M4: fail tickets on terminal health states

The criterion required the token to bind the session turn number. Three
independent refusals from the implementer showed why that is not
implementable at this layer: MessageService owns acceptance but never
learns of delivery, and CompletionResolver.onDelivered runs before
SessionManager.onDelivered, so the turn number does not exist yet at the
only point the token could capture it.

Records both rejected alternatives and why, so the next reader does not
re-derive them: a target-keyed registry restores the ambiguity the token
exists to remove, and injecting a turn counter couples layers to fill a
field nothing reads yet.

bridged/bridged.example.yaml

@@ -85,6 +85,16 @@ bind:
 #   backoffMs: 60000
 #   quietNudgeCap: 3
 
+# Fleet health detection is dormant unless enabled. It reads one whole-fleet agent list per tick.
+# It can run without a webhook; bridge_list then reports healthCoverage: detection-only.
+# health:
+#   enabled: true
+#   intervalSeconds: 30              # minimum 15
+#   workingSuspectAfterSeconds: 600  # minimum 300
+#   paneProbeIntervalSeconds: 60     # minimum 60
+#   notifications:
+#     mode: disabled                 # disabled (default) or webhook
+
 # herdr Unix socket. Omit to use the client default
 # (${HERDR_SOCKET_PATH:-~/.config/herdr/herdr.sock}).
 herdrSocket: ~/.config/herdr/herdr.sock

bridged/src/main/java/dev/ltms/bridged/Bridged.java

@@ -23,6 +23,7 @@ import dev.ltms.bridged.mcp.BridgeMcp;
 import dev.ltms.bridged.mcp.ConnectionIdentity;
 import dev.ltms.bridged.metrics.BridgedMetrics;
 import dev.ltms.bridged.metrics.Metrics;
+import dev.ltms.bridged.health.FleetHealthMonitor;
 import dev.ltms.bridged.mcp.PrimaryRegistry;
 import dev.ltms.bridged.mcp.LsofPeerPidLookup;
 import dev.ltms.bridged.mcp.LsofProcessCwdLookup;
@@ -54,6 +55,7 @@ import java.util.Map;
 import java.util.Objects;
 import java.util.Set;
 import java.util.concurrent.Executors;
+import java.util.concurrent.atomic.AtomicReference;
 import java.util.concurrent.TimeUnit;
 import java.util.concurrent.atomic.AtomicReference;
 import java.util.function.Function;
@@ -253,6 +255,7 @@ public final class Bridged {
         // CB-106: a confirmed turn completion resolves a blocked send whose worker never replied.
         Rendezvous rendezvous = new Rendezvous();
         CompletionResolver completion = new CompletionResolver(agents, rendezvous);
+        AtomicReference<MessageService> messagesRef = new AtomicReference<>();
         // CB-113: deliver only to an available worker (its MCP is connected), never its boot window.
         // CB-301: the manager's presence bridge records availability and drives SPAWNING → READY.
         MemberPresence presence = sessions.asPresence();
@@ -284,6 +287,14 @@ public final class Bridged {
             public void onTurnFailed(String target) {
                 completion.onTurnFailed(target);
                 sessions.onTurnFailed(target);
+                failTarget(messagesRef, target, "worker turn failed");
+            }
+
+            @Override
+            public void onTurnFailed(String target, String reason) {
+                completion.onTurnFailed(target, reason);
+                sessions.onTurnFailed(target);
+                failTarget(messagesRef, target, reason);
             }
         };
         Injector injector = new Injector(agents, turnListener, deliverableTo(presence, leads),
@@ -347,6 +358,27 @@ public final class Bridged {
         }
         MessageService messages = new MessageService(agents, injector, rendezvous, replyInbox,
                 pushLoop, metrics);
+        messagesRef.set(messages);
+
+        // Health is a slow whole-fleet observer. Keep it separate from the 250ms delivery poller.
+        final FleetHealthMonitor healthMonitor;
+        var healthScheduler = Executors.newSingleThreadScheduledExecutor(r ->
+                Thread.ofVirtual().name("bridge-health-").unstarted(r));
+        if (cfg.health() != null && cfg.health().isEnabled()) {
+            healthMonitor = new FleetHealthMonitor(agents, sessions::roster, messages, healthScheduler,
+                    System::nanoTime, cfg.health().intervalOrDefault(), messages::abandon);
+            String coverage = FleetHealthMonitor.coverage(true,
+                    cfg.health().notifications() != null && cfg.health().notifications().configured());
+            if ("detection-only".equals(coverage)) {
+                log.warn("fleet health: {} (no notification sink configured)", coverage);
+            } else {
+                log.info("fleet health: {}", coverage);
+            }
+            healthMonitor.start();
+        } else {
+            healthMonitor = null;
+            healthScheduler.shutdownNow();
+        }
 
         // CB-520: the reply inbox only consumes for agents this gateway owns. own on acquire,
         // release on teardown. Do this before CB-516 so the inbox is owned before any reply can land.
@@ -387,7 +419,12 @@ public final class Bridged {
                 profile -> {
                     var configured = config.get().profiles().get(profile);
                     return configured == null ? null : configured.maxLoad();
-                }, () -> config.get().profiles().keySet(), System::nanoTime));
+                }, () -> config.get().profiles().keySet(), System::nanoTime),
+                new BridgeMcp.HealthCoverageSource(() -> {
+                    var health = config.get().health();
+                    return FleetHealthMonitor.coverage(health != null && health.isEnabled(),
+                            health != null && health.notifications() != null && health.notifications().configured());
+                }));
 
         // CB-559: opt-in config reload. With no `configReload:` block nothing is constructed, so an
         // upgraded daemon behaves exactly as before — the file is read once at boot and never again.
@@ -408,6 +445,7 @@ public final class Bridged {
             messages.close();
             pushLoop.close();
             if (heartbeat != null) heartbeat.close(); // CB-551: stop the idle-lead heartbeat scheduler
+            if (healthMonitor != null) healthMonitor.stop();
             if (configWatcher != null) configWatcher.stop(); // CB-559: stop polling the config file
             mcp.close();
             if (reaper != null) reaper.stop();
@@ -452,6 +490,14 @@ public final class Bridged {
         return target -> presence.isPresent(target) || leads.get().containsKey(target);
     }
 
