Classify named AMQP recovery errors

Merge remote-tracking branch 'origin/main' into worker/errscan-bed2ca-2
Merge #356 : name the AMQP connection in its own log lines
2026-09-05 06:05:52 +07:00 · 2026-09-05 06:02:29 +07:00 · 2026-09-05 06:01:30 +07:00 · 2026-09-05 05:59:22 +07:00 · 2026-09-05 05:55:22 +07:00 · 2026-09-05 05:53:44 +07:00
22 changed files with 2515 additions and 191 deletions
@@ -27,14 +27,14 @@ import java.util.function.Supplier;
 *
 * <ul>
 *   <li><strong>Hot</strong> — re-read per use, so a reload takes effect on the next spawn:
- *       {@code fleet:} (every role pool, {@code charters}, and {@code tabLabel}),
- *       {@code placement:}, and an existing profile's {@code weight} / {@code maxLoad}. Those
- *       three are read through a supplier on {@code CompositePeerLauncher}, which is what makes
- *       them hot — not the fact that they are config. <strong>This does NOT include
- *       {@code fleet.leaders}</strong>: {@code Fleetd.main} reads {@code cfg.fleet().leaders()}
- *       once at startup to build the {@code LeadTabScanner} and the {@code LeadLauncher}, and
- *       neither is reconstructed on reload — so a lead added, removed, or re-{@code tab}'d under
- *       {@code fleet.leaders} needs a restart, the same as any deferred key below.</li>
+ *       {@code placement:}, and an existing profile's {@code weight} / {@code maxLoad}. Both are
+ *       read through a supplier on {@code CompositePeerLauncher}, which is what makes them hot —
+ *       not the fact that they are config. Most of {@code fleet:} — every role pool
+ *       ({@code architects}/{@code developers}/{@code reviewers}), {@code charters}, and
+ *       {@code tabLabel} — is read the same live way, through the same supplier
+ *       ({@code () -> config.get().fleet()}). <strong>But {@code fleet:} as a whole is NOT in this
+ *       class</strong>: {@code fleet.leaders} inside the same key is frozen, which is exactly what
+ *       makes {@code fleet:} split rather than hot — see below.</li>
 *   <li><strong>Deferred</strong> — accepted into the new snapshot, but the wiring built at startup
 *       keeps the old value until a restart: {@code lifecycle:}, {@code leadHeartbeat:},
 *       {@code spawnReadyTimeoutMs} / {@code spawnReadyPollMs}, {@code quarantineCooldownSeconds}
@@ -71,12 +71,12 @@ import java.util.function.Supplier;
 *       {@code HerdrPeerLauncher} takes {@code Map.copyOf(profiles)} at construction and resolves
 *       each spawn out of that copy, so those never reach a launch until the daemon restarts. A
 *       reload logs these rather than pretending they applied.</li>
- *   <li><strong>Split</strong> (fleetd #330) — read <em>both</em> ways at different sites, so the
- *       key does not fit any class above as a whole: {@code health:} and {@code coordinator:}.
- *       Each is read off the startup snapshot to build a long-lived object, and read live off
- *       {@link #get()} at a different, unrelated site — so half of a reload's effect already
- *       applies while the other half waits for a restart, and a bare "config reloaded" would
- *       under-claim by exactly that half.
+ *   <li><strong>Split</strong> (fleetd #330; extended to a third key by fleetd #333) — read
+ *       <em>both</em> ways at different sites, so the key does not fit any class above as a whole:
+ *       {@code health:}, {@code coordinator:} and {@code fleet:}. Each is read off the startup
+ *       snapshot to build a long-lived object, and read live off {@link #get()} at a different,
+ *       unrelated site — so half of a reload's effect already applies while the other half waits
+ *       for a restart, and a bare "config reloaded" would under-claim by exactly that half.
 *       <ul>
 *         <li>{@code health:} — the monitor itself ({@code enabled}, {@code intervalSeconds},
 *             {@code workingSuspectAfterSeconds}) is built once at {@code Fleetd.java:556-563}
@@ -92,34 +92,75 @@ import java.util.function.Supplier;
 *             URI env-var <em>name</em> that {@code MemberEnvAllowList} keeps out of a member's
 *             environment is read live off {@link #get()} on every spawn
 *             ({@code HerdrPeerLauncher.java:1530}), so it already applies.</li>
+ *         <li>{@code fleet:} (fleetd #333) — {@code fleet.leaders} is the frozen half:
+ *             {@code Fleetd.java:281} reads {@code cfg.fleet().leaders()} off the startup snapshot
+ *             for two long-lived objects built right after it and never rebuilt — the
+ *             {@code LeadTabScanner}'s {@code tab label → lead name} map ({@code Fleetd.java:301},
+ *             wired into {@code CallerResolver.withLeadsAndMembers} at {@code Fleetd.java:620/624},
+ *             which is how a caller's pane is recognised as a lead at all) and, when herdr answered
+ *             at startup, {@code LeadLauncher(...).ensureLeads()} ({@code Fleetd.java:315}), which
+ *             auto-launches each declared lead up to its {@code instances} count. So a lead added,
+ *             removed, or given a new {@code tab:} label under {@code fleet.leaders} needs a
+ *             restart — until then it is invisible to identity resolution, and this is exactly the
+ *             scenario fleetd #333 named: an operator edits a lead's {@code tab:} to match a
+ *             renamed pane, sees "config reloaded", and the pane keeps resolving as a worker,
+ *             because {@code CallerResolver} is still matching against the old label. The live
+ *             half is the rest of {@code fleet:} — {@code architects}/{@code developers}/
+ *             {@code reviewers}, {@code charters}, {@code tabLabel} — read live through the same
+ *             {@code CompositePeerLauncher} supplier the Hot bullet above names, so a reload that
+ *             only touches those already applies with nothing to report. Because the hot and frozen
+ *             halves of {@code fleet:} are disjoint sub-fields rather than the same fields read two
+ *             ways (contrast {@code coordinator.uriEnv} above), {@link #changedSplitKeys} compares
+ *             {@code fleet.leaders} alone, not the whole {@code Fleet} record — comparing the whole
+ *             record would report "split" for a {@code tabLabel}-only change that is actually fully
+ *             hot, over-claiming in exactly the direction this class exists to avoid under-claiming
+ *             in.</li>
 *       </ul>
 *       A split change is still accepted — {@link Outcome#applied()} stays {@code true}, the same
 *       as a deferred change — because the live half genuinely took effect; refusing the whole
 *       reload would leave the operator worse off than today. {@link Outcome#split()} names the
 *       key and says which half is which each time, rather than trying to score "how changed" a
 *       mixed key is or handle "both halves changed in one reload" as a special case.</li>
- *   <li><strong>Cold</strong> — cannot change at all under a running daemon: {@code bind:},
- *       {@code herdrSocket:}, {@code broker:} and {@code auth:}. The socket is bound, the broker
+ *   <li><strong>Cold</strong> — cannot change at all under a running daemon. All five of
+ *       {@link #COLD_KEYS}: {@code bind:}, {@code herdrSocket:}, {@code memberHerdrSocket:},
+ *       {@code broker:} and {@code auth:}. The sockets are already connected, the broker
 *       connection is open, and the auth mode decides who may reach the port that is already
- *       listening.</li>
+ *       listening. This bullet omitted {@code memberHerdrSocket:} until fleetd #333 — say "all
+ *       five of COLD_KEYS" rather than re-listing them, so prose and set cannot drift again.</li>
 * </ul>
 *
- * <p><strong>The denominator, measured on 2026-09-04 (fleetd #330).</strong> {@code FleetConfig} has
- * 22 top-level record components. Two of them are named nowhere in this file, and the reason is the
- * same for both: {@code memberCredentials} and {@code memberLoginShell} are <strong>hot</strong> and
- * correctly absent — both are read live off {@code config.get()} at spawn time
- * ({@code Fleetd.java:198, 205, 729} and {@code HerdrPeerLauncher#configuredMemberLoginShell}), so a
- * reload takes effect on the next spawn with no entry needed here.
+ * <p><strong>The denominator, measured on 2026-09-04 (fleetd #330; recounted for fleetd #333).</strong>
+ * {@code FleetConfig} has 22 top-level record components: 5 cold, 11 deferred, 3 split, 3
+ * hot-excluded. Three of them are named nowhere in this file, and the reason is the same for all
+ * three: {@code placement}, {@code memberCredentials} and {@code memberLoginShell} are
+ * <strong>hot</strong> and correctly absent — all three are read live off {@code config.get()}
+ * (placement through the {@code CompositePeerLauncher} supplier the Hot bullet names;
+ * {@code memberCredentials}/{@code memberLoginShell} at spawn time, {@code Fleetd.java:198, 205, 729}
+ * and {@code HerdrPeerLauncher#configuredMemberLoginShell}), so a reload takes effect on the next
+ * spawn with no entry needed here.
 * {@code health} and {@code coordinator} used to be a third kind — <strong>undecided</strong>, not
 * hot — until fleetd #330 added the <strong>split</strong> class above and gave them a home. A
 * reload touching either used to report a bare "config reloaded", which under-claimed; now it names
- * the key and says which half is which.
+ * the key and says which half is which. {@code fleet} was the same story in reverse: fleetd #330's
+ * own fact-find named {@code health}/{@code coordinator} as "the complete set of split-shaped keys"
+ * and filed {@code fleet.leaders}'s restart requirement as a documented caveat sitting in the
+ * <strong>hot-excluded</strong> escape hatch instead — correctly documented, but in the one bucket
+ * this file's own coverage test cannot check the truth of (see that test's javadoc). fleetd #333
+ * moved it into <strong>split</strong>, where {@link #changedSplitKeys} actually reports it.
 * <p>The point of writing the count down: "not mentioned in this file" looks identical for a key
- * that is correctly hot and for a key nobody triaged. Twice now — {@code worktreeGroup} (#323) and
- * {@code primary}/{@code configReload} (#326) — the second kind hid among the first. A top-level
- * coverage checker in the {@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 is correctly hot and for a key nobody triaged. Three times now — {@code worktreeGroup} (#323),
+ * {@code primary}/{@code configReload} (#326), and {@code fleet.leaders} sitting in the escape hatch
+ * (#333) — the second kind hid among the first. A top-level coverage checker in the
+ * {@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}/{@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
@@ -153,7 +194,26 @@ public final class ConfigRef implements Supplier<FleetConfig> {
     * {@code true}) and reported by name, with a message naming which half is live and which needs
     * a restart.
     */
-    static final Set<String> SPLIT_KEYS = Set.of("health", "coordinator");
+    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;
@@ -283,8 +343,14 @@ public final class ConfigRef implements Supplier<FleetConfig> {
        return out;
    }

-    /** Cold keys whose value differs between the running config and the candidate. */
-    private static List<String> changedColdKeys(FleetConfig old, FleetConfig fresh) {
+    /**
+     * Cold keys whose value differs between the running config and the candidate.
+     *
+     * <p>Package-private (not {@code private}) so {@code ConfigRefTopLevelReportingCoverageTest}
+     * can call it directly with a reflection-built {@code FleetConfig} pair, the same reason
+     * {@link #sameLaunchSettings} is package-private — see that test's class doc (fleetd #333).
+     */
+    static List<String> changedColdKeys(FleetConfig old, FleetConfig fresh) {
        List<String> changed = new ArrayList<>();
        if (!Objects.equals(old.bind(), fresh.bind())) {
            changed.add("bind");
@@ -306,8 +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");
@@ -389,12 +463,24 @@ public final class ConfigRef implements Supplier<FleetConfig> {

    /**
     * Split keys whose value differs between the running config and the candidate — see the class
-     * doc's Split bullet (fleetd #330). Unlike {@link #changedDeferredKeys}, this does not try to
-     * tell which sub-field moved: any change to {@code health:} or {@code coordinator:} gets the
-     * same fixed message, because the message already names both halves every time, so there is no
-     * "which half changed" question left for the caller to answer.
+     * doc's Split bullet (fleetd #330, extended for {@code fleet:} by fleetd #333). Unlike
+     * {@link #changedDeferredKeys}, this does not try to tell which sub-field moved for {@code
+     * health:} or {@code coordinator:}: any change to either gets the same fixed message, because
+     * the message already names both halves every time, so there is no "which half changed"
+     * question left for the caller to answer. {@code fleet:} is different on purpose — see below.
+     *
+     * <p>Package-private (not {@code private}) so {@code ConfigRefTopLevelReportingCoverageTest}
+     * can call it directly with a reflection-built {@code FleetConfig} pair, the same reason
+     * {@link #sameLaunchSettings} is package-private — see that test's class doc (fleetd #333). That
+     * test exists because membership in {@link #SPLIT_KEYS} proves nothing about this method on its
+     * own: fleetd #333 measured that dropping the {@code coordinator} branch out of this method
+     * while leaving {@code "coordinator"} in {@code SPLIT_KEYS} left the whole suite green except a
+     * hand-written {@code ConfigRefTest} case — neither {@code ConfigRefTopLevelCoverageTest} (it
+     * only reads the set) nor the "kept in step" assert below (it only checks the reported keys are
+     * a SUBSET of {@code SPLIT_KEYS}, never that every {@code SPLIT_KEYS} member has a branch here)
+     * would have caught it.
     */
-    private static List<String> changedSplitKeys(FleetConfig old, FleetConfig fresh) {
+    static List<String> changedSplitKeys(FleetConfig old, FleetConfig fresh) {
        List<String> changed = new ArrayList<>();
        if (!Objects.equals(old.health(), fresh.health())) {
            changed.add("health: the monitor itself (enabled, interval, workingSuspectAfter) is "
@@ -406,13 +492,42 @@ public final class ConfigRef implements Supplier<FleetConfig> {
                    + "opened once and needs a restart; the broker URI env-var name kept out of a "
                    + "member's environment is read live on every spawn and already applied");
        }
+        // fleetd #333: unlike health/coordinator above, most of `fleet:` (architects, developers,
+        // reviewers, charters, tabLabel) is genuinely hot — ConfigRefTest.aHotChangeIsAppliedAndRead-
+        // ThroughGet and aCharterChangeIsHotAndReachesTheLiveConfig prove it reaches the live config
+        // with no restart note. Only fleet.leaders is frozen (Fleetd.java:281 reads
+        // cfg.fleet().leaders() off the startup snapshot to build both the LeadTabScanner's
+        // tab-label-to-name map, wired into CallerResolver.withLeadsAndMembers at Fleetd.java:620/624,
+        // and — when herdr answered — LeadLauncher(...).ensureLeads() at Fleetd.java:315, which
+        // auto-launches each lead up to its `instances` count; neither is rebuilt on reload). So this
+        // compares fleet.leaders alone, not the whole Fleet record: comparing the whole record would
+        // report "split" for a tabLabel-only or charters-only change that is actually fully hot,
+        // which is the over-claim mirror of the under-claim bug this class exists to prevent.
+        if (!Objects.equals(leadersOf(old), leadersOf(fresh))) {
+            changed.add("fleet: fleet.leaders (each lead's tab, workspace, cwd, profile and "
+                    + "instances count) is read once at startup to build the LeadTabScanner's "
+                    + "identity map and to auto-launch leads, and neither is rebuilt on reload, so a "
+                    + "lead added, removed, or given a new tab: label needs a restart — until then it "
+                    + "stays unrecognised, and a caller from its new tab resolves as a worker, not a "
+                    + "lead; the rest of fleet: (architects, developers, reviewers, charters, "
+                    + "tabLabel) is read live through the supplier on CompositePeerLauncher and "
+                    + "already applied");
+        }
        // Kept in step with SPLIT_KEYS the same way changedColdKeys is kept in step with COLD_KEYS —
-        // every message here must be traceable to one of the two split keys the class doc documents.
+        // every message here must be traceable to one of the split keys the class doc documents.
+        // NOTE what this does NOT prove, per the javadoc above: it does not catch a SPLIT_KEYS
+        // member with no branch above at all, only a branch whose message is mis-worded relative to
+        // the set. ConfigRefTopLevelReportingCoverageTest is what proves the former.
        assert changed.stream().allMatch(m -> SPLIT_KEYS.stream().anyMatch(k -> m.startsWith(k + ":")))
                : "a split entry was reported that does not start with a SPLIT_KEYS name: " + changed;
        return changed;
    }

+    /** {@code cfg.fleet().leaders()}, defensively, in case a caller hands in a non-defaulted config. */
+    private static Map<String, FleetConfig.Leader> leadersOf(FleetConfig cfg) {
+        return cfg.fleet() == null ? Map.of() : cfg.fleet().leaders();
+    }
+
    /**
     * {@link FleetConfig.Profile} record components deliberately left out of
     * {@link #sameLaunchSettings} because they are read <em>live</em>, not baked in at spawn — see
@@ -1,10 +1,12 @@
 package dev.ltms.fleet.inject;

