","pane":{"pane_id":"w9:pW","workspace_id":"w9",
                           "tab_id":"w9:t2","agent_status":"idle"}}""");
@@ -360,9 +406,13 @@ public final class FakeHerdr implements HerdrClient {
                                   "foreground_processes":[]}}""");
                 }
                 case "pane.close" -> {
-                    if (paneCloseErrorCode != null) {
-                        throw new HerdrException("herdr error [" + paneCloseErrorCode + "]: pane.close failed",
-                                paneCloseErrorCode, null);
+                    Object paneIdParam = params instanceof Map<?, ?> m ? m.get("pane_id") : null;
+                    String perPaneCode = paneIdParam == null ? null
+                            : paneCloseErrorCodeFor.get(String.valueOf(paneIdParam));
+                    String code = perPaneCode != null ? perPaneCode : paneCloseErrorCode;
+                    if (code != null) {
+                        throw new HerdrException("herdr error [" + code + "]: pane.close failed",
+                                code, null);
                     }
                     yield mapper.readTree("{\"type\":\"ok\"}");
                 }

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

@@ -925,6 +925,71 @@ class ClaudeCodeLauncherTest {
         assertTrue(herdr.called("pane.close"), "stop via handle.id() must close the pane");
     }
 
+    /** A tab-placement launcher with {@code memberCredentials policy=allow-list} under a zsh shell — the
+     * combination that makes {@link HerdrPeerLauncher#spawn} generate a real ZDOTDIR, so {@code
+     * releaseZdotdir}'s effect (the directory's deletion) is observable from a test. */
+    private ClaudeCodeLauncher serviceWithAllowList(FakeHerdr herdr) {
+        FleetConfig.Profile cfg = new FleetConfig.Profile(
+                "ltms-local", "http://gx00.gw:8000", "coder", null, "FLEETD_WORKER_TOKEN",
+                List.of("ccs", "ltms-local"), "tab", "fleetd-workers",
+                "worker: {profile} #{n}", null, null, null);
+        Supplier<FleetConfig.MemberCredentials> creds = () -> new FleetConfig.MemberCredentials(
+                FleetConfig.MemberCredentials.POLICY_ALLOW_LIST, List.of(), List.of(), null);
+        Function<String, String> env = name -> "SHELL".equals(name) ? "/bin/zsh" : null;
+        return new ClaudeCodeLauncher(new AgentControl(herdr), new WorkspaceControl(herdr),
+                new SubscriptionGuard(Set.of("gx00.gw")), Map.of(cfg.profile(), cfg), cfg.profile(),
+                env, 0, System::currentTimeMillis, () -> { }, null, creds);
+    }
+
+    /**
+     * fleetd #293: {@code stop()} used to run {@code spaces.closeTab} bare — any non-{@code
+     * *_not_found} herdr error propagated straight out of {@code stop()}, skipping {@code
+     * releaseZdotdir} entirely (the pane was already closed by that point, so the tab-close failure
+     * is cosmetic, not a real teardown failure). Proves both halves of the fix: {@code stop()} no
+     * longer throws for this failure, and {@code releaseZdotdir} still runs — observed here by the
+     * generated ZDOTDIR actually being deleted, since {@code releaseZdotdir}'s last line is {@code
+     * EnvAllowListScrub.deleteRecursively(dir)}.
+     */
+    @Test
+    @SuppressWarnings("unchecked")
+    void stopStillReleasesZdotdirWhenCloseTabFailsWithANonNotFoundCode() {
+        FakeHerdr herdr = new FakeHerdr();
+        ClaudeCodeLauncher svc = serviceWithAllowList(herdr);
+        PeerHandle handle = svc.spawn(new SpawnRequest(null, null, null));
+        Map<String, Object> tabCreateParams = (Map<String, Object>) herdr.lastCall("tab.create").params();
+        Map<String, String> tabEnv = (Map<String, String>) tabCreateParams.get("env");
+        String zdotdir = tabEnv.get("ZDOTDIR");
+        assertNotNull(zdotdir, "policy=allow-list under a zsh shell must have generated a ZDOTDIR: " + tabEnv);
+        Path dir = Path.of(zdotdir);
+        assertTrue(Files.isDirectory(dir), "the generated ZDOTDIR must exist before stop(): " + dir);
+        herdr.tabCloseFailsForTab("w9:t2", "internal_error");
+
+        Logger logger = (Logger) LoggerFactory.getLogger(HerdrPeerLauncher.class);
+        ListAppender<ILoggingEvent> appender = new ListAppender<>();
+        appender.start();
+        logger.addAppender(appender);
+        try {
+            assertDoesNotThrow(() -> svc.stop(handle.id()),
+                    "fleetd #293: a failing tab.close is cosmetic — it must not propagate out of stop()");
+        } finally {
+            logger.detachAppender(appender);
+        }
+
+        assertTrue(herdr.called("tab.close"), "tab.close was still attempted");
+        assertFalse(Files.exists(dir),
+                "releaseZdotdir must still run and delete the generated ZDOTDIR despite the tab.close "
+                        + "failure: " + dir);
+        String warn = appender.list.stream()
+                .filter(e -> e.getLevel().equals(Level.WARN))
+                .map(ILoggingEvent::getFormattedMessage)
+                .filter(m -> m.contains("tab.close") && m.contains("w9:t2"))
+                .findFirst()
+                .orElse(null);
+        assertNotNull(warn, "the failing tab.close must be logged at WARN naming the tab id — a "
+                + "silently swallowed failure with no message is not an improvement. Log lines: "
+                + appender.list.stream().map(ILoggingEvent::getFormattedMessage).toList());
+    }
+
     // --- CB-519: host-unique id, decoupled from the pane coordinate ------------------------------
 