+    /** Fail outstanding tickets without changing the member lifecycle or worktree state. */
+    private static void failTarget(AtomicReference<MessageService> messagesRef, String target, String reason) {
+        MessageService messages = messagesRef.get();
+        if (messages != null) {
+            messages.abandon(target, reason);
+        }
+    }
+
     /**
      * Poll herdr's {@code ping} until it answers or {@link #HERDR_WAIT_SECONDS} elapses (CB-504).
      *

bridged/src/main/java/dev/ltms/bridged/config/BridgedConfig.java

@@ -73,6 +73,7 @@ public record BridgedConfig(
         Primary primary,
         Fleet fleet,
         LeadHeartbeat leadHeartbeat,
+        Health health,
         String placement,
         Auth auth,
         ConfigReload configReload) {
@@ -83,7 +84,16 @@ public record BridgedConfig(
                          Integer spawnReadyPollMs, Broker broker, Primary primary, Fleet fleet,
                          LeadHeartbeat leadHeartbeat, String placement, Auth auth) {
         this(bind, herdrSocket, profiles, guard, worktreeRoot, lifecycle, spawnReadyTimeoutMs,
-                spawnReadyPollMs, broker, primary, fleet, leadHeartbeat, placement, auth, null);
+                spawnReadyPollMs, broker, primary, fleet, leadHeartbeat, null, placement, auth, null);
+    }
+
+    /** Back-compat form before the optional {@code health:} block was added. */
+    public BridgedConfig(Bind bind, String herdrSocket, Map<String, Profile> profiles, Guard guard,
+                         String worktreeRoot, Lifecycle lifecycle, Integer spawnReadyTimeoutMs,
+                         Integer spawnReadyPollMs, Broker broker, Primary primary, Fleet fleet,
+                         LeadHeartbeat leadHeartbeat, String placement, Auth auth, ConfigReload configReload) {
+        this(bind, herdrSocket, profiles, guard, worktreeRoot, lifecycle, spawnReadyTimeoutMs,
+                spawnReadyPollMs, broker, primary, fleet, leadHeartbeat, null, placement, auth, configReload);
     }
 
     /**
@@ -377,6 +387,17 @@ public record BridgedConfig(
                             boolean clearAfterTurn) {
     }
 
+    /** Optional fleet detection. A missing block stays dormant. */
+    @JsonIgnoreProperties(ignoreUnknown = true)
+    public record Health(Boolean enabled, Integer intervalSeconds, Integer workingSuspectAfterSeconds,
+                         Integer paneProbeIntervalSeconds, Notifications notifications) {
+        public boolean isEnabled() { return Boolean.TRUE.equals(enabled); }
+        public int intervalOrDefault() { return Math.max(15, intervalSeconds == null ? 30 : intervalSeconds); }
+        public record Notifications(String mode) {
+            public boolean configured() { return "webhook".equalsIgnoreCase(mode); }
+        }
+    }
+
     /**
      * External AMQP broker for durable, cross-restart reply delivery (CB-307 Stage 2). Its mere
      * presence swaps the in-memory {@code ReplyInbox} for the AMQP-backed adapter; absent, bridged
@@ -812,7 +833,7 @@ public record BridgedConfig(
     private static final Set<String> KNOWN_TOP_LEVEL_KEYS = Set.of(
             "bind", "herdrSocket", "profiles", "guard", "worktreeRoot",
             "lifecycle", "spawnReadyTimeoutMs", "spawnReadyPollMs", "broker", "primary", "fleet",
-            "leadHeartbeat", "placement", "auth", "configReload");
+            "leadHeartbeat", "health", "placement", "auth", "configReload");
 
     /** Load and validate config from {@code path}. */
     public static BridgedConfig load(Path path) {
@@ -1082,7 +1103,7 @@ public record BridgedConfig(
         // defaults the fields of a block that IS present. Defaulting it here would start watching
         // the file for every config that never asked to be watched.
         return new BridgedConfig(b, herdrSocket, profiles, g, worktreeRoot, l, timeout, pollMs,
-                broker, primary, f, leadHeartbeat, placementOrDefault, a, configReload);
+                broker, primary, f, leadHeartbeat, health, placementOrDefault, a, configReload);
     }
 
     /**

bridged/src/main/java/dev/ltms/bridged/health/FleetHealthMonitor.java

@@ -0,0 +1,119 @@
+package dev.ltms.bridged.health;
+
+import dev.ltms.bridged.herdr.Agent;
+import dev.ltms.bridged.herdr.AgentControl;
+import dev.ltms.bridged.herdr.AgentStatus;
+import dev.ltms.bridged.msg.MessageService;
+import dev.ltms.bridged.session.MemberSession;
+import org.slf4j.Logger;
+import org.slf4j.LoggerFactory;
+
+import java.util.HashMap;
+import java.util.HashSet;
+import java.util.List;
+import java.util.Map;
+import java.util.concurrent.ScheduledExecutorService;
+import java.util.concurrent.TimeUnit;
+import java.util.function.LongSupplier;
+import java.util.function.Supplier;
+import java.util.function.BiConsumer;
+
+/** Slow whole-fleet evidence collection. It is deliberately separate from the delivery poller. */
+public final class FleetHealthMonitor {
+    private static final Logger log = LoggerFactory.getLogger(FleetHealthMonitor.class);
+    private final AgentControl agents;
+    private final Supplier<List<MemberSession>> roster;
+    private final MessageService messages;
+    private final ScheduledExecutorService scheduler;
+    private final LongSupplier clock;
+    private final long intervalSeconds;
+    private final BiConsumer<String, String> failTarget;
+    private final Map<String, HealthPrior> priors = new HashMap<>();
+    private final Map<String, HealthState> states = new HashMap<>();
+
+    // These facts need the evidence publishers introduced by later M4 units. They are not negatives.
+    private static final boolean NOT_YET_OBSERVED = false;
+
+    public FleetHealthMonitor(AgentControl agents, Supplier<List<MemberSession>> roster, MessageService messages,
+                               ScheduledExecutorService scheduler, LongSupplier clock, long intervalSeconds) {
+        this(agents, roster, messages, scheduler, clock, intervalSeconds, (_, _) -> { });
+    }
+
+    public FleetHealthMonitor(AgentControl agents, Supplier<List<MemberSession>> roster, MessageService messages,
+                               ScheduledExecutorService scheduler, LongSupplier clock, long intervalSeconds,
+                               BiConsumer<String, String> failTarget) {
+        this.agents = agents;
+        this.roster = roster;
+        this.messages = messages;
+        this.scheduler = scheduler;
+        this.clock = clock;
+        this.intervalSeconds = intervalSeconds;
+        this.failTarget = failTarget;
+    }
+
+    /** Pure per-member decision seam. */
+    static HealthDecision decide(HealthSnapshot snapshot, HealthPrior prior, long nowNanos) {
+        return FleetHealth.decide(snapshot, prior, nowNanos);
+    }
+
+    public void start() { scheduler.schedule(this::tick, intervalSeconds, TimeUnit.SECONDS); }
+    public void stop() { scheduler.shutdownNow(); }
+
+    // Package-private so tests can run one tick without waiting.
+    void tick() {
+        try {
+            List<Agent> agentsNow = agents.list(); // Exactly one list call for this complete observation.
+            List<MemberSession> rosterNow = roster.get(); // One in-memory roster snapshot for this tick.
+            Map<String, Agent> live = new HashMap<>();
+            for (Agent agent : agentsNow) live.put(agent.terminalId(), agent);
+            HashSet<String> current = new HashSet<>();
+            for (MemberSession session : rosterNow) {
+                current.add(session.terminalId());
+                Agent agent = live.get(session.terminalId());
+                AgentStatus status = agent == null ? AgentStatus.UNKNOWN : agent.status();
+                boolean accepted = messages.hasAcceptedDelivery(session.terminalId());
+                HealthSnapshot snapshot = new HealthSnapshot(session.state(), status, accepted, NOT_YET_OBSERVED,
+                        messages.hasInboxMessage(session.terminalId()), agent != null, NOT_YET_OBSERVED,
+                        NOT_YET_OBSERVED, NOT_YET_OBSERVED, NOT_YET_OBSERVED, NOT_YET_OBSERVED, NOT_YET_OBSERVED);
+                HealthDecision decision = decide(snapshot, priors.getOrDefault(session.terminalId(), HealthPrior.NONE),
+                        clock.getAsLong());
+                priors.put(session.terminalId(), decision.prior());
+                if (decision.state() == HealthState.GONE || decision.state() == HealthState.NEVER_READY) {
+                    failTarget.accept(session.terminalId(), "fleet health detected " + decision.state());
+                }
+                reportTransition(session.terminalId(), decision.state());
+            }
+            priors.keySet().retainAll(current);
+            states.keySet().retainAll(current);
+        } catch (Throwable error) {
+            // A list failure is health evidence, and must never kill the monitor's only scheduler task.
+            log.warn("fleet health collection failed; will retry next tick", error);
+        } finally {
+            if (!scheduler.isShutdown()) {
+                scheduler.schedule(this::tick, intervalSeconds, TimeUnit.SECONDS);
+            }
+        }
+    }
+
+    void reportTransition(String target, HealthState next) {
+        HealthState previous = states.put(target, next);
+        if (previous == next) return;
+        if (fault(next)) {
+            log.warn("fleet health member={} state={} previous={}", target, next, previous);
+        } else if (previous != null && fault(previous)) {
+            log.info("fleet health member={} recovered state={} previous={}", target, next, previous);
+        }
+    }
+
+    private static boolean fault(HealthState state) {
+        return switch (state) {
+            case NEVER_READY, GONE, TURN_BOUNDARY_LOST, ERROR_ON_SCREEN, STALL_SUSPECTED,
+                    MUTE, REPLY_STRANDED, DELEGATION_ORPHANED, CONTROL_LINK_DOWN -> true;
+            default -> false;
+        };
+    }
+
+    public static String coverage(boolean enabled, boolean notificationConfigured) {
+        return !enabled ? "off" : notificationConfigured ? "full" : "detection-only";
+    }
+}