 /**
- * Notified when {@link CompletionResolver} actually delivers a typed backend-error classification
- * to a waiting send (fleetd#201 / #227) — never on a race that lost. {@link CompletionResolver}
- * calls this only after {@code Rendezvous.resolveFailure} returns {@code true} for that exact
- * waiter, mirroring the win-only race rule {@link ExhaustionSink} already uses.
+ * Notified when {@link CompletionResolver} has a backend-error match at the start of a pane line,
+ * or both a match and its too-fast crash signature, for a waiting send (fleetd#201 / #227). A text
+ * match inside ordinary pane prose can be a member's report about an error, so it fails the send
+ * without notifying this sink.
+ * {@link CompletionResolver} calls this only after {@code Rendezvous.resolveFailure} returns
+ * {@code true} for that exact waiter, mirroring the win-only race rule {@link ExhaustionSink} uses.
 *
 * <p>The public send result is unchanged by this classification — it is still a failed send
 * ({@code Rendezvous.Kind#FAILED}); this sink is the internal seam a later stage (fleetd#201 Unit
@@ -380,16 +380,19 @@ public final class CompletionResolver implements TurnListener {
                        + "matched the profile's exhausted pattern): {}", target, reason);
                // CB-578 stage B: only on the resolution that actually won the race — a late
                // duplicate must never quarantine a credential twice for one refusal.
-                exhaustionSink.onExhausted(target, reason);
+                if (startsWithExhaustion(matchedLine, exhausted)) {
+                    exhaustionSink.onExhausted(target, reason);
+                }
            }
            return;
        }
        // fleetd#164 (part 2) / fleetd#201: a scrape that read cleanly and produced content still
        // isn't a real reply when that content is the backend's own rejection (e.g. an HTTP 400
        // before the worker did any work). Classify it as a failure naming the member, rather than
-        // handing the caller a scrape that reads like a completed answer, and — only on the
-        // resolution that actually wins the race, mirroring the exhaustion sink above — notify the
-        // typed backend-error sink so a later stage can act on repeated failures.
+        // handing the caller a scrape that reads like a completed answer. A text match alone is not
+        // enough to notify the typed backend-error sink: this assistant block can be a member's
+        // normal prose about an error. A line that starts with the error match is stronger evidence;
+        // the too-fast path below also has its crash signature before it records a credential failure.
        String backendError = firstMatchingLine(assistantBlock, backendErrorPatternOrFallback(target));
        if (backendError != null) {
            // Carry the whole scrape, not just the matched line. The pattern is a heuristic: a member
@@ -401,9 +404,9 @@ public final class CompletionResolver implements TurnListener {
            if (rendezvous.resolveFailure(waiter, reason)) {
                inFlight.remove(target, turn);
                log.warn("failing send to {} via turn-stall fallback: {}", target, reason);
-                // fleetd#201 Unit 1: only on the resolution that actually won the race — a late
-                // duplicate must never double-count one backend failure.
-                backendErrorSink.onBackendError(target, backendError, reason);
+                if (startsWithBackendError(backendError, backendErrorPatternOrFallback(target))) {
+                    backendErrorSink.onBackendError(target, backendError, reason);
+                }
            }
            return;
        }
@@ -477,7 +480,9 @@ public final class CompletionResolver implements TurnListener {
                        + "usable assistant block; no fleet_reply): {}", target, reason);
                // CB-578 stage B: only on the resolution that actually won the race — a late
                // duplicate must never quarantine a credential twice for one refusal.
-                exhaustionSink.onExhausted(target, reason);
+                if (startsWithExhaustion(matchedLine, exhausted)) {
+                    exhaustionSink.onExhausted(target, reason);
+                }
            }
            return true;
        }
@@ -492,8 +497,9 @@ public final class CompletionResolver implements TurnListener {
            if (rendezvous.resolveFailure(waiter, reason)) {
                inFlight.remove(target, turn);
                log.warn("failing send to {} via turn-stall fallback from the raw scrape: {}", target, reason);
-                // fleetd#201 Unit 1: only on the resolution that actually won the race.
-                backendErrorSink.onBackendError(target, backendError, reason);
+                if (startsWithBackendError(backendError, backendErrorPatternOrFallback(target))) {
+                    backendErrorSink.onBackendError(target, backendError, reason);
+                }
            }
            return true;
        }
@@ -540,10 +546,10 @@ public final class CompletionResolver implements TurnListener {
     * {@link #MIN_TURN_NANOS} — a crash signature (e.g. a backend HTTP 400 before the worker did
     * anything) that a bare {@code BUSY -> DONE} transition cannot be told apart from a genuinely
     * fast completion. Runs the same backend-error classification the normal and raw-scrape paths
-     * apply, against whatever is on screen right now: a match is a typed failure that notifies
-     * {@link #backendErrorSink} (only on the resolution that wins the race); a non-match stays the
-     * original generic too-fast failure, naming the member and both timings, with whatever the pane
-     * shows appended so the caller sees the cause, not just "it failed".
+     * apply, against whatever is on screen right now: a match together with the too-fast crash
+     * signature notifies {@link #backendErrorSink} (only on the resolution that wins the race). A
+     * non-match stays the original generic too-fast failure, naming the member and both timings,
+     * with whatever the pane shows appended so the caller sees the cause, not just "it failed".
     */
    private void failTooFast(String target, InFlight turn, CompletableFuture<Rendezvous.Resolution> waiter,
                              long elapsedNanos) {
@@ -611,6 +617,68 @@ public final class CompletionResolver implements TurnListener {
        return null;
    }

+    /**
+     * True when nothing before the match on this pane line ends a sentence — that is, the match is
+     * still inside the line's first sentence rather than inside prose a member wrote about it.
+     * Used to decide whether an exhaustion match may quarantine a credential (fleetd #348).
+     *
+     * <p><strong>Why this is looser than {@link #startsWithBackendError}.</strong> An
+     * {@code exhaustedPattern} is written per profile and may name only the decisive words of a
+     * provider message — {@code "usage limit has been reached"} without its leading {@code "The"}.
+     * A start-of-line check would then reject the genuine refusal. That is the false negative
+     * fleetd #348's invariant 1 calls the worse direction: an unrecorded exhaustion leaves the
+     * fleet spawning into a credential with no capacity, and a quarantine runs 1800s against the
+     * backend-error cooldown's fixed 60s.
+     *
+     * <p>This rule accepts a superset of what a start-of-line check accepts: if the match begins
+     * right after the chrome, there is nothing in front of it, so there is no sentence ending
+     * either. So moving to it cannot add a false negative.
+     *
+     * <p><strong>No chrome skipping here, deliberately.</strong> The first version of this method
+     * copied {@code startsWithBackendError}'s leading-chrome loop. Measured on merge: deleting that
+     * loop left all 1369 tests green, and it must — the scan only looks for {@code . ! ?}, and no
+     * terminal chrome character is one of those. A step that cannot change the result is worse than
+     * no step, because the next reader takes it as evidence that chrome was handled.
+     *
+     * <p>It stays a heuristic. Prose whose <em>first</em> sentence carries the pattern still
+     * notifies the sink, and a genuine refusal behind an earlier full stop (a hostname, a version
+     * number) still does not. Both are known and neither is fixed here.
+     */
+    private static boolean startsWithExhaustion(String line, Pattern pattern) {
+        var matcher = pattern.matcher(line);
+        if (!matcher.find()) {
+            return false;
+        }
+        for (int prefix = 0; prefix < matcher.start(); prefix++) {
+            if (".!?".indexOf(line.charAt(prefix)) >= 0) {
+                return false;
+            }
+        }
+        return true;
+    }
+
+    /**
+     * 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) {
+        int i = 0;
+        while (i < line.length() && !Character.isLetterOrDigit(line.charAt(i))) {
+            i++;
+        }
+        return pattern.matcher(line.substring(i)).lookingAt();
+    }
+
    /**
     * Coverage summary for the CB-578 stage A exhausted-pattern classification, logged at startup
     * the way {@link dev.ltms.fleet.health.FleetHealthMonitor#coverage} is — so an operator can
@@ -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;
@@ -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);
        }
    }

@@ -8,6 +8,7 @@ import com.rabbitmq.client.DeliverCallback;
 import com.rabbitmq.client.Recoverable;
 import com.rabbitmq.client.RecoveryListener;
 import com.rabbitmq.client.Return;
+import com.rabbitmq.client.impl.DefaultExceptionHandler;
 import org.slf4j.Logger;
 import org.slf4j.LoggerFactory;

@@ -153,17 +154,22 @@ public final class AmqpReplyInbox implements ReplyInbox, AutoCloseable {
    /** As {@link #open(String)}, with an explicit consumer prefetch (CB-527: caps the held backlog per target). */
    public static AmqpReplyInbox open(String uri, int prefetch) {
        try {
-            ConnectionFactory factory = new ConnectionFactory();
-            factory.setUri(uri);
-            // Self-heal transient blips; topology recovery re-declares queues and re-attaches consumers.
-            factory.setAutomaticRecoveryEnabled(true);
-            factory.setTopologyRecoveryEnabled(true);
-            return new AmqpReplyInbox(factory.newConnection("fleetd-reply-inbox"), prefetch);
+            return new AmqpReplyInbox(connectionFactory(uri).newConnection(AmqpConnectionFailureLogger.REPLY_INBOX), prefetch);
        } catch (Exception e) {
            throw new IllegalStateException("cannot connect to AMQP broker at " + uri, e);
        }
    }

+    static ConnectionFactory connectionFactory(String uri) throws Exception {
+        ConnectionFactory factory = new ConnectionFactory();
+        factory.setUri(uri);
+        // Self-heal transient blips; topology recovery re-declares queues and re-attaches consumers.
+        factory.setAutomaticRecoveryEnabled(true);
+        factory.setTopologyRecoveryEnabled(true);
+        factory.setExceptionHandler(new AmqpConnectionFailureLogger(AmqpConnectionFailureLogger.REPLY_INBOX, log));
+        return factory;
+    }
+
    /** Wrap an already-open connection with {@link #DEFAULT_PREFETCH} (injection seam for the contract test). */
    AmqpReplyInbox(Connection connection) {
        this(connection, DEFAULT_PREFETCH);
@@ -571,3 +577,44 @@ public final class AmqpReplyInbox implements ReplyInbox, AutoCloseable {
        }
    }
 }
+
+/**
+ * Keeps RabbitMQ's forgiving exception behaviour while adding the connection identity that its
+ * default logger drops. Package-private so both AMQP connections use the same two names.
+ */
+final class AmqpConnectionFailureLogger extends DefaultExceptionHandler {
+
+    static final String REPLY_INBOX = "fleetd-reply-inbox";
+    static final String LEAD_MAILBOX = "fleetd-lead-mailbox";
+
+    private final String connectionName;
+    private final Logger logger;
+
+    AmqpConnectionFailureLogger(String connectionName, Logger logger) {
+        this.connectionName = connectionName;
+        this.logger = logger;
+    }
+
+    String connectionName() {
+        return connectionName;
+    }
+
+    @Override
+    protected void log(String message, Throwable cause) {
+        if (isSocketClosedOrConnectionReset(cause)) {
+            logger.warn("AMQP connection {}: {} (Exception message: {})", connectionName, message, cause.getMessage());
+        } else {
+            logger.error("AMQP connection {}: {}", connectionName, message, cause);
+        }
+    }
+
+    private static boolean isSocketClosedOrConnectionReset(Throwable cause) {
+        // Deliberate copy of ForgivingExceptionHandler's private static helper; check it on amqp-client upgrades.
+        if (!(cause instanceof IOException)) {
+            return false;
+        }
+        return "Connection reset".equals(cause.getMessage())
+                || "Socket closed".equals(cause.getMessage())
+                || "Connection reset by peer".equals(cause.getMessage());
+    }
+}
@@ -122,17 +122,22 @@ public final class LeadMailbox implements LeadChannel, AutoCloseable {
    /** As {@link #open(String, String)}, with an explicit consumer prefetch. */
    public static LeadMailbox open(String uri, String selfCoordId, int prefetch) {
        try {
-            ConnectionFactory factory = new ConnectionFactory();
-            factory.setUri(uri);
-            // Self-heal transient blips; topology recovery re-declares the queue and re-attaches the consumer.
-            factory.setAutomaticRecoveryEnabled(true);
-            factory.setTopologyRecoveryEnabled(true);
-            return new LeadMailbox(factory.newConnection("fleetd-lead-mailbox"), selfCoordId, prefetch);
+            return new LeadMailbox(connectionFactory(uri).newConnection(AmqpConnectionFailureLogger.LEAD_MAILBOX), selfCoordId, prefetch);
        } catch (Exception e) {
            throw new IllegalStateException("cannot connect to AMQP coordination broker at " + uri, e);
        }
    }