     @Test

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

@@ -154,7 +154,7 @@ class AmqpReplyInboxContractTest {
     }
 
     @Test
-    void releaseCancelsConsumerAndClearsHeld() throws Exception {
+    void releaseCancelsConsumerAndRequeuesHeldDeliveryForRecovery() throws Exception {
         String target = "worker-release-" + System.nanoTime();
         try (AmqpReplyInbox inbox = AmqpReplyInbox.open(uri())) {
             inbox.own(target);
@@ -164,6 +164,22 @@ class AmqpReplyInboxContractTest {
             inbox.release(target);
             assertTrue(inbox.peek(target).isEmpty(),
                     "release clears the local held snapshot");
+
+            // fleetd #298: release() must not just drop the local record — the broker delivery was
+            // never acked, so cancelling the consumer alone leaves it unacked-but-orphaned on the
+            // still-open channel unless release() nacks it back with requeue=true. Prove the message
+            // is genuinely recoverable, not merely absent from peek: re-own the same target and
+            // confirm the broker redelivers it to the fresh consumer.
+            inbox.own(target);
+            List<ReplyInbox.InboxMessage> recovered = awaitPeek(inbox, target);
+            assertEquals(1, recovered.size(),
+                    "a reply held (but undrained) at release() time must still be recoverable — "
+                            + "release() must requeue it, not silently drop it while the broker still "
+                            + "considers it outstanding");
+            assertEquals("m1", recovered.getFirst().msgId());
+            assertEquals("release me", recovered.getFirst().content());
+
+            inbox.ack(target, "m1");
         }
     }
 

fleetd/src/test/java/dev/ltms/fleet/session/SessionManagerTest.java

@@ -1069,6 +1069,79 @@ class SessionManagerTest {
         assertEquals(1, paneCloseCallsFor(herdr, "w9:pRoot_3"), "the third pane is stopped");
     }
 