bridged/src/main/java/dev/ltms/bridged/inject/CompletionResolver.java

@@ -137,7 +137,13 @@ public final class CompletionResolver implements TurnListener {
     @Override
     public void onTurnFailed(String target) {
         InFlight turn = inFlight.get(target);
-        Thread.ofVirtual().name("turn-failed-" + target).start(() -> fail(target, turn));
+        Thread.ofVirtual().name("turn-failed-" + target).start(() -> fail(target, turn, null));
+    }
+
+    @Override
+    public void onTurnFailed(String target, String reason) {
+        InFlight turn = inFlight.get(target);
+        Thread.ofVirtual().name("turn-failed-" + target).start(() -> fail(target, turn, reason));
     }
 
     /** Synchronous resolve (the unit-testable core of {@link #onTurnComplete}). */
@@ -192,6 +198,11 @@ public final class CompletionResolver implements TurnListener {
 
     /** Synchronous fail (the unit-testable core of {@link #onTurnFailed}). */
     void fail(String target, InFlight turn) {
+        fail(target, turn, null);
+    }
+
+    /** Synchronous fail with an optional reason supplied by a dropped worker queue. */
+    void fail(String target, InFlight turn, String explicitReason) {
         // A never-delivered readiness failure has no in-flight record but still has a blocked send;
         // fall back to the currently-registered waiter (unambiguous — that send never completed, so
         // no next turn exists to confuse it with).
@@ -201,16 +212,18 @@ public final class CompletionResolver implements TurnListener {
             inFlight.remove(target, turn); // nobody blocked on this worker — nothing to fail
             return;
         }
-        String reason;
-        try {
-            reason = clip(agents.read(target, SCRAPE_SOURCE));
-        } catch (RuntimeException e) {
-            reason = "";
-        }
-        if (reason.isBlank()) {
-            // No screen to scrape — either the worker is stuck (CB-109) or gone (CB-110).
-            reason = "worker did not reply; its turn ended in an unrecoverable state "
-                    + "(worker unreachable or stuck)";
+        String reason = explicitReason;
+        if (reason == null || reason.isBlank()) {
+            try {
+                reason = clip(agents.read(target, SCRAPE_SOURCE));
+            } catch (RuntimeException e) {
+                reason = "";
+            }
+            if (reason.isBlank()) {
+                // No screen to scrape — either the worker is stuck (CB-109) or gone (CB-110).
+                reason = "worker did not reply; its turn ended in an unrecoverable state "
+                        + "(worker unreachable or stuck)";
+            }
         }
         if (rendezvous.resolveFailure(waiter, reason)) {
             inFlight.remove(target, turn);

bridged/src/main/java/dev/ltms/bridged/inject/Injector.java

@@ -410,8 +410,8 @@ public final class Injector {
         for (Pending p : pending) {
             p.delivered().completeExceptionally(cause);
         }
-        if (hadDeliveredTurn) {
-            turnListener.onTurnFailed(target);
-        }
+        // A queued send has no in-flight record, while a delivered turn does. CompletionResolver
+        // handles both forms and resolves its waiter at most once.
+        turnListener.onTurnFailed(target, cause.getMessage());
     }
 }

bridged/src/main/java/dev/ltms/bridged/inject/TurnListener.java

@@ -41,6 +41,14 @@ public interface TurnListener {
     default void onTurnFailed(String target) {
     }
 
+    /**
+     * As {@link #onTurnFailed(String)}, carrying the reason a worker became unreachable. The default
+     * keeps existing listeners working while allowing the completion resolver to report a useful cause.
+     */
+    default void onTurnFailed(String target, String reason) {
+        onTurnFailed(target);
+    }
+
     /**
      * A message was just delivered into {@code target}'s pane (CB-115). Fired so the completion
      * resolver can snapshot the pane's pre-turn content: a later {@link #onTurnComplete} whose

bridged/src/main/java/dev/ltms/bridged/logging/McpCancelledNotificationFilter.java