+    static ConnectionFactory connectionFactory(String uri) throws Exception {
+        ConnectionFactory factory = new ConnectionFactory();
+        factory.setUri(uri);
+        // Self-heal transient blips; topology recovery re-declares queues and re-attaches consumers.
+        factory.setAutomaticRecoveryEnabled(true);
+        factory.setTopologyRecoveryEnabled(true);
+        factory.setExceptionHandler(new AmqpConnectionFailureLogger(AmqpConnectionFailureLogger.LEAD_MAILBOX, log));
+        return factory;
+    }
+
    /** Wrap an already-open connection with {@link #DEFAULT_PREFETCH} (injection seam for tests). */
    LeadMailbox(Connection connection, String selfCoordId) {
        this(connection, selfCoordId, DEFAULT_PREFETCH);
@@ -466,8 +466,18 @@ public final class MessageService {
        if (candidates.size() == 1) {
            Task orphan = candidates.get(0);
            if (orphan.future.complete(new Reply(Outcome.REPLIED, content))) {
-                if (orphan.turnId != null) {
-                    asyncTasksByTurn.remove(orphan.turnId, orphan);
+                // fleetd #329 (F3): read orphan.turnId once. It used to be read twice, under no
+                // lock — once for this null check, once as the removal key — the same double-read
+                // shape fleetd #324 fixed in finishAsyncTask. clearAsyncQuestion's unlocked
+                // forgetTurn=true path (ask()'s timeout cleanup) can null the field between the two
+                // reads; capturing it once removes the torn read here too.
+                String turnId = orphan.turnId;
+                if (turnId != null) {
+                    if (replyOrphanTurnIdRaceHookForTest != null) {
+                        // Test-only (fleetd #329, F3): see the field's own javadoc.
+                        replyOrphanTurnIdRaceHookForTest.run();
+                    }
+                    asyncTasksByTurn.remove(turnId, orphan);
                }
                count(FleetMetrics.REPLIES, "path", "async-recovered");
                return true;                   // the ticket itself took it — no inbox stranding at all
@@ -699,27 +709,47 @@ public final class MessageService {
            boolean isRecovery = task == recoveryTask && recovered != null;
            Reply outcome = isRecovery ? recovered : new Reply(Outcome.WORKER_FAILED, reason);
            String turnId = task.turnId;
-            if (task.future.complete(outcome)) {
-                if (outcome.outcome() == Outcome.WORKER_FAILED) {
-                    asyncFailed = true;
+            // fleetd #335: completing THIS task's future must not depend on any other task's
+            // cleanup succeeding — every task in `matching` is owed its own outcome regardless of
+            // what happens below, so decide and record that before doing anything that can throw.
+            boolean completedHere = task.future.complete(outcome);
+            if (completedHere && outcome.outcome() == Outcome.WORKER_FAILED) {
+                asyncFailed = true;
+            }
+            try {
+                if (abandonCleanupHookForTest != null) {
+                    // Test-only (fleetd #335 site 1): see the field's own javadoc.
+                    abandonCleanupHookForTest.run();
                }
-                if (turnId != null) {
-                    // #275: whether this task was swept out of ASKING or was already answered and
-                    // only waiting on its resumed turn's real reply (#137), nothing will ever
-                    // complete this turnId now — drop it from this class's own bookkeeping AND the
-                    // reverse-rendezvous itself, so hasAsyncQuestion(target) stops reporting a turn
-                    // that is actually done, and a late answer() sees it as lapsed rather than
-                    // resolving a question nothing is listening for any more.
-                    asyncTasksByTurn.remove(turnId, task);
-                    rendezvous.closeAsk(turnId);
+                if (completedHere) {
+                    if (turnId != null) {
+                        // #275: whether this task was swept out of ASKING or was already answered and
+                        // only waiting on its resumed turn's real reply (#137), nothing will ever
+                        // complete this turnId now — drop it from this class's own bookkeeping AND the
+                        // reverse-rendezvous itself, so hasAsyncQuestion(target) stops reporting a turn
+                        // that is actually done, and a late answer() sees it as lapsed rather than
+                        // resolving a question nothing is listening for any more.
+                        asyncTasksByTurn.remove(turnId, task);
+                        rendezvous.closeAsk(turnId);
+                    }
+                } else if (isRecovery) {
+                    // The recovered reply was already drained out of the inbox, but this task resolved
+                    // through another path (e.g. a concurrent reply() or a second abandon() racing this
+                    // one) between us choosing it and completing it here. Put the reply back rather than
+                    // lose it silently — it may still belong to some other still-open task, or the next
+                    // caller that drains this target's inbox.
+                    inbox.publish(target, UUID.randomUUID().toString(), recovered.text());
                }
-            } else if (isRecovery) {
-                // The recovered reply was already drained out of the inbox, but this task resolved
-                // through another path (e.g. a concurrent reply() or a second abandon() racing this
-                // one) between us choosing it and completing it here. Put the reply back rather than
-                // lose it silently — it may still belong to some other still-open task, or the next
-                // caller that drains this target's inbox.
-                inbox.publish(target, UUID.randomUUID().toString(), recovered.text());
+            } catch (RuntimeException e) {
+                // fleetd #335: inbox.publish reaches a broker (AmqpReplyInbox throws
+                // IllegalStateException on an unroutable/unconfirmed/interrupted publish) and this
+                // loop has no other teardown path — a caller on the release path, or the health
+                // monitor's GONE/NEVER_READY sweep. Losing this exception uncaught would abort the
+                // loop and leave every task still to come in `matching` PENDING forever (fleetd
+                // #335 site 1). completedHere is already recorded above, so only this task's
+                // best-effort bookkeeping is lost — log it and let the loop reach the rest.
+                log.error("abandon: per-task cleanup failed for ticket {} (target {}, turnId {})",
+                        task.ticket, target, turnId, e);
            }
        }
        if (failed) {
@@ -853,16 +883,26 @@ 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) {
+                        if (timeoutCancellationRaceHookForTest != null) {
+                            // Test-only (fleetd #345): see the field's own javadoc.
+                            timeoutCancellationRaceHookForTest.run();
+                        }
+                        // 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(
@@ -925,13 +965,41 @@ public final class MessageService {
            return new AskResult(AskOutcome.ANSWERED, answer);
        } catch (TimeoutException e) {
            log.debug("fleet_ask from {} went unanswered in {}ms", workerSession, timeoutMillis);
-            // fleetd #307: mark the task BEFORE clearAsyncQuestion(forgetTurn=true) below drops it out of
-            // asyncTasksByTurn and nulls its turnId — that forgetting is deliberate and stays (it is
-            // what keeps the target from staying BUSY forever), but it would otherwise also erase
-            // askAnsweredAsyncTasks' only signal that the worker's eventual real fleet_reply still
-            // belongs to this task, stranding it in the inbox with a false "never replied" verdict.
-            markAskTimedOut(ticket.turnId());
-            clearAsyncQuestion(ticket.turnId(), true);
+            // Only the fresh owner tears down the shared turn (mirrors the finally block below).
+            // A duplicate's own timeoutMillis says nothing about whether the SHARED ask is actually
+            // done — it must leave the close/forget bookkeeping to the fresh owner, exactly as it
+            // already leaves closeAsk to it.
+            if (ticket.fresh()) {
+                // fleetd #334: close the ask turn BEFORE forgetting this task's turnId mapping below.
+                // Before this fix the order was reversed — the mapping was forgotten here first, and
+                // rendezvous.closeAsk only ran afterward, in the shared finally. A primary's answer()
+                // call racing this exact timeout could then find rendezvous.askSession(turnId) still
+                // non-null (the ask still "answerable") after the Task mapping was already gone:
+                // answer()'s own asyncTasksByTurn lookup returned null, its task != null guard skipped
+                // the completion, and the async ticket sat at PENDING forever even though answer()
+                // itself reported the worker's real reply. Closing here first removes that window:
+                // any answer() call that still observes askSession(turnId) != null is necessarily
+                // racing a point BEFORE the forgetting below runs (both happen on this one thread, in
+                // this order, with nothing that yields in between), so the Task mapping is still there
+                // for it to find; any call that observes askSession(turnId) == null now correctly
+                // bails out STALE_TURN (see answer()'s own top check) before ever reaching
+                // asyncTasksByTurn. rendezvous.closeAsk is idempotent — a no-op once the turn is
+                // already removed, see its own javadoc — so the shared finally below re-running it
+                // for this same fresh call is harmless.
+                rendezvous.closeAsk(ticket.turnId());
+                if (askTimeoutRaceHookForTest != null) {
+                    // Test-only (fleetd #334): see the field's own javadoc.
+                    askTimeoutRaceHookForTest.run();
+                }
+                // fleetd #307: mark the task BEFORE clearAsyncQuestion(forgetTurn=true) below drops it
+                // out of asyncTasksByTurn and nulls its turnId — that forgetting is deliberate and
+                // stays (it is what keeps the target from staying BUSY forever), but it would
+                // otherwise also erase askAnsweredAsyncTasks' only signal that the worker's eventual
+                // real fleet_reply still belongs to this task, stranding it in the inbox with a false
+                // "never replied" verdict.
+                markAskTimedOut(ticket.turnId());
+                clearAsyncQuestion(ticket.turnId(), true);
+            }
            return new AskResult(AskOutcome.TIMED_OUT, null);
        } catch (ExecutionException e) {
            Throwable cause = e.getCause();
@@ -1002,13 +1070,58 @@ public final class MessageService {
                // Measured when #282 was merged: this guard is DEFENCE IN DEPTH, not the thing
                // that makes the chained ask work. ask() calls markAsyncQuestion (:860) before
                // resolveQuestion (:861), so by the time this thread wakes, the task has already
-                // moved to the new turnId and finishAsyncTask(oldTurnId, ...) finds nothing. Removing
-                // this guard alone leaves the test green. Keep it anyway: it mirrors sendAsync's
-                // sibling guard, and that sibling's own comment (:1017) warns the two orderings are
-                // not something to rely on. Do NOT delete it as dead code without re-checking that
-                // ordering, and do not treat it as the sole protection either.
-                if (result.outcome() != Outcome.QUESTION) {
-                    finishAsyncTask(turnId, result);
+                // moved to the new turnId — this outcome check, not a Task lookup, is what tells the
+                // two cases apart (see fleetd #329 below). Removing this guard alone leaves the test
+                // green. Keep it anyway: it mirrors sendAsync's sibling guard, and that sibling's own
+                // comment (:1017) warns the two orderings are not something to rely on. Do NOT delete
+                // it as dead code without re-checking that ordering, and do not treat it as the sole
+                // protection either.
+                //
+                // fleetd #329 (F1): complete the SAME Task object this method already looked up at
+                // :991, rather than re-resolving it from turnId a second time. The old
+                // finishAsyncTask(turnId, result) did its own asyncTasksByTurn.get(turnId) here, and
+                // that second lookup races ask()'s own timeout path: ask()'s ticket.answer().get(...)
+                // can time out (or lose that exact race) at essentially the same instant this method's
+                // rendezvous.answerAsk(turnId, ...) above already succeeded, running
+                // markAskTimedOut + clearAsyncQuestion(turnId, true) with no lock at all — which
+                // forgets turnId (removes it from asyncTasksByTurn, nulls Task.turnId) before this
+                // thread ever gets here. The worker's real reply then arrived, this method's own wait
+                // woke up with it, and the by-turnId lookup found nothing: the async ticket's future
+                // was never completed, so fleet_poll{ticket} stayed PENDING forever even though
+                // answer() itself correctly returned REPLIED. Reusing the reference captured at :991
+                // — before that race window opens — sidesteps the second lookup entirely: it is the
+                // identical Task whatever asyncTasksByTurn or Task.turnId say by the time we reach
+                // this point, so finishAsyncTask(Task, Reply) can still complete its future and detach
+                // it using whatever turnId it now reads. The #282 chained-ask case is unaffected
+                // because it is still gated purely by result.outcome() == QUESTION above, which does
+                // not depend on this lookup — widening what "task" means here cannot complete a ticket
+                // the chained ask deliberately left open.
+                //
+                // A null task is NOT only "this was never an async ticket". That reading was in this
+                // comment when #329 merged and it is wrong; it is still not the whole story after
+                // #334. A genuine async ticket can still land here with task == null — a blocking
+                // (wait:true) send's fleet_ask never has a Task at all, so that case is expected and
+                // fine. What #334 fixed was a SECOND, unintended way to get here with task == null:
+                // ask()'s timeout path used to run clearAsyncQuestion(turnId, true) — which drops the
+                // asyncTasksByTurn entry — in its catch block, while rendezvous.closeAsk(turnId) ran
+                // later, in its finally. Between those two the ask was still answerable but the map
+                // entry was already gone, so the lookup at :1053 returned null and this ticket was
+                // never completed. Measured on 2026-09-04: a probe firing only that first half before
+                // answer() ran printed "answer=REPLIED phase=PENDING reply=null" — the same stranded
+                // ticket #329 set out to fix, one step earlier in the same race; the probe used
+                // forgetTurnForTest, which omits ask()'s markAskTimedOut, and that omission does not
+                // change the outcome, because askTimedOut is read only by askAnsweredAsyncTasks, and
+                // reply() never reaches it while this method's own waiter is live. #334's fix
+                // reorders ask()'s timeout catch to run closeAsk before the forgetting (see the
+                // fresh-owner block there), which removes this path entirely rather than narrowing
+                // it further: once closeAsk has run, rendezvous.askSession(turnId) is null and
+                // answer() returns STALE_TURN from its own top check, before it ever reaches this
+                // lookup — see aLateAnswerDuringAskTimeoutTeardownStillCompletesTheAsyncTicket in
+                // MessageServiceTest, which pins the exact window with askTimeoutRaceHookForTest.
+                // So by the time this line runs, task == null means only the ordinary blocking-send
+                // case (or #329's own already-fixed race elsewhere) — not this one.
+                if (result.outcome() != Outcome.QUESTION && task != null) {
+                    finishAsyncTask(task, result);
                }
                return result;
            } catch (TimeoutException e) {
@@ -1059,14 +1172,25 @@ public final class MessageService {
            // this fires exactly once, from whichever path completes it: finishAsyncTask(task, result)
            // below on any non-QUESTION outcome of send() — a worker's fleet_reply, the CB-106
            // completion fallback, a CB-109 wedge, TIMED_OUT, BUSY, or BACKEND_EXHAUSTED — the same
-            // finishAsyncTask reached via answer()'s finishAsyncTask(turnId, result) once a QUESTION
+            // finishAsyncTask reached via answer()'s own finishAsyncTask(task, result) once a QUESTION
            // is resolved, completeExceptionally(t) just below when send() itself throws, or a CB-516
            // abandon() on teardown. Without this, MessageService.reply's rendezvous fast path (the
            // one an async ticket always takes) never told the push loop anything happened — see the
            // class javadoc on sendAsync/CB-107.
            task.future.whenComplete((reply, ex) -> {
                boolean failed = ex != null || reply == null || !reply.completed();
-                pushLoop.onTicketTerminal(ticket, target, failed);
+                try {
+                    pushLoop.onTicketTerminal(ticket, target, failed);
+                } catch (Throwable t) {
+                    // fleetd #335 (site 2): this stage's own CompletableFuture is discarded, so an
+                    // uncaught throw here (e.g. a RejectedExecutionException from
+                    // ReplyPushLoop's scheduler, already shut down while this in-flight send's
+                    // whenComplete fires during the daemon's own shutdown sequence — messages.close()
+                    // only stops accepting NEW async work, it does not cancel a delivery already
+                    // running) vanishes with no log line and no metric, and the push loop never learns
+                    // the ticket went terminal — the exact thing this hook exists to tell it.
+                    log.error("push loop failed to learn ticket {} (target {}) went terminal", ticket, target, t);
+                }
            });
        }
        asyncExecutor.submit(() -> {
@@ -1079,7 +1203,22 @@ public final class MessageService {
                    finishAsyncTask(task, result);
                }
            } catch (Throwable t) {
-                task.future.completeExceptionally(t);
+                // fleetd #329 (F2): finishAsyncTask above already completes task.future — on its very
+                // first line — before doing anything else, so anything that throws afterward (inside
+                // finishAsyncTask's own cleanup, or from a future addition to this try block) lands
+                // here with the future already resolved. completeExceptionally on an already-completed
+                // future is a silent no-op: it returns false and does nothing, so without the check
+                // below the exception simply vanished — no log, no metric, nothing. Measured (see the
+                // ticket): temporarily reintroducing the fleetd #324 NPE reproduced 19 real exceptions
+                // on the ordinary path, with 74/74 tests staying green and not one log line produced.
+                // Do not stop completing the future first — finishAsyncTask completing before it
+                // cleans up is what makes a late failure harmless to the ticket's own result — only
+                // add the missing visibility for the case where that step, or whatever ran after it,
+                // has already lost the race to report through the future.
+                if (!task.future.completeExceptionally(t)) {
+                    log.error("async send {} -> {} threw after its ticket was already resolved",
+                            ticket, target, t);
+                }
            }
        });
        pruneTerminalTickets();
@@ -1247,6 +1386,10 @@ public final class MessageService {
     */
    private void finishAsyncTask(Task task, Reply result) {
        task.future.complete(result);
+        if (afterFinishAsyncTaskCompleteHookForTest != null) {
+            // Test-only (fleetd #329, F2): see the field's own javadoc.
+            afterFinishAsyncTaskCompleteHookForTest.run();
+        }
        String turnId = task.turnId;
        if (turnId != null) {
            if (finishAsyncTaskRaceHook != null) {
@@ -1276,6 +1419,67 @@ public final class MessageService {
        this.finishAsyncTaskRaceHook = hook;
    }

+    /**
+     * Null in production; test seam for fleetd #329 (F2) — invoked from {@link #finishAsyncTask(Task,
+     * Reply)} unconditionally, immediately after {@code task.future.complete(result)} runs (before
+     * {@link Task#turnId} is even read, so it fires regardless of whether this task was ever asked).
+     * A test installs this to force an exception into exactly the shape fleetd #329 identified:
+     * something throws inside {@link #sendAsync}'s executor task after the async ticket's future is
+     * already resolved, so the surrounding {@code catch (Throwable t)} can only report failure through
+     * {@code completeExceptionally} — a silent no-op on an already-completed future. Engineering a
+     * real exception to land in that exact post-completion window is what the ticket itself had to do
+     * by temporarily deleting a production guard (fleetd #324's {@code turnId != null} check); this
+     * hook drives the identical shape deterministically instead.
+     */
+    private volatile Runnable afterFinishAsyncTaskCompleteHookForTest;
+
+    /**
+     * Test-only (fleetd #329, F2): install {@link #afterFinishAsyncTaskCompleteHookForTest}.
+     * Package-private so the test, in the same package, can reach it without widening any production
+     * API.
+     */
+    void setAfterFinishAsyncTaskCompleteHookForTest(Runnable hook) {
+        this.afterFinishAsyncTaskCompleteHookForTest = hook;
+    }
+
+    /**
+     * Null in production; test seam for fleetd #345 — invoked in {@link #send}'s timeout path after
+     * {@link Injector.Delivery#completion()} reports incomplete and before {@link Injector#cancel}
+     * takes the target monitor. A test installs this to make {@code onStatus} pick the exact queued
+     * delivery up in that window, so {@code cancel} returns {@link Injector.Cancellation#DELIVERED}.
+     * This deterministically covers the caller's need to use that result rather than relying on a
+     * timing-sensitive real race.
+     */
+    private volatile Runnable timeoutCancellationRaceHookForTest;
+
+    /**
+     * Test-only (fleetd #345): install {@link #timeoutCancellationRaceHookForTest}. Package-private
+     * so the test, in the same package, can reach it without widening any production API.
+     */
+    void setTimeoutCancellationRaceHookForTest(Runnable hook) {
+        this.timeoutCancellationRaceHookForTest = hook;
+    }
+
+    /**
+     * Null in production; test seam for fleetd #329 (F3) — invoked from {@link #reply} right after
+     * the single local read of {@code orphan.turnId} passes its null-check and before that (now-local)
+     * value is used as the {@code asyncTasksByTurn} removal key. Mirrors {@link
+     * #finishAsyncTaskRaceHook} exactly, for the structurally identical double-read fleetd #324 fixed
+     * in {@link #finishAsyncTask}: a test installs this to force, deterministically, {@code
+     * clearAsyncQuestion}'s unlocked {@code forgetTurn=true} path nulling {@link Task#turnId} in that
+     * exact window, and to confirm the single-read fix tolerates it (the captured local is used
+     * unconditionally, so a hook that nulls the field afterward cannot affect this call).
+     */
+    private volatile Runnable replyOrphanTurnIdRaceHookForTest;
+
+    /**
+     * Test-only (fleetd #329, F3): install {@link #replyOrphanTurnIdRaceHookForTest}. Package-private
+     * so the test, in the same package, can reach it without widening any production API.
+     */
+    void setReplyOrphanTurnIdRaceHookForTest(Runnable hook) {
+        this.replyOrphanTurnIdRaceHookForTest = hook;
+    }
+
    /**
     * Test-only (fleetd #324): run the exact production cleanup {@link #ask}'s own timeout path runs
     * unlocked — {@link #clearAsyncQuestion(String, boolean)} with {@code forgetTurn=true} — so a test
@@ -1285,12 +1489,54 @@ public final class MessageService {
        clearAsyncQuestion(turnId, true);
    }