+    /**
+     * fleetd #290: the #283 fix above closed the one trigger this suite used for {@code
+     * reapIdle}'s own per-session try/catch (CB-581) — a worktree-removal failure is now caught
+     * and logged inside {@code release()} itself, so it never reaches {@code reapIdle}'s guard at
+     * all. This test restores coverage of that guard using the trigger the ticket names: {@code
+     * release()} calls {@code launcher.stop(paneId)} with no try/catch around it, so a failing
+     * {@code pane.close} propagates straight out of {@code release()} uncaught. {@link
+     * FakeHerdr#paneCloseFailsForPane} (added for this ticket) makes exactly the middle session's
+     * stop fail, while the other two still succeed, so this proves {@code reapIdle} keeps reaping
+     * the rest of the roster rather than aborting the whole pass.
+     */
+    @Test
+    void reapIdleSurvivesOneSessionWhoseLauncherStopFails() {
+        long[] clock = {0};
+        FakeHerdr herdr = new FakeHerdr();
+        RecordingWorktrees worktrees = new RecordingWorktrees();
+        SessionManager sessions = sessionManager(herdr, worktrees, () -> clock[0]);
+        MemberSession a = sessions.acquire("ltms-local", null, "/caller/proj", null,
+                new WorktreeRequest("cb-290a", null));
+        MemberSession b = sessions.acquire("ltms-local", null, "/caller/proj", null,
+                new WorktreeRequest("cb-290b", null));
+        MemberSession c = sessions.acquire("ltms-local", null, "/caller/proj", null,
+                new WorktreeRequest("cb-290c", null));
+        sessions.asPresence().markPresent(a.terminalId());
+        sessions.asPresence().markPresent(b.terminalId());
+        sessions.asPresence().markPresent(c.terminalId());
+        // Only the middle session's herdr pane fails to close — a and c stop normally. This is the
+        // trigger reapIdle's own guard is for, now that #283 closed the worktree-removal trigger.
+        herdr.paneCloseFailsForPane("w9:pRoot_2", "internal_error");
+
+        LoggerContext ctx = (LoggerContext) LoggerFactory.getILoggerFactory();
+        ch.qos.logback.classic.Logger sessionLog =
+                (ch.qos.logback.classic.Logger) LoggerFactory.getLogger(SessionManager.class);
+        ListAppender<ILoggingEvent> appender = new ListAppender<>();
+        appender.setContext(ctx);
+        appender.start();
+        sessionLog.addAppender(appender);
+        sessionLog.setLevel(Level.WARN);
+        int reaped;
+        try {
+            clock[0] = 100;
+            reaped = sessions.reapIdle(10);
+
+            String warn = appender.list.stream()
+                    .filter(e -> e.getLevel().equals(Level.WARN))
+                    .map(ILoggingEvent::getFormattedMessage)
+                    .filter(m -> m.contains("reap failed") && m.contains(b.paneId()))
+                    .findFirst()
+                    .orElse("no reap-failed WARN logged for the failing session");
+            assertTrue(warn.contains(b.terminalId()), "the WARN names the failed session's terminal: " + warn);
+        } finally {
+            sessionLog.detachAppender(appender);
+        }
+
+        assertEquals(2, reaped,
+                "the middle session's launcher.stop failure is not counted as reaped, but must not "
+                        + "abort reaping the other two");
+        assertTrue(sessions.get(a.paneId()).isEmpty(), "the first session is still released");
+        assertTrue(sessions.get(c.paneId()).isEmpty(),
+                "the third session is still reached and released — proves the pass did not abort "
+                        + "when the middle session's release() threw");
+        assertTrue(sessions.get(b.paneId()).isEmpty(),
+                "the middle session is still deregistered — release() removes it from the registry "
+                        + "before launcher.stop() runs, regardless of whether stop() then throws");
+        assertEquals(1, paneCloseCallsFor(herdr, "w9:pRoot_1"), "the first pane is stopped");
+        assertEquals(1, paneCloseCallsFor(herdr, "w9:pRoot_2"),
+                "the middle pane's stop was attempted, even though it failed");
+        assertEquals(1, paneCloseCallsFor(herdr, "w9:pRoot_3"), "the third pane is stopped");
+        assertEquals(List.of(a.worktree(), c.worktree()), worktrees.removeCalls().stream().sorted().toList(),
+                "the middle session's worktree removal never runs — release() throws before reaching "
+                        + "it — while the other two, unaffected, still have theirs removed");
+    }
+
     @Test
     void unchangedRegressionCleanCompletedReleaseStillRemovesTheWorktree() {
         FakeHerdr herdr = new FakeHerdr();
@@ -1083,6 +1156,41 @@ class SessionManagerTest {
                 "COMPLETED release of a clean worktree still removes it");
     }
 
+    /**
+     * fleetd #293: {@code HerdrPeerLauncher.stop()} used to run {@code spaces.closeTab} bare — a
+     * failing {@code tab.close} (any code other than {@code *_not_found}) propagated straight out
+     * of {@code stop()}. {@code SessionManager.release} calls {@code launcher.stop(paneId)} with
+     * no try/catch (fleetd #283 wrapped the WORKTREE-removal step further down, not this one), so
+     * the throw happened <em>before</em> that worktree-removal step ever ran — and by then {@code
+     * registry.remove(paneId)} had already run, so a second {@code stop} is a no-op: the worktree
+     * leaked with no retry path. The pane itself is already closed by the time {@code tab.close}
+     * runs, so its failure is cosmetic workspace tidying, not a real teardown failure. The fix
+     * wraps {@code closeTab} inside {@code stop()} so it no longer throws for this reason; this
+     * test proves both halves at once: {@code release()} does not throw, and it still removes the
+     * worktree.
+     */
+    @Test
+    void releaseStillRemovesTheWorktreeWhenCloseTabFails() {
+        FakeHerdr herdr = new FakeHerdr();
+        RecordingWorktrees worktrees = new RecordingWorktrees();
+        SessionManager sessions = sessionManager(herdr, worktrees);
+        MemberSession s = sessions.acquire("ltms-local", null, "/caller/proj", null,
+                new WorktreeRequest("cb-293a", null));
+        // FakeHerdr's pane.get always answers with tab_id "w9:t2" for a tab-placement spawn.
+        herdr.tabCloseFailsForTab("w9:t2", "internal_error");
+
+        assertDoesNotThrow(() -> sessions.release(s.paneId()),
+                "a failing tab.close is cosmetic (the pane is already closed by then) — it must not "
+                        + "propagate out of release()");
+
+        assertTrue(herdr.called("tab.close"), "tab.close was still attempted");
+        assertEquals(List.of(s.worktree()), worktrees.removeCalls(),
+                "release() must still remove the worktree even though tab.close failed — this is "
+                        + "the leak fleetd #293 reports: before the fix, release() never reached this "
+                        + "step at all");
+        assertTrue(sessions.get(s.paneId()).isEmpty(), "the session is still deregistered");
+    }
+
     @Test
     void unchangedRegressionDirtyCompletedReleaseStillPreservesTheWorktree() {
         FakeHerdr herdr = new FakeHerdr();