@@ -0,0 +1,30 @@
+package dev.ltms.bridged.logging;
+
+import ch.qos.logback.classic.Level;
+import ch.qos.logback.classic.Logger;
+import ch.qos.logback.classic.turbo.TurboFilter;
+import ch.qos.logback.core.spi.FilterReply;
+import io.modelcontextprotocol.spec.McpSchema;
+import org.slf4j.Marker;
+
+/** Suppresses only the SDK warning for the normal MCP cancellation notification. */
+public final class McpCancelledNotificationFilter extends TurboFilter {
+
+    static final String LOGGER = "io.modelcontextprotocol.spec.McpStreamableServerSession";
+    static final String UNHANDLED_NOTIFICATION = "No handler registered for notification method: {}";
+
+    @Override
+    public FilterReply decide(Marker marker, Logger logger, Level level, String format, Object[] params,
+                              Throwable throwable) {
+        if (level == Level.WARN
+                && LOGGER.equals(logger.getName())
+                && UNHANDLED_NOTIFICATION.equals(format)
+                && params != null
+                && params.length == 1
+                && params[0] instanceof McpSchema.JSONRPCNotification notification
+                && "notifications/cancelled".equals(notification.method())) {
+            return FilterReply.DENY;
+        }
+        return FilterReply.NEUTRAL;
+    }
+}

bridged/src/main/java/dev/ltms/bridged/mcp/BridgeMcp.java

@@ -77,6 +77,7 @@ public final class BridgeMcp {
     private final CallerResolver authz; // CB-501: null → authorization not enforced (legacy)
     private final Metrics metrics;      // CB-502: null → auth failures not counted
     private final CapacitySource capacity;
+    private final HealthCoverageSource healthCoverage;
 
     /** Capacity facts used by {@code bridge_list}; production must supply the placement live count. */
     public record CapacitySource(Function<String, Integer> liveCount, Function<String, Integer> maxLoad,
@@ -86,6 +87,9 @@ public final class BridgeMcp {
         boolean available() { return !configuredProfiles.get().isEmpty(); }
     }
 
+    /** Coverage is supplied by the health wiring, not inferred from a missing dependency. */
+    public record HealthCoverageSource(Supplier<String> value) { }
+
     /**
      * @param callers resolves each call's {@link Principal}; {@code null} disables authorization.
      *                This surface needs its own enforcement: {@code /mcp} is a raw servlet on
@@ -95,8 +99,9 @@ public final class BridgeMcp {
      */
     public BridgeMcp(MessageService messages, PeerLauncher workers, SessionManager sessions,
                      ConnectionIdentity identity, MemberPresence presence, PrimaryRegistry primaryRegistry,
-                     CallerResolver callers, Metrics metrics, CapacitySource capacity) {
+                      CallerResolver callers, Metrics metrics, CapacitySource capacity, HealthCoverageSource healthCoverage) {
         this.capacity = capacity;
+        this.healthCoverage = healthCoverage;
         McpJsonMapper json = new JacksonMcpJsonMapperSupplier().get();
         this.transport = HttpServletStreamableServerTransportProvider.builder()
                 .jsonMapper(json)
@@ -210,7 +215,7 @@ public final class BridgeMcp {
                 .toolCall(listTool(), (exchange, _) -> {
                     McpSchema.CallToolResult denied = deny(exchange, Authz.Action.READ, null);
                     if (denied != null) return denied;
-                    return listFleet(workers, sessions, messages, capacity,
+                    return listFleet(workers, sessions, messages, capacity, healthCoverage,
                             callers == null ? Map.of() : callers.leads(),
                             callerTerminal(exchange));
                 })
@@ -724,11 +729,11 @@ public final class BridgeMcp {
      */
     static McpSchema.CallToolResult listFleet(PeerLauncher workers, SessionManager sessions,
                                               Map<String, String> leads, String selfTerm) {
-        return listFleet(workers, sessions, null, CapacitySource.none(), leads, selfTerm);
+        return listFleet(workers, sessions, null, CapacitySource.none(), new HealthCoverageSource(() -> "off"), leads, selfTerm);
     }
 
     static McpSchema.CallToolResult listFleet(PeerLauncher workers, SessionManager sessions, MessageService messages,
-                                              CapacitySource capacity,
+                                               CapacitySource capacity, HealthCoverageSource healthCoverage,
                                               Map<String, String> leads, String selfTerm) {
         try {
             Map<String, Agent> live = workers.list().stream()
@@ -747,6 +752,7 @@ public final class BridgeMcp {
             roster.stream().map(MemberSession::profile).forEach(profiles::add);
             Map<String, Object> result = new LinkedHashMap<>();
             result.put("leads", leadRows); result.put("members", out);
+            result.put("healthCoverage", healthCoverage.value().get());
             if (capacity.available()) result.put("capacity", profiles.stream()
                     .map(profile -> capacityView(profile, capacity.liveCount(), capacity.maxLoad(), roster, messages,
                             capacity.clock().getAsLong())).toList());

bridged/src/main/java/dev/ltms/bridged/msg/MessageService.java

@@ -170,8 +170,9 @@ public final class MessageService {
     private final Metrics metrics; // CB-502: nullable — no registry in unit tests
     private final ConcurrentHashMap<String, ReentrantLock> sessionLocks = new ConcurrentHashMap<>();
     private final ConcurrentHashMap<String, Task> tasks = new ConcurrentHashMap<>();
-    /** The async task that currently owns a target's send lock. */
-    private final ConcurrentHashMap<String, Task> asyncTasksByTarget = new ConcurrentHashMap<>();
+    /** Async task that owns each exact forward rendezvous waiter. */
+    private final ConcurrentHashMap<CompletableFuture<Rendezvous.Resolution>, Task> asyncTasksByWaiter =
+            new ConcurrentHashMap<>();
     /** Async tickets paused on a specific {@code bridge_ask} turn. */
     private final ConcurrentHashMap<String, Task> asyncTasksByTurn = new ConcurrentHashMap<>();
     private final AtomicLong ticketSeq = new AtomicLong();
@@ -300,14 +301,18 @@ public final class MessageService {
      */
     public boolean abandon(String target, String reason) {
         CompletableFuture<Rendezvous.Resolution> waiter = rendezvous.currentWaiter(target);
-        if (waiter == null || waiter.isDone()) {
-            return false; // nobody is blocked on this worker — nothing to abandon
+        boolean failed = waiter != null && !waiter.isDone() && rendezvous.resolveFailure(waiter, reason);
+        boolean asyncFailed = false;
+        for (Task task : tasks.values()) {
+            if (target.equals(task.target) && task.question == null
+                    && task.future.complete(new Reply(Outcome.WORKER_FAILED, reason))) {
+                asyncFailed = true;
+            }
         }
-        boolean failed = rendezvous.resolveFailure(waiter, reason);
         if (failed) {
             log.warn("abandoning the blocked send to {}: {}", target, reason);
         }
-        return failed;
+        return failed || asyncFailed;
     }
 