-    /** Complete the async ticket correlated to a specific answered turn. */
-    private void finishAsyncTask(String turnId, Reply result) {
-        Task task = asyncTasksByTurn.get(turnId);
-        if (task != null) {
-            finishAsyncTask(task, result);
-        }
+    /**
+     * Null in production; test seam for fleetd #335 (site 1) — invoked from {@link #abandon(String,
+     * String, boolean)}'s per-task loop, once per task, right before that task's own cleanup
+     * (turnId bookkeeping, or the stranded-reply put-back) runs. A test installs this to inject a
+     * throw at that exact point deterministically.
+     *
+     * <p>The one production call there that can really throw is {@code inbox.publish} in the
+     * put-back branch — {@link AmqpReplyInbox#publish} reaches a broker and throws {@link
+     * IllegalStateException} on an unroutable, unconfirmed, or interrupted publish — but reaching
+     * that branch requires a second completion of the very task {@code abandon} is about to
+     * complete to win the race first (see the branch's own comment), and the #137 follow-up
+     * investigation above already found the combination this needs (a stranded reply coinciding
+     * with an open matching task) unreachable through the public API, not merely hard to time.
+     * This hook reproduces the resulting shape — a per-task cleanup throw — directly, the same
+     * technique {@link #finishAsyncTaskRaceHook} and {@link
+     * #afterFinishAsyncTaskCompleteHookForTest} already use for their own hard-to-time races.
+     */
+    private volatile Runnable abandonCleanupHookForTest;
+
+    /**
+     * Test-only (fleetd #335, site 1): install {@link #abandonCleanupHookForTest}. Package-private
+     * so the test, in the same package, can reach it without widening any production API.
+     */
+    void setAbandonCleanupHookForTest(Runnable hook) {
+        this.abandonCleanupHookForTest = hook;
+    }
+
+    /**
+     * Null in production; test seam for fleetd #334 — invoked from {@link #ask}'s {@code
+     * TimeoutException} catch, only for the fresh owner, right after {@code rendezvous.closeAsk}
+     * has run and before {@link #markAskTimedOut} / {@link #clearAsyncQuestion} forget this task's
+     * turnId mapping. A test installs this to call {@link #answer} for the very same {@code turnId}
+     * synchronously from inside that exact window, deterministically reproducing the race a real
+     * concurrent {@code answer()} call could otherwise only win by timing luck: with the ask already
+     * closed, that call must see {@code rendezvous.askSession(turnId) == null} and return {@link
+     * Outcome#STALE_TURN} immediately, never reaching {@code asyncTasksByTurn} at all — proving the
+     * window fleetd #334 describes (mapping forgotten while the ask was still "answerable") is
+     * closed, rather than merely narrowed the way fleetd #329 narrowed the sibling race in {@link
+     * #finishAsyncTask}.
+     */
+    private volatile Runnable askTimeoutRaceHookForTest;
+
+    /**
+     * Test-only (fleetd #334): install {@link #askTimeoutRaceHookForTest}. Package-private so the
+     * test, in the same package, can reach it without widening any production API.
+     */
+    void setAskTimeoutRaceHookForTest(Runnable hook) {
+        this.askTimeoutRaceHookForTest = hook;
    }

    /** A new send must not open a waiter while an async ticket owns this worker's paused turn. */
@@ -638,6 +638,79 @@ class ConfigRefTest {
        assertEquals("mac-b", ref.get().coordinator().selfId());
    }

+    /**
+     * fleetd #333: {@code fleet:} is split too — {@code fleet.leaders} is read only at
+     * {@code Fleetd.java:281} to build the {@code LeadTabScanner}'s identity map (fed into
+     * {@code CallerResolver}) and to auto-launch leads via {@code LeadLauncher.ensureLeads()}, and
+     * neither is rebuilt on reload, while the rest of {@code fleet:} (role pools, charters,
+     * tabLabel) is read live through the {@code CompositePeerLauncher} supplier. Before this fix
+     * {@code fleet} sat in the coverage checker's hot-excluded escape hatch, so a reload that
+     * changed only {@code fleet.leaders} reported a bare "config reloaded" — exactly the
+     * under-claim the split class exists to prevent for {@code health}/{@code coordinator}.
+     */
+    @Test
+    void changingFleetLeadersIsReportedAsSplit(@TempDir Path dir) throws Exception {
+        Path f = dir.resolve("fleetd.yaml");
+        Files.writeString(f, yaml("""
+                fleet:
+                  leaders:
+                    opus:
+                      tab: "lead: opus-a"
+                      profile: sonnet
+                """));
+        ConfigRef ref = refFor(f);
+
+        Files.writeString(f, yaml("""
+                fleet:
+                  leaders:
+                    opus:
+                      tab: "lead: opus-b"
+                      profile: sonnet
+                """));
+        ConfigRef.Outcome out = ref.reload();
+
+        assertTrue(out.applied());
+        assertTrue(out.deferred().isEmpty(), out.deferred().toString());
+        assertEquals(1, out.split().size(), out.split().toString());
+        assertTrue(out.split().getFirst().startsWith("fleet:"), out.split().toString());
+        assertTrue(out.split().getFirst().contains("restart"), out.split().toString());
+        assertTrue(out.split().getFirst().contains("live"), out.split().toString());
+        assertTrue(out.summary().contains("partially live"), out.summary());
+        // The snapshot still carries the new value — the LeadTabScanner's identity map and
+        // LeadLauncher's auto-launch are what wait for a restart; a reload rebuilds neither.
+        assertEquals("lead: opus-b", ref.get().fleet().leaders().get("opus").tab());
+    }
+
+    /**
+     * fleetd #333: {@code fleet.leaders} is the ONLY frozen part of {@code fleet:}. A reload that
+     * changes {@code tabLabel} (or charters, or a role pool) without touching {@code fleet.leaders}
+     * must stay fully hot with nothing reported — proving {@link ConfigRef#changedSplitKeys}
+     * compares {@code fleet.leaders} specifically rather than the whole {@code Fleet} record, which
+     * would over-claim "needs a restart" for a change that is genuinely all live (the mirror
+     * mistake of the under-claim this class exists to prevent).
+     */
+    @Test
+    void changingFleetTabLabelWithoutLeadersStaysFullyHot(@TempDir Path dir) throws Exception {
+        Path f = dir.resolve("fleetd.yaml");
+        Files.writeString(f, yaml("""
+                fleet:
+                  tabLabel: "{role}: {profile} #{n}"
+                """));
+        ConfigRef ref = refFor(f);
+
+        Files.writeString(f, yaml("""
+                fleet:
+                  tabLabel: "[{profile}] {role}"
+                """));
+        ConfigRef.Outcome out = ref.reload();
+
+        assertTrue(out.applied());
+        assertTrue(out.deferred().isEmpty(), out.deferred().toString());
+        assertTrue(out.split().isEmpty(), out.split().toString());
+        assertEquals("config reloaded", out.summary());
+        assertEquals("[{profile}] {role}", ref.get().fleet().tabLabel());
+    }
+
    /**
     * A split change must not refuse the reload (invariant 2 of fleetd #330) and
     * {@code Outcome.applied()} must stay {@code true} (invariant 3) — unlike a cold change, the
@@ -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
@@ -76,20 +76,24 @@ class ConfigRefTopLevelCoverageTest {
     *   <li>{@code placement} — read live by the placement policy on every spawn (class doc, Hot
     *       bullet; {@code ConfigRefTest.aConsumerHoldingTheRefSeesTheNewValue} proves it
     *       behaviourally).</li>
-     *   <li>{@code fleet} — role pools, {@code charters} and {@code tabLabel} are read live through
-     *       the supplier on {@code CompositePeerLauncher} (class doc, Hot bullet). The one
-     *       documented exception, {@code fleet.leaders}, is read only at startup and genuinely needs
-     *       a restart — a real gap in {@code changedDeferredKeys}, but one the class doc already
-     *       carries and that fleetd #330 explicitly did not re-open (its "seven readers" fact-find
-     *       named {@code health}/{@code coordinator} as the complete set of split-shaped keys, not
-     *       {@code fleet}). Reported as a caveat, not fixed here.</li>
     *   <li>{@code memberCredentials} — read live at {@code Fleetd.java:198, 205, 729}.</li>
     *   <li>{@code memberLoginShell} — read live at
     *       {@code HerdrPeerLauncher#configuredMemberLoginShell}.</li>
     * </ul>
+     *
+     * <p>{@code fleet} used to sit here too, on the strength of most of it (role pools, charters,
+     * tabLabel) being read the same live way — but {@code fleet.leaders} inside the same key is
+     * read only at startup and genuinely needs a restart, which is a real gap this escape hatch
+     * cannot represent: it excuses a whole top-level key from reload bookkeeping, and {@code fleet}
+     * needed exactly half of it excused. fleetd #333 moved it into {@link ConfigRef#SPLIT_KEYS}
+     * instead, where {@code changedSplitKeys} reports the frozen half by name. That is the
+     * cautionary tale this test's own javadoc already told: this checker proves the record's shape
+     * is triaged, never that a bucket a key sits in is the right one — a person has to read the
+     * source, which is exactly how fleetd #333 found {@code fleet} sitting in the wrong bucket
+     * while this test stayed green throughout.</p>
     */
    private static final Set<String> HOT_EXCLUDED_TOP_LEVEL_KEYS =
-            Set.of("placement", "fleet", "memberCredentials", "memberLoginShell");
+            Set.of("placement", "memberCredentials", "memberLoginShell");

    @Test
    void everyTopLevelComponentIsAccountedForInExactlyOneClass() {
@@ -116,7 +120,7 @@ class ConfigRefTopLevelCoverageTest {