     /**
@@ -354,6 +359,11 @@ public final class MessageService {
      * never earned. {@code null} disables the hook.
      */
     public Reply send(String target, String content, long timeoutMillis, Runnable onAccepted) {
+        return send(target, content, timeoutMillis, onAccepted, null);
+    }
+
+    /** Run a send, optionally stopping an async task that teardown already failed before acceptance. */
+    private Reply send(String target, String content, long timeoutMillis, Runnable onAccepted, Task task) {
         long deadlineNanos = System.nanoTime() + timeoutMillis * 1_000_000L;
         ReentrantLock lock = sessionLocks.computeIfAbsent(target, _ -> new ReentrantLock());
 
@@ -361,6 +371,9 @@ public final class MessageService {
             return new Reply(Outcome.BUSY, null); // another send held the session the whole window
         }
         try {
+            if (task != null && task.future.isDone()) {
+                return task.future.getNow(null);
+            }
             if (hasAsyncQuestion(target)) {
                 return new Reply(Outcome.BUSY, null); // the worker's current turn is paused for its lead
             }
@@ -372,6 +385,9 @@ public final class MessageService {
             // failed send leaves no stale waiter behind.
             CompletableFuture<Rendezvous.Resolution> reply = rendezvous.open(target);
             try {
+                if (task != null) {
+                    asyncTasksByWaiter.put(reply, task);
+                }
                 // The send has won the lock; the accepted-delivery hook records delegator ownership
                 // here (CB-548). It runs BEFORE enqueue so a throwing hook — onAccepted is now a
                 // public callback — fails the send without queuing a message that would orphan.
@@ -395,6 +411,7 @@ public final class MessageService {
                     throw new IllegalStateException("interrupted awaiting reply from " + target, e);
                 }
             } finally {
+                asyncTasksByWaiter.remove(reply);
                 rendezvous.close(target, reply);
             }
         } finally {
@@ -420,11 +437,15 @@ public final class MessageService {
         if (ticket.fresh()) {
             // Register the reverse waiter first, then surface the question — so the answer, which can
             // arrive the instant the primary reacts, always finds an open waiter to resolve.
+            CompletableFuture<Rendezvous.Resolution> waiter = rendezvous.currentWaiter(workerSession);
+            Task task = markAsyncQuestion(waiter, question, ticket.turnId());
             if (!rendezvous.resolveQuestion(workerSession, question, ticket.turnId())) {
+                if (task != null) {
+                    clearAsyncQuestion(ticket.turnId(), true);
+                }
                 rendezvous.closeAsk(ticket.turnId());
                 return new AskResult(AskOutcome.NO_WAITER, null); // no primary is blocked on this worker
             }
-            markAsyncQuestion(workerSession, question, ticket.turnId());
         }
         try {
             String answer = ticket.answer().get(timeoutMillis, TimeUnit.MILLISECONDS);
@@ -523,21 +544,15 @@ public final class MessageService {
         tasks.put(ticket, task);
         asyncExecutor.submit(() -> {
             try {
-                Runnable trackingAccepted = () -> {
-                    if (onAccepted != null) {
-                        onAccepted.run();
-                    }
-                    asyncTasksByTarget.put(target, task);
-                };
-                Reply result = send(target, content, ASYNC_TIMEOUT_MS, trackingAccepted);
+                Reply result = send(target, content, ASYNC_TIMEOUT_MS, onAccepted, task);
                 if (result.outcome() == Outcome.QUESTION) {
-                    asyncTasksByTarget.remove(target, task);
+                    // Keep the accepted owner until answer() finishes it. markAsyncQuestion may run
+                    // just after resolveQuestion wakes this thread.
                 } else {
                     finishAsyncTask(task, result);
                 }
             } catch (Throwable t) {
                 task.future.completeExceptionally(t);
-                asyncTasksByTarget.remove(target, task);
             }
         });
         pruneTerminalTickets();
@@ -599,13 +614,14 @@ public final class MessageService {
     }
 
     /** Record the active question for an async ticket; blocking sends have no entry and stay unchanged. */
-    private void markAsyncQuestion(String target, String text, String turnId) {
-        Task task = asyncTasksByTarget.get(target);
+    private Task markAsyncQuestion(CompletableFuture<Rendezvous.Resolution> waiter, String text, String turnId) {
+        Task task = waiter == null ? null : asyncTasksByWaiter.get(waiter);
         if (task != null) {
             task.question = new Reply(Outcome.QUESTION, text, turnId);
             task.turnId = turnId;
             asyncTasksByTurn.put(turnId, task);
         }
+        return task;
     }
 
     /** Clear an answered or lapsed question, but only when it matches the ticket's current turn. */
@@ -623,7 +639,6 @@ public final class MessageService {
     /** Complete and detach an async ticket after its worker's actual terminal reply. */
     private void finishAsyncTask(Task task, Reply result) {
         task.future.complete(result);
-        asyncTasksByTarget.remove(task.target, task);
         if (task.turnId != null) {
             asyncTasksByTurn.remove(task.turnId, task);
         }

bridged/src/main/resources/logback.xml

@@ -1,4 +1,13 @@
 <configuration>
+    <!--
+      CB-575: MCP SDK 2.0.0 has no public notification registration API. It registers only
+      notifications/initialized and notifications/roots/list_changed, so other client notifications
+      still warn when unhandled. Clients may legitimately send notifications/cancelled; suppress only
+      that SDK WARN because it would devalue the action-needed WARN level used by M4 fleet health.
+      If a later SDK handles cancellation, this filter simply stops matching and can be removed.
+    -->
+    <turboFilter class="dev.ltms.bridged.logging.McpCancelledNotificationFilter"/>
+
     <appender name="STDOUT" class="ch.qos.logback.core.ConsoleAppender">
         <encoder>
             <pattern>%d{HH:mm:ss.SSS} %-5level [%thread] %logger{28} - %msg%n</pattern>

bridged/src/test/java/dev/ltms/bridged/health/FleetHealthMonitorTest.java

@@ -0,0 +1,81 @@
+package dev.ltms.bridged.health;
+
+import ch.qos.logback.classic.Logger;
+import ch.qos.logback.classic.spi.ILoggingEvent;
+import ch.qos.logback.core.read.ListAppender;
+import dev.ltms.bridged.herdr.AgentControl;
+import dev.ltms.bridged.herdr.FakeHerdr;
+import dev.ltms.bridged.inject.Injector;
+import dev.ltms.bridged.msg.InMemoryReplyInbox;
+import dev.ltms.bridged.msg.MessageService;
+import dev.ltms.bridged.msg.Rendezvous;
+import dev.ltms.bridged.session.SessionManager;
+import dev.ltms.bridged.member.ClaudeCodeLauncher;
+import dev.ltms.bridged.herdr.WorkspaceControl;
+import dev.ltms.bridged.guard.SubscriptionGuard;
+import dev.ltms.bridged.config.BridgedConfig;
+import org.junit.jupiter.api.Test;
+import org.slf4j.LoggerFactory;
+
+import java.util.Map;
+import java.util.Set;
+import java.util.concurrent.Executors;
+
+import static org.junit.jupiter.api.Assertions.assertEquals;
+
+class FleetHealthMonitorTest {
+    @Test void oneTickUsesOneFleetListForAnyRosterSize() {
+        FakeHerdr herdr = new FakeHerdr();
+        AgentControl agents = new AgentControl(herdr);
+        BridgedConfig.Profile profile = new BridgedConfig.Profile("test", "http://test:1", null,
+                null, null, null, null, null, null, null, null, null);
+        ClaudeCodeLauncher launcher = new ClaudeCodeLauncher(agents, new WorkspaceControl(herdr),
+                new SubscriptionGuard(Set.of("test")), Map.of("test", profile), "test", _ -> "token");
+        SessionManager sessions = new SessionManager(launcher);
+        sessions.acquire("test", null, null, null);
+        sessions.acquire("test", null, null, null);
+        herdr.calls.clear();
+        MessageService messages = new MessageService(agents, new Injector(agents), new Rendezvous(), new InMemoryReplyInbox());
+        var scheduler = Executors.newSingleThreadScheduledExecutor();
+        FleetHealthMonitor monitor = new FleetHealthMonitor(agents, sessions::roster, messages, scheduler, () -> 1, 60);
+        monitor.tick();
+        monitor.stop();
+        assertEquals(1, herdr.calls.stream().filter(call -> call.method().equals("agent.list")).count());
+    }
+
+    @Test void failedTickDoesNotStopTheNextTick() {
+        FakeHerdr herdr = new FakeHerdr().healthy(false);
+        AgentControl agents = new AgentControl(herdr);
+        var scheduler = Executors.newSingleThreadScheduledExecutor();
+        FleetHealthMonitor monitor = new FleetHealthMonitor(agents, java.util.List::of,
+                new MessageService(agents, new Injector(agents), new Rendezvous(), new InMemoryReplyInbox()),
+                scheduler, () -> 1, 60);
+        monitor.tick();
+        herdr.healthy(true);
+        monitor.tick();
+        monitor.stop();
+        assertEquals(2, herdr.calls.stream().filter(call -> call.method().equals("agent.list")).count());
+    }
+
+    @Test void faultTransitionLogsOnlyOnceUntilItChanges() {
+        Logger logger = (Logger) LoggerFactory.getLogger(FleetHealthMonitor.class);
+        ListAppender<ILoggingEvent> appender = new ListAppender<>();
+        appender.start();
+        logger.addAppender(appender);
+        try {
+            FakeHerdr herdr = new FakeHerdr();
+            AgentControl agents = new AgentControl(herdr);
+            var scheduler = Executors.newSingleThreadScheduledExecutor();
+            FleetHealthMonitor monitor = new FleetHealthMonitor(agents, java.util.List::of,
+                    new MessageService(agents, new Injector(agents), new Rendezvous(), new InMemoryReplyInbox()),
+                    scheduler, () -> 1, 60);
+            monitor.reportTransition("term_a", HealthState.TURN_BOUNDARY_LOST);
+            monitor.reportTransition("term_a", HealthState.TURN_BOUNDARY_LOST);
+            monitor.stop();
+            assertEquals(1, appender.list.stream().filter(event -> event.getFormattedMessage()
+                    .contains("member=term_a state=TURN_BOUNDARY_LOST")).count());
+        } finally {
+            logger.detachAppender(appender);
+        }
+    }
+}

bridged/src/test/java/dev/ltms/bridged/inject/InjectorTest.java

@@ -259,6 +259,7 @@ class InjectorTest {
     private static final class Captor implements TurnListener {
         final List<String> completed = new ArrayList<>();
         final List<String> failed = new ArrayList<>();
+        final List<String> failureReasons = new ArrayList<>();
 
         @Override
         public void onTurnComplete(String target) {
@@ -269,6 +270,12 @@ class InjectorTest {
         public void onTurnFailed(String target) {
             failed.add(target);
         }
+
+        @Override
+        public void onTurnFailed(String target, String reason) {
+            failed.add(target);
+            failureReasons.add(reason);
+        }
     }
 
     // ~30s of unknown at the 250ms prod poll interval; enough onStatus samples to trip the stall.
@@ -322,6 +329,23 @@ class InjectorTest {
         assertTrue(f.isCompletedExceptionally(), "queued waiters unblock when the worker vanishes");
     }
 
+    @Test
+    void dropPassesTheRealCauseForQueuedAndDeliveredWork() {
+        Captor cap = new Captor();
+        Injector inj = new Injector(new AgentControl(herdr), cap);
+        CompletableFuture<Void> delivered = inj.enqueue(T, "delivered");
+        CompletableFuture<Void> queued = inj.enqueue(T, "queued");
+
+        inj.onStatus(T, AgentStatus.IDLE);    // deliver the first message
+        inj.onStatus(T, AgentStatus.WORKING); // its turn is now in flight; one remains queued
+        inj.drop(T, new HerdrException("agent target sol not found", "agent_not_found", null));
+
+        assertEquals(List.of(T), cap.failed, "drop signals one turn failure for both affected states");
+        assertEquals(List.of("agent target sol not found"), cap.failureReasons);
+        assertTrue(delivered.isDone(), "the delivered future has already completed");
+        assertTrue(queued.isCompletedExceptionally(), "the queued future fails with the drop cause");
+    }
+
     @Test
     void dropFailsTheTurnOfADeliveredMessageWhenTheWorkerVanishes() {
         // CB-110: the message was delivered (no longer queued), so failing queued waiters alone would

bridged/src/test/java/dev/ltms/bridged/logging/McpCancelledNotificationFilterTest.java