        // The escape hatch is pinned. Growing it requires editing this line — a visible, deliberate
        // diff, not a quiet one. See the field javadoc above for what "belongs here" actually means.
-        assertEquals(Set.of("placement", "fleet", "memberCredentials", "memberLoginShell"), hot,
+        assertEquals(Set.of("placement", "memberCredentials", "memberLoginShell"), hot,
                "HOT_EXCLUDED_TOP_LEVEL_KEYS changed. A component belongs here ONLY if it is read "
                        + "live off the config supplier, never because adding it makes this test "
                        + "pass. If you are adding one to silence this test, that is fleetd #323 "
@@ -0,0 +1,284 @@
+package dev.ltms.fleet.config;
+
+import org.junit.jupiter.api.Test;
+
+import java.lang.reflect.Constructor;
+import java.lang.reflect.RecordComponent;
+import java.util.ArrayList;
+import java.util.Arrays;
+import java.util.LinkedHashMap;
+import java.util.List;
+import java.util.Map;
+import java.util.Set;
+import java.util.TreeSet;
+
+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}, {@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}/{@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>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". 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". */
+    private static final Map<String, Object> ALT = altValues();
+
+    private static Map<String, Object> baseValues() {
+        Map<String, Object> v = new LinkedHashMap<>();
+        v.put("bind", new FleetConfig.Bind("127.0.0.1", 8765));
+        v.put("herdrSocket", "~/.config/herdr/a.sock");
+        v.put("memberHerdrSocket", "~/.config/herdr/member-a.sock");
+        v.put("profiles", Map.of());
+        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", 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", 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", 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", "group-a");
+        v.put("memberLoginShell", null);
+        assertNamesMatchComponents(v);
+        return v;
+    }
+
+    private static Map<String, Object> altValues() {
+        Map<String, Object> v = new LinkedHashMap<>();
+        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");
+        // 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", 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
+        // ConfigRefTest.changingFleetTabLabelWithoutLeadersStaysFullyHot) and would wrongly fail this
+        // test — that is by design, not a gap: this map exists to prove fleet.leaders is covered.
+        v.put("fleet", new FleetConfig.Fleet(
+                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", 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", 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", "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) {
+            names.add(rc.getName());
+        }
+        assertEquals(names, new TreeSet<>(values.keySet()),
+                "this test's value map has drifted from FleetConfig's actual top-level components — "
+                        + "update BASE/ALT alongside the record");
+    }
+
+    private static FleetConfig configOf(Map<String, Object> values) throws ReflectiveOperationException {
+        Class<?>[] types = Arrays.stream(COMPONENTS).map(RecordComponent::getType).toArray(Class<?>[]::new);
+        Object[] args = Arrays.stream(COMPONENTS).map(rc -> values.get(rc.getName())).toArray();
+        Constructor<FleetConfig> ctor = FleetConfig.class.getDeclaredConstructor(types);
+        return ctor.newInstance(args);
+    }
+
+    /** {@code BASE} with exactly one named top-level component swapped for its {@code ALT} value. */
+    private static FleetConfig mutate(String componentName) throws ReflectiveOperationException {
+        Map<String, Object> values = new LinkedHashMap<>(BASE);
+        values.put(componentName, ALT.get(componentName));
+        return configOf(values);
+    }
+
+    /**
+     * The mechanism fleetd #333 F2 asked for, applied to {@link ConfigRef#COLD_KEYS}: mutate each
+     * cold key in isolation and prove {@code changedColdKeys} actually names it, not just that
+     * {@code COLD_KEYS} claims it does.
+     */
+    @Test
+    void everyColdKeyIsActuallyReportedByChangedColdKeys() throws ReflectiveOperationException {
+        FleetConfig base = configOf(BASE);
+        List<String> uncovered = new ArrayList<>();
+        for (String key : new TreeSet<>(ConfigRef.COLD_KEYS)) {
+            FleetConfig mutated = mutate(key);
+            if (!ConfigRef.changedColdKeys(base, mutated).contains(key)) {
+                uncovered.add(key);
+            }
+        }
+        System.out.printf(
+                "ConfigRef.changedColdKeys reporting coverage — %d COLD_KEYS, %d verified%n",
+                ConfigRef.COLD_KEYS.size(), ConfigRef.COLD_KEYS.size() - uncovered.size());
+        assertEquals(List.of(), uncovered,
+                "these keys are in ConfigRef.COLD_KEYS but mutating them alone produces no matching "
+                        + "entry from changedColdKeys — a set entry with no comparison behind it: "
+                        + uncovered);
+    }
+
+    /**
+     * The same mechanism applied to {@link ConfigRef#SPLIT_KEYS}: mutate each split key in
+     * isolation and prove {@code changedSplitKeys} actually names it (message starting with
+     * {@code "<key>:"}), not just that {@code SPLIT_KEYS} claims it does. This is the exact check
+     * that would have failed fleetd #333's own reproduction — dropping the {@code coordinator}
+     * branch from {@code changedSplitKeys} while {@code "coordinator"} stayed in {@code SPLIT_KEYS}
+     * — see the mutation proof in the fleetd #333 PR description; the earlier checkers here (the
+     * shape test and the in-method assert) do not.
+     */
+    @Test
+    void everySplitKeyIsActuallyReportedByChangedSplitKeys() throws ReflectiveOperationException {
+        FleetConfig base = configOf(BASE);
+        List<String> uncovered = new ArrayList<>();
+        for (String key : new TreeSet<>(ConfigRef.SPLIT_KEYS)) {
+            FleetConfig mutated = mutate(key);
+            List<String> split = ConfigRef.changedSplitKeys(base, mutated);
+            if (split.stream().noneMatch(s -> s.startsWith(key + ":"))) {
+                uncovered.add(key);
+            }
+        }
+        System.out.printf(
+                "ConfigRef.changedSplitKeys reporting coverage — %d SPLIT_KEYS, %d verified%n",
+                ConfigRef.SPLIT_KEYS.size(), ConfigRef.SPLIT_KEYS.size() - uncovered.size());
+        assertEquals(List.of(), uncovered,
+                "these keys are in ConfigRef.SPLIT_KEYS but mutating them alone produces no matching "
+                        + "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);
+    }
+}
@@ -0,0 +1,192 @@
+package dev.ltms.fleet.config;
+
+import org.junit.jupiter.api.Test;
+
+import java.lang.reflect.Constructor;
+import java.lang.reflect.RecordComponent;
+import java.util.ArrayList;
+import java.util.Arrays;
+import java.util.LinkedHashMap;
+import java.util.List;
+import java.util.Locale;
+import java.util.Map;
+import java.util.Objects;
+import java.util.Set;
+import java.util.TreeSet;
+
+import static org.junit.jupiter.api.Assertions.assertEquals;
+
+/**
+ * Guards against a "defect factory" built into this file's own established pattern, found live
+ * while building the (parked) idle-sleep-guard PR: every time a component is added to
+ * {@link FleetConfig}, the record grows by one arg AND a new back-compat constructor is added at
+ * the OLD arity, so existing callers keep compiling. That is correct and required — see the
+ * constructor ladder just below the record header. But {@link #withDefaults()}'s own {@code return
+ * new FleetConfig(...)} call sits in this same file, written at a literal argument count. The very
+ * next time a component is added, the freshly-added back-compat constructor at the OLD arity
+ * silently captures that stale call, because it is now a legal overload at that arg count too. It
+ * compiles. Every other test passes, because nothing else exercises the new field. The new
+ * component is defaulted away — {@code null}, or whatever that back-compat overload defaults it to
+ * — on every {@link FleetConfig#load}. Measured, not theoretical: this exact sequence happened
+ * live when the {@code idleSleepGuard} component was added on a sibling branch; it was caught only
+ * because that branch's own new tests happened to assert on the new field's value.
+ *
+ * <p>This test proves the opposite property, and does it in a way that survives the next field
+ * being added without being rewritten: reflectively enumerate {@link FleetConfig}'s own record
+ * components (never a hardcoded count — the arity is exactly what changes over time), build one
+ * config through the true canonical constructor with a real, distinctive, non-null value in EVERY
+ * component (reusing the exact reflective-construction pattern
+ * {@link ConfigRefTopLevelReportingCoverageTest} already established for this file:
+ * {@code getDeclaredConstructor(exact record-component types)}, which resolves the canonical
+ * constructor by its true shape, not by binding to whichever overload happens to match arg count —
+ * the same way Jackson resolves it), call the real {@link FleetConfig#withDefaults()}, and assert
+ * every one of those values survives unchanged.
+ *
+ * <p>Why this is a valid check for every component, not just some: {@link #withDefaults()}'s own
+ * comments document that it only ever REPLACES a component when the incoming value is {@code null}
+ * (or blank, for {@code placement}) — {@code broker}/{@code primary}/{@code leadHeartbeat}/
+ * {@code configReload}/{@code coordinator}/{@code worktreeGroup}/{@code memberLoginShell} are left
+ * as-is unconditionally, and {@code bind}/{@code guard}/{@code lifecycle}/{@code auth}/
+ * {@code fleet}/{@code quarantineCooldownSeconds}/{@code memberCredentials}/{@code placement} are
+ * replaced only on null/blank input. A value that is never null or blank going in must therefore
+ * never change coming out, for every current component. No exclusion is needed today.
+ *
+ * <p>{@link #EXCLUDED_FROM_SURVIVAL_CHECK} exists anyway, kept deliberately empty and size-pinned
+ * by {@link #exclusionListSizeIsPinned()}: a future component that {@code withDefaults()} is
+ * <em>documented</em> to transform unconditionally (unlike every field today) would legitimately
+ * need one. Pinning the size at 0 means growing that set to make a failure go away is itself a
+ * visible diff to this test, not a silent one — a checker that can be silenced by adding to its
+ * own escape hatch is not a checker.
+ */
+class FleetConfigWithDefaultsPreservesEveryComponentTest {
+
+    private static final RecordComponent[] COMPONENTS = FleetConfig.class.getRecordComponents();
+
+    /** See the class javadoc — deliberately empty today; grow it only with a matching justification. */
+    private static final Set<String> EXCLUDED_FROM_SURVIVAL_CHECK = Set.of();
+
+    /** One real, distinctive, non-null (non-blank where blankness would mean "unset") value per component. */
+    private static Map<String, Object> baseValues() {
+        Map<String, Object> v = new LinkedHashMap<>();
+        v.put("bind", new FleetConfig.Bind("127.0.0.1", 8765));
+        v.put("herdrSocket", "~/.config/herdr/guard.sock");
+        v.put("memberHerdrSocket", "~/.config/herdr/member-guard.sock");
+        v.put("profiles", Map.of("sonnet", minimalProfile("sonnet")));
+        v.put("guard", new FleetConfig.Guard(List.of("host-guard")));
+        v.put("worktreeRoot", "/wt/guard");
+        v.put("lifecycle", new FleetConfig.Lifecycle(300, 5, 30, true));
+        v.put("spawnReadyTimeoutMs", 12_345);
+        v.put("spawnReadyPollMs", 234);
+        v.put("broker", new FleetConfig.Broker("amqp://guard", null, 7));
+        v.put("primary", new FleetConfig.Primary("term-guard", 4, 4000));
+        v.put("fleet", new FleetConfig.Fleet(
+                Map.of("opus", new FleetConfig.Leader("sonnet", "lead: opus-guard", 1, null, 10,
+                        "claude", null, null, null)),
+                Map.of(), Map.of(), Map.of(), Map.of(), "{role}: {profile} #{n}"));
+        v.put("leadHeartbeat", new FleetConfig.LeadHeartbeat(301, 61_000L, 4));
+        v.put("health", new FleetConfig.Health(true, 31, 601, 61, null));
+        v.put("placement", "round-robin");
+        v.put("auth", new FleetConfig.Auth("loopback-trust", null));
+        v.put("configReload", new FleetConfig.ConfigReload(true, 11));
+        v.put("quarantineCooldownSeconds", 1801);
+        v.put("memberCredentials", new FleetConfig.MemberCredentials(
+                FleetConfig.MemberCredentials.POLICY_DENY_BY_DEFAULT,
+                List.of("git"), List.of("git", "ssh"), null));
+        v.put("coordinator", new FleetConfig.Coordinator("amqp://coord-guard", null, "self-guard", 3));
+        v.put("worktreeGroup", "group-guard");
+        v.put("memberLoginShell", "/bin/zsh");
+        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);
+    }
+
+    /**
+     * Guards {@link #baseValues()} itself against drifting from the record's real shape — the same
+     * assurance {@link ConfigRefTopLevelReportingCoverageTest} already relies on. This is what makes
+     * "no hardcoded arity" true in practice: forgetting to add a new component here fails this
+     * assertion by name, rather than silently checking one component fewer than the record has.
+     */
+    private static void assertNamesMatchComponents(Map<String, Object> values) {
+        Set<String> names = new TreeSet<>();
+        for (RecordComponent rc : COMPONENTS) {
+            names.add(rc.getName());
+        }
+        assertEquals(names, new TreeSet<>(values.keySet()),
+                "this test's value map has drifted from FleetConfig's actual top-level components — "
+                        + "update baseValues() alongside the record");
+    }
+
+    /**
+     * Builds a {@link FleetConfig} through the TRUE canonical constructor — resolved by the record's
+     * own component types, not by argument count — so this never accidentally exercises a
+     * back-compat overload the way a literal {@code new FleetConfig(...)} call risks doing.
+     */
+    private static FleetConfig configOf(Map<String, Object> values) throws ReflectiveOperationException {
+        Class<?>[] types = Arrays.stream(COMPONENTS).map(RecordComponent::getType).toArray(Class<?>[]::new);
+        Object[] args = Arrays.stream(COMPONENTS).map(rc -> values.get(rc.getName())).toArray();
+        Constructor<FleetConfig> ctor = FleetConfig.class.getDeclaredConstructor(types);
+        return ctor.newInstance(args);
+    }
+
+    @Test
+    void exclusionListSizeIsPinned() {
+        assertEquals(0, EXCLUDED_FROM_SURVIVAL_CHECK.size(),
+                "EXCLUDED_FROM_SURVIVAL_CHECK grew from 0 — every entry needs a justification in "
+                        + "this test class's javadoc AND this assertion re-pinned to the new size; a "
+                        + "growing exclusion list that silences failures on its own is not a guard");
+    }
+
+    /**
+     * The mutation this is built to catch: make {@code withDefaults()}'s final constructor call
+     * literal at some arg count, add one more component to the record with a new back-compat
+     * constructor at the old arity, and the stale call silently rebinds. Every component here is
+     * real and non-null (non-blank for the one String — {@code placement} — where blank has
+     * meaning), so none of it should be replaced by {@code withDefaults()}; any component that
+     * comes back different was silently dropped.
+     */
+    @Test
+    void everyComponentGivenARealValueSurvivesWithDefaults() throws ReflectiveOperationException {
+        Map<String, Object> base = baseValues();
+        FleetConfig config = configOf(base);
+        FleetConfig defaulted = config.withDefaults();
+
+        List<String> dropped = new ArrayList<>();
+        int checked = 0;
+        for (RecordComponent rc : COMPONENTS) {
+            String name = rc.getName();
+            if (EXCLUDED_FROM_SURVIVAL_CHECK.contains(name)) {
+                continue;
+            }
+            checked++;
+            Object expected = base.get(name);
+            Object actual;
+            try {
+                actual = rc.getAccessor().invoke(defaulted);
+            } catch (ReflectiveOperationException e) {
+                throw new RuntimeException("failed to read FleetConfig." + name + "()", e);
+            }
+            if (!Objects.equals(expected, actual)) {
+                dropped.add(String.format(Locale.ROOT,
+                        "%s: withDefaults() was given a real, non-null value (%s) for '%s' but "
+                                + "returned %s — a component silently dropped by withDefaults(), the "
+                                + "shape of the defect this test exists to catch (its final "
+                                + "\"return new FleetConfig(...)\" call binding to a back-compat "
+                                + "constructor instead of the true canonical one)",
+                        name, expected, name, actual));
+            }
+        }
+
+        System.out.printf(Locale.ROOT,
+                "FleetConfig.withDefaults() component-survival coverage — %d components, %d checked, "
+                        + "%d excluded, %d survived%n",
+                COMPONENTS.length, checked, EXCLUDED_FROM_SURVIVAL_CHECK.size(), checked - dropped.size());
+        assertEquals(List.of(), dropped,
+                "withDefaults() silently dropped these real, given components: " + dropped);
+    }
+}
@@ -484,6 +484,27 @@ class CompletionResolverTest {

    // --- CB-578 stage A: backend-exhausted classification ---------------------------------

+    @Test
+    void aNormalMemberReportMentioningTheExhaustionPatternDoesNotNotifyTheSink() {
+        String block = "⏺ I reviewed capacity handling. The usage limit has been reached means no more work can start.\n❯ ";
+        FakeHerdr herdr = new FakeHerdr().readText(block);
+        Rendezvous rendezvous = new Rendezvous();
+        ExhaustedPatternLookup patterns = target -> Pattern.compile("usage limit has been reached");
+        java.util.List<String> notified = new java.util.ArrayList<>();
+        ExhaustionSink sink = (target, reason, profile) -> notified.add(target + ": " + reason);
+        CompletionResolver resolver = new CompletionResolver(new AgentControl(herdr), rendezvous, patterns, sink);
+
+        var waiter = rendezvous.open("term_a");
+        resolver.resolve("term_a", new CompletionResolver.InFlight(waiter, null));
+
+        assertEquals(Rendezvous.Kind.BACKEND_EXHAUSTED, waiter.getNow(null).kind(),
+                "a matching report still fails the send as exhausted");
+        assertTrue(waiter.getNow(null).text().contains("I reviewed capacity handling."),
+                "the exhausted result keeps the whole matched pane line");
+        assertTrue(notified.isEmpty(),
+                "a normal report mentioning an exhaustion pattern must not quarantine a credential");
+    }
+
    @Test
    void classifiesAMatchingScrapeAsBackendExhaustedInsteadOfACompletedReply() {
        String block = "⏺ Working on it...\nThe usage limit has been reached. Try again later.\n❯ ";
@@ -534,6 +555,53 @@ class CompletionResolverTest {
                "the sink is told the matched reason: " + notified.get(0));
    }

+    @Test
+    void aRealExhaustionBehindTerminalChromeStillNotifiesTheSink() {
+        String block = "⏺ │ The usage limit has been reached. Try again later.\n❯ ";
+        FakeHerdr herdr = new FakeHerdr().readText(block);
+        Rendezvous rendezvous = new Rendezvous();
+        ExhaustedPatternLookup patterns = target -> Pattern.compile("usage limit has been reached");
+        java.util.List<String> notified = new java.util.ArrayList<>();
+        ExhaustionSink sink = (target, reason, profile) -> notified.add(target + ": " + reason);
+        CompletionResolver resolver = new CompletionResolver(new AgentControl(herdr), rendezvous, patterns, sink);
+
+        var waiter = rendezvous.open("term_a");
+        resolver.resolve("term_a", new CompletionResolver.InFlight(waiter, null));
+
+        assertEquals(Rendezvous.Kind.BACKEND_EXHAUSTED, waiter.getNow(null).kind(),
+                "a real exhaustion must still fail the send as exhausted");
+        assertEquals(1, notified.size(),
+                "a real exhaustion behind terminal chrome must reach the sink");
+    }
+
+    /**
+     * The live fleet configures {@code exhaustedPattern: "The usage limit has been reached"} — with
+     * the leading {@code "The"}. Every other test here uses a pattern without it, which is the shape
+     * that made fleetd #348 need a looser rule than a start-of-line check. This pins the deployed
+     * shape as well, so a later tightening of {@link CompletionResolver} cannot silently stop
+     * recording the exhaustion this fleet actually reports.
+     *
+     * <p>What it does not prove: that this is the only pattern shape an operator will write.
+     */
+    @Test
+    void anExhaustionPatternCarryingItsLeadingWordsStillNotifiesTheSink() {
+        String block = "⏺ │ The usage limit has been reached. Try again later.\n❯ ";
+        FakeHerdr herdr = new FakeHerdr().readText(block);
+        Rendezvous rendezvous = new Rendezvous();
+        ExhaustedPatternLookup patterns = target -> Pattern.compile("The usage limit has been reached");
+        java.util.List<String> notified = new java.util.ArrayList<>();
+        ExhaustionSink sink = (target, reason, profile) -> notified.add(target + ": " + reason);
+        CompletionResolver resolver = new CompletionResolver(new AgentControl(herdr), rendezvous, patterns, sink);
+
+        var waiter = rendezvous.open("term_a");
+        resolver.resolve("term_a", new CompletionResolver.InFlight(waiter, null));
+
+        assertEquals(Rendezvous.Kind.BACKEND_EXHAUSTED, waiter.getNow(null).kind(),
+                "the live pattern shape must still fail the send as exhausted");
+        assertEquals(1, notified.size(),
+                "the live pattern shape must still reach the sink");
+    }
+
    @Test
    void aLosingBackendExhaustedClassificationNeverNotifiesTheExhaustionSink() {
        // The waiter was already resolved (e.g. by the worker's own reply) before this scrape landed —
@@ -804,7 +872,28 @@ class CompletionResolverTest {
    // --- fleetd#201 Unit 1: target-keyed backend-error pattern + typed sink ----------------------

    @Test
-    void aConfiguredBackendErrorPatternClassifiesAMatchAsAFailureAndNotifiesTheSinkOnce() {
+    void aNormalMemberReportMentioningTheFallbackErrorPatternFailsButDoesNotNotifyTheSink() {
+        String block = "⏺ I checked the retry path. An API Error: makes it back off.\n❯ ";
+        FakeHerdr herdr = new FakeHerdr().readText(block);
+        Rendezvous rendezvous = new Rendezvous();
+        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(), BackendErrorPatternLookup.legacy(), 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 scrape mentioning the pattern must still fail the send");
+        assertTrue(waiter.getNow(null).text().contains("I checked the retry path. An API Error: makes it back off."),
+                "the failure must keep the whole pane tail");
+        assertTrue(notified.isEmpty(),
+                "a normal report mentioning the fallback pattern must not record a credential failure");
+    }
+
+    @Test
+    void aConfiguredBackendErrorPatternAtTheStartOfALineClassifiesAMatchAndNotifiesTheSinkOnce() {
        String block = "⏺ 503 Service Unavailable: upstream credential rejected\n❯ ";
        FakeHerdr herdr = new FakeHerdr().readText(block);
        Rendezvous rendezvous = new Rendezvous();
@@ -924,7 +1013,7 @@ class CompletionResolverTest {
    }