@@ -0,0 +1,37 @@
+package dev.ltms.bridged.logging;
+
+import ch.qos.logback.classic.Level;
+import ch.qos.logback.classic.Logger;
+import ch.qos.logback.classic.spi.ILoggingEvent;
+import ch.qos.logback.core.read.ListAppender;
+import io.modelcontextprotocol.spec.McpSchema;
+import org.junit.jupiter.api.Test;
+import org.slf4j.LoggerFactory;
+
+import java.util.Map;
+
+import static org.junit.jupiter.api.Assertions.assertEquals;
+import static org.junit.jupiter.api.Assertions.assertTrue;
+
+class McpCancelledNotificationFilterTest {
+
+    @Test
+    void suppressesOnlyTheCancelledNotificationWarning() {
+        Logger logger = (Logger) LoggerFactory.getLogger(McpCancelledNotificationFilter.LOGGER);
+        ListAppender<ILoggingEvent> appender = new ListAppender<>();
+        appender.start();
+        logger.addAppender(appender);
+        try {
+            logger.warn(McpCancelledNotificationFilter.UNHANDLED_NOTIFICATION,
+                    new McpSchema.JSONRPCNotification("notifications/cancelled", Map.of("requestId", 7)));
+            logger.warn(McpCancelledNotificationFilter.UNHANDLED_NOTIFICATION,
+                    new McpSchema.JSONRPCNotification("notifications/progress", Map.of("progress", 1)));
+
+            assertEquals(1, appender.list.size());
+            assertEquals(Level.WARN, appender.list.getFirst().getLevel());
+            assertTrue(appender.list.getFirst().getFormattedMessage().contains("notifications/progress"));
+        } finally {
+            logger.detachAppender(appender);
+        }
+    }
+}

bridged/src/test/java/dev/ltms/bridged/mcp/BridgeMcpAuthzTest.java

@@ -72,7 +72,7 @@ class BridgeMcpAuthzTest {
                 new PrimaryRegistry(null),
                 enforce ? CallerResolver.withLeadsAndMembers(identity, false, null,
                 Map::of, new MemberRegistry(null)) : null,
-                metrics, BridgeMcp.CapacitySource.none());
+                metrics, BridgeMcp.CapacitySource.none(), new BridgeMcp.HealthCoverageSource(() -> "off"));
         return mcp;
     }
 

bridged/src/test/java/dev/ltms/bridged/mcp/BridgeMcpTest.java

@@ -474,7 +474,7 @@ class BridgeMcpTest {
         sessions.acquire("ltms-local", null, null, null);
         McpSchema.CallToolResult res = BridgeMcp.listFleet(workerService(h, "http://gx00.gw:8000", Set.of("gx00.gw")),
                 sessions, null, new BridgeMcp.CapacitySource(profile -> 2, profile -> 2,
-                () -> Set.of("ltms-local"), () -> 0), Map.of(), "");
+                () -> Set.of("ltms-local"), () -> 0), new BridgeMcp.HealthCoverageSource(() -> "off"), Map.of(), "");
         String out = textOf(res);
         assertTrue(out.contains("\"maxLoad\":2"), out);
         assertTrue(out.contains("\"live\":2"), out);
@@ -487,7 +487,7 @@ class BridgeMcpTest {
         SessionManager sessions = new SessionManager(workerService(h, "http://gx00.gw:8000", Set.of("gx00.gw")));
         String out = textOf(BridgeMcp.listFleet(workerService(h, "http://gx00.gw:8000", Set.of("gx00.gw")),
                 sessions, null, new BridgeMcp.CapacitySource(profile -> 0, profile -> 2,
-                () -> Set.of("terra"), () -> 0), Map.of(), ""));
+                () -> Set.of("terra"), () -> 0), new BridgeMcp.HealthCoverageSource(() -> "off"), Map.of(), ""));
         assertTrue(out.contains("\"profile\":\"terra\""), out);
         assertTrue(out.contains("\"live\":0"), out);
         assertTrue(out.contains("\"free\":2"), out);
@@ -499,7 +499,7 @@ class BridgeMcpTest {
         FakeHerdr h = new FakeHerdr();
         String out = textOf(BridgeMcp.listFleet(workerService(h, "http://gx00.gw:8000", Set.of("gx00.gw")),
                 new SessionManager(workerService(h, "http://gx00.gw:8000", Set.of("gx00.gw"))), null,
-                BridgeMcp.CapacitySource.none(), Map.of(), ""));
+                BridgeMcp.CapacitySource.none(), new BridgeMcp.HealthCoverageSource(() -> "off"), Map.of(), ""));
         assertFalse(out.contains("\"capacity\":"), out);
     }
 

bridged/src/test/java/dev/ltms/bridged/msg/MessageServiceTest.java

@@ -120,6 +120,36 @@ class MessageServiceTest {
         assertFalse(reply.completed());
     }
 
+    @Test
+    void droppedQueuedAndDeliveredTurnsExposeTheRealCauseExactlyOnce() throws Exception {
+        CompletableFuture<MessageService.Reply> first = sendAsync();
+        awaitWaiting();
+        injector.onStatus(T, AgentStatus.IDLE);    // first delivery
+        injector.onStatus(T, AgentStatus.WORKING); // first turn in flight
+
+        CompletableFuture<Void> queued = injector.enqueue(T, "second task");
+        CompletableFuture<Rendezvous.Resolution> waiter = rendezvous.currentWaiter(T);
+        injector.drop(T, new HerdrException("agent target sol not found", "agent_not_found", null));
+
+        MessageService.Reply reply = first.get(5, TimeUnit.SECONDS);
+        assertEquals(MessageService.Outcome.WORKER_FAILED, reply.outcome());
+        assertEquals("agent target sol not found", reply.text());
+        assertTrue(queued.isCompletedExceptionally(), "the queued delivery future also fails");
+        assertFalse(rendezvous.resolveFailure(waiter, "second failure"), "the waiter fails exactly once");
+    }
+
+    @Test
+    void noDropReasonKeepsTheExistingFallbackText() throws Exception {
+        herdr.readText("");
+        CompletableFuture<Rendezvous.Resolution> waiter = rendezvous.open(T);
+
+        completion.onTurnFailed(T);
+
+        Rendezvous.Resolution resolution = waiter.get(2, TimeUnit.SECONDS);
+        assertEquals("worker did not reply; its turn ended in an unrecoverable state "
+                + "(worker unreachable or stuck)", resolution.text());
+    }
+
     // --- bridge_ask reverse rendezvous (CB-205) ------------------------------------------------
 