    @Test
-    void aConfiguredPatternAlsoClassifiesTheRawScrapeFallbackAndNotifiesTheSink() {
+    void aConfiguredPatternAtTheStartOfALineAlsoClassifiesTheRawScrapeFallbackAndNotifiesTheSink() {
        // No ⏺ marker and leading TUI chrome ⇒ lastAssistantBlock() yields "", so classification must
        // fall back to the raw scrape (fleetd#211) — and it must use the configured pattern too.
        String block = """
@@ -949,6 +1038,40 @@ 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
@@ -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());
@@ -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 {
@@ -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
@@ -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
@@ -0,0 +1,134 @@
+package dev.ltms.fleet.msg;
+
+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 com.rabbitmq.client.Channel;
+import com.rabbitmq.client.ConnectionFactory;
+import com.rabbitmq.client.impl.DefaultExceptionHandler;
+import org.junit.jupiter.api.Test;
+import org.slf4j.LoggerFactory;
+
+import java.io.IOException;
+import java.lang.reflect.Proxy;
+import java.util.concurrent.atomic.AtomicInteger;
+
+import static org.junit.jupiter.api.Assertions.assertEquals;
+import static org.junit.jupiter.api.Assertions.assertFalse;
+import static org.junit.jupiter.api.Assertions.assertInstanceOf;
+import static org.junit.jupiter.api.Assertions.assertNotEquals;
+import static org.junit.jupiter.api.Assertions.assertTrue;
+
+class AmqpConnectionFailureLoggerTest {
+
+    @Test
+    void installedHandlersLogTheirOwnConnectionNamesAtErrorWithTheCause() throws Exception {
+        ConnectionFactory inboxFactory = AmqpReplyInbox.connectionFactory("amqp://127.0.0.1");
+        ConnectionFactory mailboxFactory = LeadMailbox.connectionFactory("amqp://127.0.0.1");
+
+        AmqpConnectionFailureLogger inboxHandler = installedStrictHandler(inboxFactory, "reply inbox");
+        AmqpConnectionFailureLogger mailboxHandler = installedStrictHandler(mailboxFactory, "lead mailbox");
+        assertEquals(AmqpConnectionFailureLogger.REPLY_INBOX, inboxHandler.connectionName());
+        assertEquals(AmqpConnectionFailureLogger.LEAD_MAILBOX, mailboxHandler.connectionName());
+
+        ListAppender<ILoggingEvent> inboxEvents = attach(AmqpReplyInbox.class);
+        ListAppender<ILoggingEvent> mailboxEvents = attach(LeadMailbox.class);
+        IllegalStateException inboxFailure = new IllegalStateException("inbox failure");
+        IllegalStateException mailboxFailure = new IllegalStateException("mailbox failure");
+        try {
+            inboxHandler.handleUnexpectedConnectionDriverException(null, inboxFailure);
+            mailboxHandler.handleConnectionRecoveryException(null, mailboxFailure);
+
+            assertError(inboxEvents, "AMQP connection fleetd-reply-inbox: An unexpected connection driver error occurred",
+                    inboxFailure, "inbox failure line");
+            assertError(mailboxEvents, "AMQP connection fleetd-lead-mailbox: Caught an exception during connection recovery!",
+                    mailboxFailure, "mailbox recovery line");
+        } finally {
+            detach(AmqpReplyInbox.class, inboxEvents);
+            detach(LeadMailbox.class, mailboxEvents);
+        }
+    }
+
+    @Test
+    void connectionResetKeepsForgivingHandlerWarningSemantics() {
+        AmqpConnectionFailureLogger handler = new AmqpConnectionFailureLogger(
+                AmqpConnectionFailureLogger.REPLY_INBOX, LoggerFactory.getLogger(AmqpReplyInbox.class));
+        ListAppender<ILoggingEvent> events = attach(AmqpReplyInbox.class);
+        try {
+            handler.handleUnexpectedConnectionDriverException(null, new IOException("Connection reset"));
+            assertEquals(1, events.list.size(), "the handler must still log a reset");
+            ILoggingEvent event = events.list.getFirst();
+            assertEquals(Level.WARN, event.getLevel(), "ForgivingExceptionHandler logs connection resets at WARN");
+            assertEquals("AMQP connection fleetd-reply-inbox: An unexpected connection driver error occurred "
+                            + "(Exception message: Connection reset)", event.getFormattedMessage());
+            assertTrue(event.getThrowableProxy() == null, "ForgivingExceptionHandler does not attach a reset stack trace");
+        } finally {
+            detach(AmqpReplyInbox.class, events);
+        }
+    }
+
+    @Test
+    void connectionNamesStayDistinct() {
+        assertNotEquals(AmqpConnectionFailureLogger.REPLY_INBOX, AmqpConnectionFailureLogger.LEAD_MAILBOX,
+                "reply-inbox and lead-mailbox failures must be distinguishable");
+    }
+
+    @Test
+    void strictConsumerExceptionStillClosesItsChannel() {
+        AtomicInteger closes = new AtomicInteger();
+        Channel channel = (Channel) Proxy.newProxyInstance(getClass().getClassLoader(), new Class<?>[] {Channel.class},
+                (_, method, _) -> switch (method.getName()) {
+                    case "close" -> {
+                        closes.incrementAndGet();
+                        yield null;
+                    }
+                    case "toString" -> "test-channel";
+                    default -> throw new UnsupportedOperationException(method.getName());
+                });
+        AmqpConnectionFailureLogger handler = new AmqpConnectionFailureLogger(
+                AmqpConnectionFailureLogger.REPLY_INBOX, LoggerFactory.getLogger(AmqpReplyInbox.class));
+
+        handler.handleConsumerException(channel, new IllegalStateException("consumer failed"), null, "tag", "handleDelivery");
+
+        assertEquals(1, closes.get(), "DefaultExceptionHandler must close a channel after a consumer exception");
+    }
+
+    @Test
+    void handlerOnlyChangesDefaultHandlerLogging() {
+        assertEquals(DefaultExceptionHandler.class,
+                AmqpConnectionFailureLogger.class.getSuperclass());
+        assertFalse(java.util.Arrays.stream(AmqpConnectionFailureLogger.class.getDeclaredMethods())
+                        .anyMatch(method -> method.getName().startsWith("handle")),
+                "all exception-handling methods must remain inherited from DefaultExceptionHandler");
+    }
+
+    private static ListAppender<ILoggingEvent> attach(Class<?> owner) {
+        Logger logger = (Logger) LoggerFactory.getLogger(owner);
+        logger.setLevel(Level.DEBUG);
+        ListAppender<ILoggingEvent> appender = new ListAppender<>();
+        appender.start();
+        logger.addAppender(appender);
+        return appender;
+    }
+
+    private static void detach(Class<?> owner, ListAppender<ILoggingEvent> appender) {
+        ((Logger) LoggerFactory.getLogger(owner)).detachAppender(appender);
+    }
+
+    private static AmqpConnectionFailureLogger installedStrictHandler(ConnectionFactory factory, String connection) {
+        assertInstanceOf(DefaultExceptionHandler.class, factory.getExceptionHandler(),
+                connection + " must keep DefaultExceptionHandler: replacing the strict handler with a forgiving one "
+                        + "changes when a channel is closed");
+        return assertInstanceOf(AmqpConnectionFailureLogger.class, factory.getExceptionHandler());
+    }
+
+    private static void assertError(ListAppender<ILoggingEvent> events, String message, Throwable cause, String name) {
+        assertEquals(1, events.list.size(), name);
+        ILoggingEvent event = events.list.getFirst();
+        assertEquals(Level.ERROR, event.getLevel(), name);
+        assertEquals(message, event.getFormattedMessage(), name);
+        assertEquals(cause.toString(), event.getThrowableProxy().getClassName() + ": "
+                + event.getThrowableProxy().getMessage(), name);
+    }
+}
@@ -1,5 +1,9 @@
 package dev.ltms.fleet.msg;

+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 dev.ltms.fleet.herdr.AgentControl;
 import dev.ltms.fleet.herdr.AgentStatus;
 import dev.ltms.fleet.herdr.FakeHerdr;
@@ -11,6 +15,7 @@ import dev.ltms.fleet.mcp.PrimaryRegistry;
 import dev.ltms.fleet.inject.Injector;
 import org.junit.jupiter.api.BeforeEach;
 import org.junit.jupiter.api.Test;
+import org.slf4j.LoggerFactory;

 import java.util.concurrent.CompletableFuture;
 import java.util.concurrent.TimeUnit;
@@ -253,7 +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));

@@ -372,6 +377,126 @@ class MessageServiceTest {
        assertEquals(MessageService.Outcome.QUESTION, q.outcome());
    }

+    /**
+     * fleetd #334. {@code ask()}'s {@code TimeoutException} catch used to forget this task's
+     * {@code turnId} mapping ({@code clearAsyncQuestion(turnId, true)}) BEFORE closing the ask
+     * ({@code rendezvous.closeAsk}, in the shared {@code finally}). A primary's {@code answer()}
+     * call racing that exact window found the ask still "answerable" ({@code
+     * rendezvous.askSession(turnId)} still non-null) while the {@code Task} was already forgotten,
+     * so its {@code task != null} guard skipped the completion and the async ticket sat at
+     * {@code PENDING} forever even though {@code answer()} itself reported a result. The fix
+     * (closing the ask first) makes this window impossible: a racing {@code answer()} call either
+     * still finds the ask open (and the {@code Task} mapping guaranteed intact) or finds it already
+     * closed (and bails {@code STALE_TURN} before ever touching the {@code Task}). This test pins
+     * the exact window with {@code askTimeoutRaceHookForTest} and proves both invariants the ticket
+     * named: (1) a late/racing {@code answer()} sees the ask as already lapsed ({@code STALE_TURN}),
+     * never made answerable again, and (2) the async ticket still resolves {@code DONE} once the
+     * worker's real {@code fleet_reply} lands — it is never stranded {@code PENDING}.
+     */
+    /**
+     * fleetd #334 gated the ask-timeout teardown on {@code ticket.fresh()}, matching the {@code
+     * finally} block that already did. This pins that gate. A coalesced duplicate passes its own
+     * {@code timeoutMillis}, which says nothing about whether the shared ask is done — so a
+     * duplicate timing out first must leave the fresh owner's still-open ask answerable.
+     *
+     * <p>Measured on merge: without this test, removing the {@code ticket.fresh()} gate left all
+     * 1371 tests green. The gate shipped with the reorder and nothing held it there.
+     *
+     * <p>What this does not prove: anything about the ordering inside the gate — that is
+     * {@code aLateAnswerDuringAskTimeoutTeardownStillCompletesTheAsyncTicket}'s job.
+     */
+    @Test
+    void aCoalescedDuplicateAskTimingOutLeavesTheFreshOwnersAskOpen() throws Exception {
+        String ticket = messages.sendAsync(T, "long task");
+        awaitWaiting();
+        injector.onStatus(T, AgentStatus.IDLE);    // deliver
+        injector.onStatus(T, AgentStatus.WORKING); // worker picks it up, then pauses to ask
+
+        CompletableFuture<MessageService.AskResult> fresh =
+                CompletableFuture.supplyAsync(() -> messages.ask(T, "which config file?", 5000));
+
+        MessageService.TaskView asking = null;
+        long deadline = System.currentTimeMillis() + 2000;
+        while ((asking == null || asking.phase() != MessageService.Phase.ASKING)
+                && System.currentTimeMillis() < deadline) {
+            asking = messages.poll(ticket);
+            //noinspection BusyWait
+            Thread.sleep(5);
+        }
+        assertNotNull(asking, "the fresh owner's question must surface before the duplicate asks");
+        String turnId = asking.turnId();
+        assertNotNull(turnId, "an ASKING view carries the turnId to answer on");
+
+        // A coalesced duplicate on the same session, with its own much shorter timeout.
+        MessageService.AskResult duplicate = messages.ask(T, "which config file?", 100);
+        assertEquals(MessageService.AskOutcome.TIMED_OUT, duplicate.outcome(),
+                "the duplicate's own timeout elapses first");
+
+        CompletableFuture<MessageService.Reply> answered =
+                CompletableFuture.supplyAsync(() -> messages.answer(turnId, "fleetd.yaml", 500));
+
+        MessageService.AskResult a = fresh.get(5, TimeUnit.SECONDS);
+        assertEquals(MessageService.AskOutcome.ANSWERED, a.outcome(),
+                "a duplicate's timeout must not lapse the ask the fresh owner still holds");
+        assertEquals("fleetd.yaml", a.answer());
+        assertFalse(answered.get(5, TimeUnit.SECONDS).outcome() == MessageService.Outcome.STALE_TURN,
+                "the answer must not be rejected as stale");
+    }
+
+    @Test
+    void aLateAnswerDuringAskTimeoutTeardownStillCompletesTheAsyncTicket() throws Exception {
+        String ticket = messages.sendAsync(T, "long task");
+        awaitWaiting();
+        injector.onStatus(T, AgentStatus.IDLE);    // deliver
+        injector.onStatus(T, AgentStatus.WORKING); // worker picks it up, then pauses to ask
+
+        CompletableFuture<MessageService.AskResult> ask =
+                CompletableFuture.supplyAsync(() -> messages.ask(T, "which config file?", 200));
+
+        // Wait for the question to actually surface (poll sees ASKING) before racing the timeout.
+        MessageService.TaskView asking = null;
+        long deadline = System.currentTimeMillis() + 2000;
+        while ((asking == null || asking.phase() != MessageService.Phase.ASKING)
+                && System.currentTimeMillis() < deadline) {
+            asking = messages.poll(ticket);
+            //noinspection BusyWait
+            Thread.sleep(5);
+        }
+        assertNotNull(asking, "the question must surface before the ask times out");
+        String turnId = asking.turnId();
+        assertNotNull(turnId, "an ASKING view carries the turnId to answer on");
+
+        CompletableFuture<MessageService.Reply> lateAnswer = new CompletableFuture<>();
+        messages.setAskTimeoutRaceHookForTest(() ->
+                lateAnswer.complete(messages.answer(turnId, "too late", 500)));
+        try {
+            MessageService.AskResult a = ask.get(5, TimeUnit.SECONDS);
+            assertEquals(MessageService.AskOutcome.TIMED_OUT, a.outcome());
+
+            MessageService.Reply late = lateAnswer.get(5, TimeUnit.SECONDS);
+            assertEquals(MessageService.Outcome.STALE_TURN, late.outcome(),
+                    "a late answer racing the timeout teardown must see the ask as already lapsed");
+
+            // The worker resumes on its own (per the ask() contract) and eventually sends its real
+            // fleet_reply; the async ticket must still resolve with it, not strand at PENDING.
+            assertTrue(messages.reply(T, "real result"), "the worker's real reply must still be accepted");
+        } finally {
+            messages.setAskTimeoutRaceHookForTest(null);
+        }
+
+        MessageService.TaskView done = null;
+        deadline = System.currentTimeMillis() + 2000;
+        while ((done == null || done.phase() == MessageService.Phase.PENDING)
+                && System.currentTimeMillis() < deadline) {
+            done = messages.poll(ticket);
+            //noinspection BusyWait
+            Thread.sleep(5);
+        }
+        assertNotNull(done);
+        assertEquals(MessageService.Phase.DONE, done.phase(), "the async ticket must not be stranded PENDING");
+        assertEquals("real result", done.reply());
+    }
+
    @Test
    void answeringAnUnknownTurnIsStale() {
        MessageService.Reply r = messages.answer(T + "#999", "too late", 500);
@@ -405,6 +530,29 @@ class MessageServiceTest {
                "a delivered send whose worker never replies times out as still working");
    }

+    /**
+     * fleetd #345. This forces the injector to pick up the exact pending delivery after {@code send}
+     * first observes its completion as incomplete, but before {@code cancel} takes the target monitor.
+     * The timeout must use {@link Injector.Cancellation#DELIVERED} from {@code cancel} and report
+     * {@link MessageService.Outcome#TIMED_OUT_WORKING}, because the text landed.
+     *
+     * <p>What this does not prove: that this precise interleaving happens by itself under production
+     * timing. The test forces it through a test-only hook; it proves the timeout caller handles the
+     * injector result when the interleaving occurs.
+     */
+    @Test
+    void sendTimeoutUsesCancellationDeliveredWhenPickupWinsTheRace() {
+        messages.setTimeoutCancellationRaceHookForTest(() -> injector.onStatus(T, AgentStatus.IDLE));
+        try {
+            MessageService.Reply reply = messages.send(T, "race delivery", 50);
+
+            assertEquals(MessageService.Outcome.TIMED_OUT_WORKING, reply.outcome(),
+                    "cancel reporting DELIVERED means the worker received the timed-out message");
+        } finally {
+            messages.setTimeoutCancellationRaceHookForTest(null);
+        }
+    }
+
    @Test
    void answerTimesOutWhenTheResumedWorkerNeverReplies() throws Exception {
        CompletableFuture<MessageService.Reply> send = sendAsync();
@@ -780,6 +928,49 @@ class MessageServiceTest {
        assertFailedTicket(third, "agent target term_a not found");
    }

+    // --- fleetd #335 (site 1): a per-task cleanup failure inside the abandon() loop must not -----
+    // strand the tasks that come after it. abandon()'s own comment on the loop documents the one
+    // real production call that can throw there (inbox.publish, in the stranded-reply put-back
+    // branch, reached when a concurrent reply() or a second abandon() races this one) — but
+    // reaching that branch requires the exact combination the #137 follow-up above already found
+    // unreachable through the public API. abandonCleanupHookForTest reproduces the resulting SHAPE
+    // (one task's cleanup throws) directly instead, the same technique this file already uses for
+    // fleetd #324/#329's own hard-to-time races.
+    @Test
+    void aPerTaskCleanupFailureDoesNotStrandTheRemainingMatchingTasks() throws Exception {
+        ListAppender<ILoggingEvent> appender = attachMessageServiceLog();
+        try {
+            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
+            String third = messages.sendAsync(T, "third task"); // parked too — the whole sweep must survive
+
+            java.util.concurrent.atomic.AtomicInteger calls = new java.util.concurrent.atomic.AtomicInteger();
+            messages.setAbandonCleanupHookForTest(() -> {
+                if (calls.getAndIncrement() == 0) {
+                    throw new RuntimeException("PROBE-335-SITE1");
+                }
+            });
+
+            assertTrue(messages.abandon(T, "agent target term_a not found"));
+
+            // Every task in the loop still gets its own outcome — the one whose cleanup threw
+            // included — even though the loop had no way to know in advance which one that would be.
+            assertFailedTicket(first, "agent target term_a not found");
+            assertFailedTicket(second, "agent target term_a not found");
+            assertFailedTicket(third, "agent target term_a not found");
+
+            assertTrue(appender.list.stream().anyMatch(e ->
+                            e.getLevel() == Level.ERROR
+                                    && e.getThrowableProxy() != null
+                                    && "PROBE-335-SITE1".equals(e.getThrowableProxy().getMessage())),
+                    "a per-task cleanup failure must still reach the log, not vanish silently");
+        } finally {
+            messages.setAbandonCleanupHookForTest(null);
+            detachMessageServiceLog(appender);
+        }
+    }
+
    // --- #137 follow-up: abandon() must not guess when more than one task is open ---------------
    //
    // A test combining a genuine stranded reply (hasStrandedReply(T)==true) with two simultaneously
@@ -1041,6 +1232,105 @@ class MessageServiceTest {
                "the reply completed its own ticket directly and never touched the inbox");
    }

+    /**
+     * fleetd #329 (F1). {@code answer()} completes the async ticket by looking {@code turnId} up in
+     * {@code asyncTasksByTurn} a SECOND time (the first is at :991, purely to re-register the
+     * {@code asyncTasksByWaiter} entry for #282's chained-ask case). That second lookup races
+     * {@code ask()}'s own unlocked timeout cleanup ({@code clearAsyncQuestion(turnId, true)}, run from
+     * {@code markAskTimedOut} + forgetting): {@code ask()}'s {@code ticket.answer().get(timeoutMillis)}
+     * can time out at essentially the same instant {@code answer()}'s {@code rendezvous.answerAsk}
+     * call above already succeeded and unblocked the worker. fleetd #324's single-read fix does not
+     * help here — that fixed a torn read of one already-held {@link MessageService} internal
+     * {@code Task}; this is a second, independent map lookup by a caller that no longer holds the
+     * {@code Task} it already found once.
+     *
+     * <p>This test does not wait for that real race to land on its own schedule — it drives the exact
+     * sequence the ticket describes (worker asks, primary answers, worker's real reply arrives) and
+     * fires the identical production cleanup {@code ask()}'s timeout path runs
+     * ({@code clearAsyncQuestion(turnId, true)}, via the {@code forgetTurnForTest} seam fleetd #324
+     * already merged) at the point between "the worker's own {@code ask()} call has unblocked" and
+     * "the worker's real {@code fleet_reply} arrives" — the exact window fleetd #329 names.
+     *
+     * <p>What this proves: given that exact interleaving, the async ticket must still resolve to the
+     * worker's real reply, not stay stuck {@link MessageService.Phase#PENDING} forever. What it does
+     * not prove: that the interleaving itself is reachable in production on its own timing — that is
+     * established by reading the code (see the ticket's "path in"), not by this test, for the same
+     * reason fleetd #324's own race test says so.
+     */
+    @Test
+    void aReplyRacingAsksTimeoutCleanupStillCompletesTheAsyncTicket() 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);
+        String turnId = asking.turnId();
+
+        CompletableFuture<MessageService.Reply> answer =
+                CompletableFuture.supplyAsync(() -> messages.answer(turnId, "config.yaml", 5000));
+        assertEquals("config.yaml", ask.get(5, TimeUnit.SECONDS).answer(),
+                "the worker's own ask() call must have already unblocked with the primary's answer "
+                        + "before we force the race below");
+
+        // ask()'s own unlocked timeout cleanup can forget this exact turnId at essentially the same
+        // instant answer() has already unblocked the worker and is now waiting on the resumed turn's
+        // real reply — reproduce that interleaving directly instead of trying to win a real race.
+        messages.forgetTurnForTest(turnId);
+
+        assertTrue(messages.reply(T, "PR opened: https://example/pulls/42"));
+
+        assertEquals(MessageService.Outcome.REPLIED, answer.get(5, TimeUnit.SECONDS).outcome(),
+                "the primary's own answer() call must still see the worker's real reply");
+        MessageService.TaskView done = awaitTicketPhase(ticket, MessageService.Phase.DONE);
+        assertEquals("PR opened: https://example/pulls/42", done.reply(),
+                "fleet_poll{ticket} must return the worker's real reply, not stay PENDING forever "
+                        + "just because ask()'s timeout cleanup forgot this turnId first");
+    }
+
+    /**
+     * fleetd #329 (F3). {@link MessageService#reply} reads the same orphan {@code Task}'s {@code
+     * turnId} twice on this path (once to check it is non-null, once as the {@code
+     * asyncTasksByTurn.remove} key) — the exact double-read shape fleetd #324 fixed in {@code
+     * finishAsyncTask}. This test drives the same real sequence as {@code
+     * aReplyAfterAnswerTimesOutStillCompletesTheAsyncTicket} (worker asks, primary answers, the
+     * primary's own bounded wait for the resumed turn expires) to reach a task with {@code turnId}
+     * genuinely stamped and no live rendezvous waiter open — the state {@code askAnsweredAsyncTasks}
+     * matches here — then fires the identical production cleanup {@code ask()}'s own timeout path
+     * runs ({@code clearAsyncQuestion(turnId, true)}, via the {@code forgetTurnForTest} seam fleetd
+     * #324 merged) at the point between the check and the removal use, via a dedicated test hook
+     * mirroring {@code finishAsyncTaskRaceHook}.
+     */
+    @Test
+    void replyToAnOrphanedTaskSurvivesTurnIdGoingNullBetweenItsTwoReads() 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);
+        String turnId = asking.turnId();
+
+        MessageService.Reply answerReply = messages.answer(turnId, "config.yaml", 150);
+        assertEquals("config.yaml", ask.get(5, TimeUnit.SECONDS).answer());
+        assertEquals(MessageService.Outcome.TIMED_OUT_WORKING, answerReply.outcome(),
+                "the primary's own bounded wait must give up first, leaving turnId stamped with no "
+                        + "live waiter — the state reply()'s F3 code path matches");
+
+        messages.setReplyOrphanTurnIdRaceHookForTest(() -> messages.forgetTurnForTest(turnId));
+        try {
+            assertTrue(messages.reply(T, "PR opened: https://example/pulls/42"));
+
+            MessageService.TaskView done = awaitTicketPhase(ticket, MessageService.Phase.DONE);
+            assertEquals("PR opened: https://example/pulls/42", done.reply(),
+                    "the ticket must still resolve to the worker's real reply despite the forced race");
+        } finally {
+            messages.setReplyOrphanTurnIdRaceHookForTest(null);
+        }
+    }
+
    /**
     * fleetd #307's ambiguity guard: an ask timeout frees its target ({@code hasAsyncQuestion}
     * becomes false the instant it lapses — proven above), so a second, independent delegation can
@@ -1151,6 +1441,63 @@ class MessageServiceTest {
        assertEquals("done", done.reply());
    }

+    // --- fleetd #329 (F2): an exception after the ticket's future completes must reach a log -----
+    //
+    // sendAsync's own executor task ends with catch (Throwable t) { task.future.completeExceptionally(t); }
+    // — but finishAsyncTask completes that same future on its first line, so anything that throws
+    // afterward hits an already-completed future: completeExceptionally returns false and does
+    // nothing, and (before this fix) nothing logged it either. Measured on the ticket: temporarily
+    // reintroducing the fleetd #324 NPE reproduced 19 real exceptions on the ordinary path with
+    // 74/74 tests staying green and zero log lines. There is no reachable production call site where
+    // finishAsyncTask throws after completing the future (fleetd #324 already closed the one that
+    // used to), so this test drives the shape directly via a dedicated test-only hook
+    // (afterFinishAsyncTaskCompleteHookForTest) rather than trying to engineer a real exception into
+    // that narrow window — the same technique fleetd #324's own race test and this ticket's F1/F3
+    // tests use for their own races.
+
+    private static ListAppender<ILoggingEvent> attachMessageServiceLog() {
+        Logger logger = (Logger) LoggerFactory.getLogger(MessageService.class);
+        ListAppender<ILoggingEvent> appender = new ListAppender<>();
+        appender.start();
+        logger.addAppender(appender);
+        return appender;
+    }
+
+    private static void detachMessageServiceLog(ListAppender<ILoggingEvent> appender) {
+        ((Logger) LoggerFactory.getLogger(MessageService.class)).detachAppender(appender);
+    }
+
+    @Test
+    void anExceptionAfterTheTicketFutureCompletesStillReachesTheLog() throws Exception {
+        ListAppender<ILoggingEvent> appender = attachMessageServiceLog();
+        try {
+            messages.setAfterFinishAsyncTaskCompleteHookForTest(() -> {
+                throw new RuntimeException("PROBE-329-F2");
+            });
+
+            String ticket = messages.sendAsync(T, "do the task");
+            awaitWaiting();
+            assertTrue(rendezvous.resolve(T, "done"));
+
+            // finishAsyncTask's own first line already completed the future with the real result
+            // BEFORE the hook threw — F2 is a visibility gap, not a correctness gap for the ticket
+            // itself, so the ticket's own outcome must be unaffected by the swallowed exception.
+            MessageService.TaskView done = awaitTicketPhase(ticket, MessageService.Phase.DONE);
+            assertEquals("done", done.reply());
+
+            assertTrue(appender.list.stream().anyMatch(e ->
+                            e.getLevel() == Level.ERROR
+                                    && e.getFormattedMessage().contains(ticket)
+                                    && e.getThrowableProxy() != null
+                                    && "PROBE-329-F2".equals(e.getThrowableProxy().getMessage())),
+                    "an exception thrown after the ticket's future already completed must still "
+                            + "reach the log, not vanish silently");
+        } finally {
+            messages.setAfterFinishAsyncTaskCompleteHookForTest(null);
+            detachMessageServiceLog(appender);
+        }
+    }
+
    // --- CB-582: fleet_status pendingAsk() ------------------------------------------------------

    @Test
@@ -1311,6 +1658,42 @@ class MessageServiceTest {
        }
    }

+    // --- fleetd #335 (site 2): task.future.whenComplete's own returned stage is discarded, so an --
+    // uncaught throw from ReplyPushLoop.onTicketTerminal used to vanish with no log line and no
+    // metric. The real production trigger is the daemon's own shutdown sequence (Fleetd's shutdown
+    // hook): messages.close() only stops the async executor from taking NEW work — it does not
+    // cancel a send already in flight — while pushLoop.close() shuts its scheduler down immediately
+    // right after, so a ticket that completes in that narrow window has onTicketTerminal's own
+    // scheduler.schedule(...) throw a real RejectedExecutionException. Reproduced here by shutting
+    // the very same scheduler down before the ticket resolves — no test-only hook needed, this
+    // reachable path throws for real.
+    @Test
+    void aTicketTerminalPushFailureDoesNotVanishSilently() throws Exception {
+        ListAppender<ILoggingEvent> appender = attachMessageServiceLog();
+        try (var wiring = wireWithPushLoop(1, 50)) {
+            String ticket = wiring.service().sendAsync(T, "long task");
+            awaitWaiting();
+            wiring.scheduler().shutdownNow(); // simulate pushLoop.close() racing an in-flight send
+            injectDelivery();
+            assertTrue(rendezvous.resolve(T, "async result"));
+
+            // The ticket's own outcome must be unaffected by the swallowed exception — finishAsyncTask
+            // completes task.future before whenComplete's action (and thus onTicketTerminal) ever runs.
+            MessageService.TaskView done = awaitTicketPhaseOn(wiring.service(), ticket, MessageService.Phase.DONE);
+            assertEquals("async result", done.reply());
+
+            assertTrue(appender.list.stream().anyMatch(e ->
+                            e.getLevel() == Level.ERROR
+                                    && e.getFormattedMessage().contains(ticket)
+                                    && e.getThrowableProxy() != null
+                                    && "java.util.concurrent.RejectedExecutionException"
+                                            .equals(e.getThrowableProxy().getClassName())),
+                    "onTicketTerminal throwing must still reach the log, not vanish silently");
+        } finally {
+            detachMessageServiceLog(appender);
+        }
+    }
+
    // --- CB-582: fleet_ask question-open nudges --------------------------------------------------

    @Test
@@ -1607,7 +1990,10 @@ 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
@@ -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
@@ -116,6 +116,50 @@ check_log_path_matches_plist() {
  ok "log path check: script and plist agree ($resolved_out)"
 }

+# Classify ERROR lines in one fresh log region. AMQP failure messages now include the connection
+# name, so a recovery can clear only errors for its own connection. A candidate with neither name
+# remains unexplained: it must never be quieted by a recovery on the other connection.
+classify_amqp_connection_errors() {
+  local log_file="$1" line pending_inbox=0 pending_lead_mailbox=0
+  REDEPLOY_ERROR_COUNT=0
+  REDEPLOY_RECOVERED_AMQP_ERRORS=0
+  REDEPLOY_UNEXPLAINED_ERRORS=0
+
+  while IFS= read -r line || [ -n "$line" ]; do
+    case "$line" in
+      *' ERROR '*|*' SEVERE '*)
+        REDEPLOY_ERROR_COUNT=$((REDEPLOY_ERROR_COUNT + 1))
+        case "$line" in
+          *'AMQP connection fleetd-reply-inbox: An unexpected connection driver error occurred'*|*'AMQP connection fleetd-reply-inbox: Caught an exception during connection recovery!'*)
+            pending_inbox=$((pending_inbox + 1))
+            ;;
+          *'AMQP connection fleetd-lead-mailbox: An unexpected connection driver error occurred'*|*'AMQP connection fleetd-lead-mailbox: Caught an exception during connection recovery!'*)
+            pending_lead_mailbox=$((pending_lead_mailbox + 1))
+            ;;
+          *'AMQP connection'*'An unexpected connection driver error occurred'*|*'AMQP connection'*'Caught an exception during connection recovery!'*)
+            REDEPLOY_UNEXPLAINED_ERRORS=$((REDEPLOY_UNEXPLAINED_ERRORS + 1))
+            ;;
+          *) REDEPLOY_UNEXPLAINED_ERRORS=$((REDEPLOY_UNEXPLAINED_ERRORS + 1)) ;;
+        esac
+        ;;
+      *'AMQP connection recovered; cleared held replies for fresh redelivery'*)
+        if [ "$pending_inbox" -gt 0 ]; then
+          pending_inbox=$((pending_inbox - 1))
+          REDEPLOY_RECOVERED_AMQP_ERRORS=$((REDEPLOY_RECOVERED_AMQP_ERRORS + 1))
+        fi
+        ;;
+      *'AMQP lead mailbox connection recovered; cleared held messages for fresh redelivery'*)
+        if [ "$pending_lead_mailbox" -gt 0 ]; then
+          pending_lead_mailbox=$((pending_lead_mailbox - 1))
+          REDEPLOY_RECOVERED_AMQP_ERRORS=$((REDEPLOY_RECOVERED_AMQP_ERRORS + 1))
+        fi
+        ;;
+    esac
+  done < "$log_file"
+
+  REDEPLOY_UNEXPLAINED_ERRORS=$((REDEPLOY_UNEXPLAINED_ERRORS + pending_inbox + pending_lead_mailbox))
+}
+
 # CB-600: sourceable for testing. When this file is SOURCED (not executed) it stops here — nothing
 # below runs — so a test harness can `source` it to call check_log_path_matches_plist (or the
 # other pure helpers above) against a throwaway plist fixture without ever reaching the mutating
@@ -361,15 +405,21 @@ tail -n "+$((RESTART_MARK + 1))" "$OUT" 2>/dev/null \
  | grep -iE 'deferred|classification:|fleet health:|coverage' | tail -8 | sed 's/^/   /' \
  || echo "   (nothing reported)"