     @Test
@@ -609,4 +639,94 @@ class MessageServiceTest {
         assertTrue(view.detail() != null && view.detail().contains("released"),
                 "and the detail says why, rather than 'worker unknown'");
     }
+
+    @Test
+    void abandonFailsEveryPendingAsyncTicketForTheReleasedTarget() throws Exception {
+        String first = messages.sendAsync(T, "first task");
+        awaitWaiting(); // first task owns the target lock and rendezvous waiter
+        String second = messages.sendAsync(T, "second task"); // parked on the same lock, not yet queued
+        String third = messages.sendAsync(T, "third task"); // a second queued ticket proves the full sweep
+
+        assertTrue(messages.abandon(T, "agent target term_a not found"));
+
+        assertFailedTicket(first, "agent target term_a not found");
+        assertFailedTicket(second, "agent target term_a not found");
+        assertFailedTicket(third, "agent target term_a not found");
+    }
+
+    @Test
+    void abandonDoesNotFailAnAsyncTicketWaitingForAnAnswer() throws Exception {
+        String ticket = messages.sendAsync(T, "task that asks");
+        awaitWaiting();
+        injectDelivery();
+
+        CompletableFuture<MessageService.AskResult> ask =
+                CompletableFuture.supplyAsync(() -> messages.ask(T, "which config?", 5000));
+        MessageService.TaskView asking = awaitTicketPhase(ticket, MessageService.Phase.ASKING);
+
+        assertFalse(messages.abandon(T, "agent target term_a not found"),
+                "an asking ticket is an active turn, not a pending send to sweep");
+        assertEquals(MessageService.Phase.ASKING, messages.poll(ticket).phase());
+
+        CompletableFuture<MessageService.Reply> answer = CompletableFuture.supplyAsync(
+                () -> messages.answer(asking.turnId(), "config.yaml", 5000));
+        assertEquals("config.yaml", ask.get(5, TimeUnit.SECONDS).answer());
+        awaitWaiting();
+        assertTrue(rendezvous.resolve(T, "done"));
+        assertEquals(MessageService.Outcome.REPLIED, answer.get(5, TimeUnit.SECONDS).outcome());
+    }
+
+    @Test
+    void unansweredAsyncQuestionReturnsTheTicketToPendingAndReleasesItsTarget() throws Exception {
+        String ticket = messages.sendAsync(T, "task that asks");
+        awaitWaiting();
+        injectDelivery();
+
+        assertEquals(MessageService.AskOutcome.TIMED_OUT,
+                messages.ask(T, "which config?", 200).outcome());
+        assertEquals(MessageService.Phase.PENDING, messages.poll(ticket).phase(),
+                "only the question wait ended; the delegated turn may still finish");
+
+        String next = messages.sendAsync(T, "next task");
+        awaitWaiting();
+        assertTrue(rendezvous.resolve(T, "done"));
+        assertEquals(MessageService.Phase.DONE, awaitTicketPhase(next, MessageService.Phase.DONE).phase());
+    }
+
+    @Test
+    void asyncQuestionBelongsToTheTaskThatOwnsItsForwardWaiter() throws Exception {
+        String first = messages.sendAsync(T, "first task");
+        awaitWaiting();
+
+        CompletableFuture<MessageService.AskResult> ask =
+                CompletableFuture.supplyAsync(() -> messages.ask(T, "which config?", 5000));
+        assertEquals(MessageService.Phase.ASKING, awaitTicketPhase(first, MessageService.Phase.ASKING).phase());
+
+        MessageService.TaskView asking = messages.poll(first);
+        CompletableFuture<MessageService.Reply> answer = CompletableFuture.supplyAsync(
+                () -> messages.answer(asking.turnId(), "config.yaml", 5000));
+        assertEquals("config.yaml", ask.get(5, TimeUnit.SECONDS).answer());
+        awaitWaiting();
+        assertTrue(rendezvous.resolve(T, "done"));
+        assertEquals(MessageService.Outcome.REPLIED, answer.get(5, TimeUnit.SECONDS).outcome());
+    }
+
+    private void assertFailedTicket(String ticket, String reason) throws Exception {
+        MessageService.TaskView view = awaitTicketPhase(ticket, MessageService.Phase.FAILED);
+        assertEquals(reason, view.detail());
+    }
+
+    private MessageService.TaskView awaitTicketPhase(String ticket, MessageService.Phase phase) throws Exception {
+        long deadline = System.currentTimeMillis() + 3000;
+        MessageService.TaskView view;
+        do {
+            view = messages.poll(ticket);
+            if (view.phase() == phase) {
+                return view;
+            }
+            Thread.sleep(5);
+        } while (System.currentTimeMillis() < deadline);
+        assertEquals(phase, view.phase());
+        return view;
+    }
 }

bridged/src/test/resources/logback-test.xml

@@ -16,6 +16,9 @@
           AuditLogTest, which attaches its own ListAppender and asserts on emitted records.
     -->
 
+    <!-- Keep the test logger behaviour aligned with the production cancellation filter. -->
+    <turboFilter class="dev.ltms.bridged.logging.McpCancelledNotificationFilter"/>
+
     <appender name="STDOUT" class="ch.qos.logback.core.ConsoleAppender">
         <encoder>
             <pattern>%d{HH:mm:ss.SSS} [%thread] %-5level %logger{36} - %msg%n</pattern>
@@ -34,4 +37,4 @@
         <appender-ref ref="STDOUT"/>
     </root>
 
-</configuration>
+</configuration>

docs/M4-Fleet-Health.md

@@ -744,8 +744,28 @@ worktree discovery.
 
 Acceptance criteria:
 
-1. Every accepted send receives a stable `TurnToken` tied to target, exact waiter, session turn,
-   delivery baseline, and task outcome.
+1. Every accepted send receives a stable `TurnToken` tied to target, exact waiter, and delivery
+   baseline.
+
+   **Corrected during implementation (2026-08-15).** This criterion first also required the session
+   turn number and the task outcome. That is not implementable at this layer, and the implementer
+   refused it three times rather than fabricate a value — correctly. The reason is an ordering fact
+   that is invisible from any single class: `MessageService` owns acceptance and holds the waiter and
+   the async `Task`, but it learns nothing about delivery, because the delivery event goes to
+   `CompletionResolver` through `TurnListener.onDelivered`. And `CompletionResolver.onDelivered` runs
+   *before* `SessionManager.onDelivered`, so the session turn number does not exist yet at the only
+   point where the token could capture it.
+
+   Two ways out were rejected. A shared registry keyed by target reintroduces exactly the "whichever
+   send happens to be waiting" ambiguity the token exists to remove — the same weak claim
+   `Rendezvous.currentWaiter` warns about. Injecting a turn counter into `MessageService` adds a
+   required cross-layer dependency to populate a field that nothing in this slice reads, which is
+   speculative coupling across a boundary already shown to be fragile.
+
+   So the token identifies the **accepted send**, and `SessionManager` keeps verifying its own
+   delivery separately. Repair (criterion 2) does need the session turn; binding it means resolving
+   that acceptance-versus-delivery ordering first, and that work belongs to the repair unit, not
+   here. The token record carries a comment saying the field is deliberately absent.
 2. Repair requires the same `BUSY` token, two raw `IDLE` or `DONE` snapshots, no conflicting
    observation, exact open waiter, successful baseline, and new recognised assistant output.
 3. Missing, failed, late, or post-restart baseline never authorises repair.