-# Errors since the restart, anchored to the marker so old noise cannot leak in.
-ERRS="$(tail -n "+$((RESTART_MARK + 1))" "$OUT" 2>/dev/null | grep -cE ' (ERROR|SEVERE) ' || true)"
+# Errors since the restart, anchored to the marker so old noise cannot leak in. Keep the fresh
+# region in a file because the classifier must preserve the order of errors and recoveries.
+FRESH_LOG="$(mktemp -t fleetd-fresh-log)"
+trap 'rm -f "$FRESH_LOG"' EXIT
+tail -n "+$((RESTART_MARK + 1))" "$OUT" > "$FRESH_LOG" 2>/dev/null || true
+classify_amqp_connection_errors "$FRESH_LOG"
 say "result"
 ok "pid $NEW_PID, jar $(jar_id)"
-if [ "${ERRS:-0}" -gt 0 ]; then
-  warn "$ERRS ERROR lines since restart:"
-  tail -n "+$((RESTART_MARK + 1))" "$OUT" | grep -E ' (ERROR|SEVERE) ' | tail -5 | sed 's/^/   /'
-else
+if [ "$REDEPLOY_ERROR_COUNT" -eq 0 ]; then
  ok "no ERROR lines since restart"
+elif [ "$REDEPLOY_UNEXPLAINED_ERRORS" -eq 0 ]; then
+  ok "$REDEPLOY_RECOVERED_AMQP_ERRORS AMQP connection reset ERROR lines recovered since restart"
+else
+  warn "$REDEPLOY_ERROR_COUNT ERROR lines since restart:"
+  grep -E ' (ERROR|SEVERE) ' "$FRESH_LOG" | tail -5 | sed 's/^/   /'
 fi
 echo
 echo "   Next: call fleet_whoami and confirm it still answers 'primary'. A lead whose tab label"
@@ -0,0 +1,238 @@
+#!/usr/bin/env bash
+# Self-contained checks for the pure log classifier in redeploy-fleetd.sh.
+
+set -euo pipefail
+
+ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+TMP="$(mktemp -d "$ROOT/.redeploy-log-test.XXXXXX")"
+trap 'rm -rf "$TMP"' EXIT
+
+# Sourcing stops before redeploy-fleetd.sh can build, stop, or start the daemon.
+source "$ROOT/scripts/redeploy-fleetd.sh"
+
+fail() {
+  printf 'FAIL: %s\n' "$*" >&2
+  return 1
+}
+
+assert_equals() {
+  local expected="$1" actual="$2" description="$3"
+  [ "$expected" = "$actual" ] || fail "$description: expected $expected, got $actual"
+}
+
+classify_fixture() {
+  local name="$1"
+  classify_amqp_connection_errors "$TMP/$name"
+}
+
+test_no_errors() {
+  cat > "$TMP/no-errors.log" <<'LOG'
+2026-09-05 12:00:00 INFO fleetd listening
+LOG
+  classify_fixture no-errors.log
+  assert_equals 0 "$REDEPLOY_ERROR_COUNT" "no-errors total"
+  assert_equals 0 "$REDEPLOY_UNEXPLAINED_ERRORS" "no-errors unexplained"
+}
+
+test_recovery_patterns_match_source() {
+  grep -F 'AMQP connection {}: {}' "$ROOT/fleetd/src/main/java/dev/ltms/fleet/msg/AmqpReplyInbox.java" > /dev/null \
+    || fail "AMQP failure pattern no longer matches source"
+  grep -F 'AMQP connection recovered; cleared held replies for fresh redelivery' \
+    "$ROOT/fleetd/src/main/java/dev/ltms/fleet/msg/AmqpReplyInbox.java" > /dev/null \
+    || fail "reply-inbox recovery pattern no longer matches source"
+  grep -F 'AMQP lead mailbox connection recovered; cleared held messages for fresh redelivery' \
+    "$ROOT/fleetd/src/main/java/dev/ltms/fleet/msg/LeadMailbox.java" > /dev/null \
+    || fail "lead-mailbox recovery pattern no longer matches source"
+}
+
+test_attributed_recovered_connection_error() {
+  cat > "$TMP/attributed-recovered.log" <<'LOG'
+2026-09-05 12:00:00 ERROR [AMQP Connection broker:5672] d.l.fleet.msg.AmqpReplyInbox - AMQP connection fleetd-reply-inbox: An unexpected connection driver error occurred
+2026-09-05 12:00:01 INFO [AMQP Connection broker:5672] d.l.fleet.msg.AmqpReplyInbox - AMQP connection recovered; cleared held replies for fresh redelivery
+LOG
+  classify_fixture attributed-recovered.log
+  assert_equals 1 "$REDEPLOY_ERROR_COUNT" "attributed-recovered total"
+  assert_equals 1 "$REDEPLOY_RECOVERED_AMQP_ERRORS" "attributed-recovered errors"
+  assert_equals 0 "$REDEPLOY_UNEXPLAINED_ERRORS" "attributed-recovered unexplained"
+}
+
+test_source_derived_error_shapes_recover_by_connection() {
+  # These ERROR shapes come from AmqpConnectionFailureLogger on main. They need a live-log check
+  # after redeploy because the new code has not yet written a production line.
+  cat > "$TMP/source-derived.log" <<'LOG'
+17:37:53.537 ERROR [AMQP Connection broker:5672] d.l.fleet.msg.AmqpReplyInbox - AMQP connection fleetd-reply-inbox: An unexpected connection driver error occurred
+17:37:54.537 ERROR [AMQP Connection broker:5672] d.l.fleet.msg.AmqpReplyInbox - AMQP connection fleetd-reply-inbox: Caught an exception during connection recovery!
+17:37:55.537 ERROR [AMQP Connection broker:5672] d.l.fleet.msg.AmqpReplyInbox - AMQP connection fleetd-reply-inbox: An unexpected connection driver error occurred
+17:37:56.537 ERROR [AMQP Connection broker:5672] d.ltms.fleet.msg.LeadMailbox - AMQP connection fleetd-lead-mailbox: An unexpected connection driver error occurred
+17:37:57.537 ERROR [AMQP Connection broker:5672] d.ltms.fleet.msg.LeadMailbox - AMQP connection fleetd-lead-mailbox: Caught an exception during connection recovery!
+17:37:58.537 ERROR [AMQP Connection broker:5672] d.ltms.fleet.msg.LeadMailbox - AMQP connection fleetd-lead-mailbox: An unexpected connection driver error occurred
+17:38:00.000 INFO  [AMQP Connection broker:5672] d.l.fleet.msg.AmqpReplyInbox - AMQP connection recovered; cleared held replies for fresh redelivery
+17:38:01.000 INFO  [AMQP Connection broker:5672] d.l.fleet.msg.AmqpReplyInbox - AMQP connection recovered; cleared held replies for fresh redelivery
+17:38:02.000 INFO  [AMQP Connection broker:5672] d.l.fleet.msg.AmqpReplyInbox - AMQP connection recovered; cleared held replies for fresh redelivery
+17:38:03.000 INFO  [AMQP Connection broker:5672] d.ltms.fleet.msg.LeadMailbox - AMQP lead mailbox connection recovered; cleared held messages for fresh redelivery
+17:38:04.000 INFO  [AMQP Connection broker:5672] d.ltms.fleet.msg.LeadMailbox - AMQP lead mailbox connection recovered; cleared held messages for fresh redelivery
+17:38:05.000 INFO  [AMQP Connection broker:5672] d.ltms.fleet.msg.LeadMailbox - AMQP lead mailbox connection recovered; cleared held messages for fresh redelivery
+LOG
+  classify_fixture source-derived.log
+  assert_equals 6 "$REDEPLOY_ERROR_COUNT" "source-derived total"
+  assert_equals 6 "$REDEPLOY_RECOVERED_AMQP_ERRORS" "source-derived recovered"
+  assert_equals 0 "$REDEPLOY_UNEXPLAINED_ERRORS" "source-derived unexplained"
+}
+
+test_cross_connection_unattributable_errors_stay_loud() {
+  # This candidate has neither stable connection name, so LeadMailbox recovery must not consume it.
+  cat > "$TMP/cross-unattributable.log" <<'LOG'
+2026-09-05 12:00:00 ERROR [AMQP Connection broker:5672] unknown - AMQP connection: An unexpected connection driver error occurred
+2026-09-05 12:00:01 ERROR [AMQP Connection broker:5672] unknown - AMQP connection: An unexpected connection driver error occurred
+2026-09-05 12:00:02 INFO  [AMQP Connection 10.10.20.13:5672] d.ltms.fleet.msg.LeadMailbox - AMQP lead mailbox connection recovered; cleared held messages for fresh redelivery
+2026-09-05 12:00:03 INFO  [AMQP Connection 10.10.20.13:5672] d.ltms.fleet.msg.LeadMailbox - AMQP lead mailbox connection recovered; cleared held messages for fresh redelivery
+LOG
+  classify_fixture cross-unattributable.log
+  assert_equals 2 "$REDEPLOY_ERROR_COUNT" "cross-unattributable total"
+  assert_equals 0 "$REDEPLOY_RECOVERED_AMQP_ERRORS" "cross-unattributable recovered"
+  assert_equals 2 "$REDEPLOY_UNEXPLAINED_ERRORS" "cross-unattributable unexplained"
+}
+
+test_attributed_cross_connection_errors_stay_loud() {
+  # LeadMailbox recovery cannot heal AmqpReplyInbox errors.
+  cat > "$TMP/cross-attributed.log" <<'LOG'
+2026-09-05 12:00:00 ERROR AmqpReplyInbox - AMQP connection fleetd-reply-inbox: An unexpected connection driver error occurred
+2026-09-05 12:00:01 ERROR AmqpReplyInbox - AMQP connection fleetd-reply-inbox: An unexpected connection driver error occurred
+2026-09-05 12:00:02 INFO LeadMailbox - AMQP lead mailbox connection recovered; cleared held messages for fresh redelivery
+2026-09-05 12:00:03 INFO LeadMailbox - AMQP lead mailbox connection recovered; cleared held messages for fresh redelivery
+LOG
+  classify_fixture cross-attributed.log
+  assert_equals 2 "$REDEPLOY_ERROR_COUNT" "cross-attributed total"
+  assert_equals 0 "$REDEPLOY_RECOVERED_AMQP_ERRORS" "cross-attributed recovered"
+  assert_equals 2 "$REDEPLOY_UNEXPLAINED_ERRORS" "cross-attributed unexplained"
+}
+
+test_attributed_unrecovered_connection_error() {
+  cat > "$TMP/unrecovered.log" <<'LOG'
+2026-09-05 12:00:00 ERROR AmqpReplyInbox - AMQP connection fleetd-reply-inbox: An unexpected connection driver error occurred
+LOG
+  classify_fixture unrecovered.log
+  assert_equals 1 "$REDEPLOY_ERROR_COUNT" "unrecovered total"
+  assert_equals 0 "$REDEPLOY_RECOVERED_AMQP_ERRORS" "unrecovered AMQP errors"
+  assert_equals 1 "$REDEPLOY_UNEXPLAINED_ERRORS" "unrecovered unexplained"
+}
+
+test_other_error_is_unexplained() {
+  cat > "$TMP/other-error.log" <<'LOG'
+2026-09-05 12:00:00 ERROR dev.ltms.fleet.Fleetd - startup failed
+2026-09-05 12:00:01 INFO dev.ltms.fleet.msg.AmqpReplyInbox - AMQP connection recovered; cleared held replies for fresh redelivery
+LOG
+  classify_fixture other-error.log
+  assert_equals 1 "$REDEPLOY_ERROR_COUNT" "other-error total"
+  assert_equals 1 "$REDEPLOY_UNEXPLAINED_ERRORS" "other-error unexplained"
+}
+
+test_recovery_requirement_mutation_is_caught() {
+  classify_amqp_connection_errors() {
+    local log_file="$1" line
+    REDEPLOY_ERROR_COUNT=0
+    REDEPLOY_RECOVERED_AMQP_ERRORS=0
+    REDEPLOY_UNEXPLAINED_ERRORS=0
+    while IFS= read -r line || [ -n "$line" ]; do
+      case "$line" in
+        *' ERROR '*|*' SEVERE '*)
+          REDEPLOY_ERROR_COUNT=$((REDEPLOY_ERROR_COUNT + 1))
+          case "$line" in
+            *'AMQP connection fleetd-reply-inbox: An unexpected connection driver error occurred'*)
+              REDEPLOY_RECOVERED_AMQP_ERRORS=$((REDEPLOY_RECOVERED_AMQP_ERRORS + 1))
+              ;;
+            *) REDEPLOY_UNEXPLAINED_ERRORS=$((REDEPLOY_UNEXPLAINED_ERRORS + 1)) ;;
+          esac
+          ;;
+      esac
+    done < "$log_file"
+  }
+
+  if test_attributed_unrecovered_connection_error > "$TMP/mutation-output" 2>&1; then
+    fail "mutation accepted an unrecovered connection error"
+  fi
+  grep -F 'FAIL: unrecovered AMQP errors: expected 0, got 1' "$TMP/mutation-output" > /dev/null \
+    || fail "mutation failed without the expected assertion"
+  printf 'Recovery mutation: FAIL: unrecovered AMQP errors: expected 0, got 1\n'
+}
+
+test_shared_counter_mutation_is_caught() {
+  classify_amqp_connection_errors() {
+    local log_file="$1" line pending=0
+    REDEPLOY_ERROR_COUNT=0
+    REDEPLOY_RECOVERED_AMQP_ERRORS=0
+    REDEPLOY_UNEXPLAINED_ERRORS=0
+    while IFS= read -r line || [ -n "$line" ]; do
+      case "$line" in
+        *' ERROR '*|*' SEVERE '*)
+          REDEPLOY_ERROR_COUNT=$((REDEPLOY_ERROR_COUNT + 1))
+          case "$line" in
+            *'AMQP connection'*'An unexpected connection driver error occurred'*|*'AMQP connection'*'Caught an exception during connection recovery!'*)
+              case "$line" in
+                *'fleetd-reply-inbox'*|*'fleetd-lead-mailbox'*) pending=$((pending + 1)) ;;
+                *) REDEPLOY_UNEXPLAINED_ERRORS=$((REDEPLOY_UNEXPLAINED_ERRORS + 1)) ;;
+              esac
+              ;;
+            *) REDEPLOY_UNEXPLAINED_ERRORS=$((REDEPLOY_UNEXPLAINED_ERRORS + 1)) ;;
+          esac
+          ;;
+        *'AMQP connection recovered; cleared held replies for fresh redelivery'*|*'AMQP lead mailbox connection recovered; cleared held messages for fresh redelivery'*)
+          if [ "$pending" -gt 0 ]; then
+            pending=$((pending - 1))
+            REDEPLOY_RECOVERED_AMQP_ERRORS=$((REDEPLOY_RECOVERED_AMQP_ERRORS + 1))
+          fi
+          ;;
+      esac
+    done < "$log_file"
+    REDEPLOY_UNEXPLAINED_ERRORS=$((REDEPLOY_UNEXPLAINED_ERRORS + pending))
+  }
+
+  if test_attributed_cross_connection_errors_stay_loud > "$TMP/shared-mutation-output" 2>&1; then
+    fail "shared counter mutation accepted cross-connection recovery"
+  fi
+  grep -F 'FAIL: cross-attributed recovered: expected 0, got 2' "$TMP/shared-mutation-output" > /dev/null \
+    || fail "shared counter mutation failed without the expected assertion"
+  printf 'Shared-counter mutation: FAIL: cross-attributed recovered: expected 0, got 2\n'
+}
+
+test_unattributable_quiet_mutation_is_caught() {
+  classify_amqp_connection_errors() {
+    local log_file="$1" line
+    REDEPLOY_ERROR_COUNT=0
+    REDEPLOY_RECOVERED_AMQP_ERRORS=0
+    REDEPLOY_UNEXPLAINED_ERRORS=0
+    while IFS= read -r line || [ -n "$line" ]; do
+      case "$line" in
+        *' ERROR '*|*' SEVERE '*)
+          REDEPLOY_ERROR_COUNT=$((REDEPLOY_ERROR_COUNT + 1))
+          case "$line" in
+            *'AMQP connection'*'An unexpected connection driver error occurred'*|*'AMQP connection'*'Caught an exception during connection recovery!'*)
+              REDEPLOY_RECOVERED_AMQP_ERRORS=$((REDEPLOY_RECOVERED_AMQP_ERRORS + 1))
+              ;;
+            *) REDEPLOY_UNEXPLAINED_ERRORS=$((REDEPLOY_UNEXPLAINED_ERRORS + 1)) ;;
+          esac
+          ;;
+      esac
+    done < "$log_file"
+  }
+
+  if test_cross_connection_unattributable_errors_stay_loud > "$TMP/unattributable-mutation-output" 2>&1; then
+    fail "unattributable mutation accepted an unknown connection"
+  fi
+  grep -F 'FAIL: cross-unattributable recovered: expected 0, got 2' "$TMP/unattributable-mutation-output" > /dev/null \
+    || fail "unattributable mutation failed without the expected assertion"
+  printf 'Unattributable mutation: FAIL: cross-unattributable recovered: expected 0, got 2\n'
+}
+
+test_no_errors
+test_recovery_patterns_match_source
+test_attributed_recovered_connection_error
+test_source_derived_error_shapes_recover_by_connection
+test_cross_connection_unattributable_errors_stay_loud
+test_attributed_cross_connection_errors_stay_loud
+test_attributed_unrecovered_connection_error
+test_other_error_is_unexplained
+test_recovery_requirement_mutation_is_caught
+test_shared_counter_mutation_is_caught
+test_unattributable_quiet_mutation_is_caught
+printf 'PASS: redeploy log classifier\n'