fleetd#341: a per-name guard so a later spawn's different unprotected name still warns

unprotectedGapLogged was one AtomicBoolean guarding two WARN branches in
logCredentialGap that name different env var names (the allow-list
keptByDerivedList branch, and warnGapUnprotected's deny-by-default /
non-zsh-fallback branch). memberCredentials is a live, re-read-per-spawn
supplier, so between two spawns a policy reload can change which names are
in the gap: spawn 1 warns about name A and trips the shared flag, and
spawn 2's gap containing a different name B never gets its WARN.

Replace the AtomicBoolean with unprotectedGapNamesWarned, a
ConcurrentHashMap-backed Set<String> guard keyed per name (same shape as
OpenCodeLauncher.modelCheckSkippedWarned), so each distinct credential-shaped
name is warned about exactly once, ever, regardless of which branch or
which spawn first reports it. allowListGapLogged (the separate INFO guard,
#192) is untouched. Neither WARN's wording changed.

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

@@ -22,19 +22,19 @@ import java.util.function.Supplier;
  * choice rather than an accident of where the field was initialised.
  *
  * <h2>Not every key can change under a running daemon</h2>
- * Keys fall into three classes, and the difference is about what already exists when the reload
+ * Keys fall into four classes, and the difference is about what already exists when the reload
  * happens — not about how important the key is.
  *
  * <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}
@@ -71,35 +71,101 @@ 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>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>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}
+ *             and never rebuilt, so a changed value needs a restart to actually start, stop, or
+ *             retime it. The coverage string {@code fleet_profiles} reports
+ *             ({@code Fleetd.java:648-650}) is read live off {@link #get()} on every call, so it
+ *             already reflects the new value.</li>
+ *         <li>{@code coordinator:} — the {@code LeadMailbox} connection ({@code uri},
+ *             {@code uriEnv}, {@code selfId}, {@code prefetch}) is opened once at
+ *             {@code Fleetd.java:502} and never reopened, so a changed value needs a restart —
+ *             {@code selfId} in particular names this daemon's own AMQP inbox queue, and a peer
+ *             lead that learned the old name would not discover a new one on its own. The broker
+ *             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. 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 #326).</strong> {@code FleetConfig} has
- * 22 top-level record components. Four of them are named nowhere in this file, and the reason
- * differs per key, so do not read "absent" as "forgotten":
- * <ul>
- *   <li>{@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.</li>
- *   <li>{@code health} and {@code coordinator} are <strong>undecided</strong>, not hot. Each is read
- *       both ways at different sites, so neither fits the three classes above as a whole key. Until
- *       that is settled a reload touching them reports a bare "config reloaded", which under-claims.
- *       Deciding it is what a top-level coverage checker (the {@link ConfigRefProfileCoverageTest}
- *       shape, one level up) is waiting on — without it, such a checker cannot express the answer.</li>
- * </ul>
- * 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.
+ * <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. {@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. 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} 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.
  *
  * <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
  * is the worst thing a reload can do to an operator debugging one. Refusing keeps the invariant that
  * the live config is always some version of the file, and the message names the keys that must
- * change through a restart.
+ * change through a restart. A split change does <em>not</em> refuse, for a different reason than a
+ * deferred change does not: its live half genuinely took effect, so refusing would throw that away
+ * and leave the operator worse off than the partial-but-honest report {@link Outcome#split()} gives.
  *
  * <p>A reload that fails to parse or fails validation is also refused, and the previous config keeps
  * running. A config file being edited is normally read once mid-save; degrading a working daemon
@@ -109,10 +175,24 @@ public final class ConfigRef implements Supplier<FleetConfig> {
 
     private static final Logger log = LoggerFactory.getLogger(ConfigRef.class);
 
-    /** Keys that cannot change under a running daemon — see the class doc. */
-    private static final Set<String> COLD_KEYS =
+    /**
+     * Keys that cannot change under a running daemon — see the class doc.
+     *
+     * <p>Package-private (not {@code private}) so {@code ConfigRefTopLevelCoverageTest} can fold it
+     * into the top-level triage it checks, the same way it reads {@link #SPLIT_KEYS}.
+     */
+    static final Set<String> COLD_KEYS =
              Set.of("bind", "herdrSocket", "memberHerdrSocket", "broker", "auth");
 
+    /**
+     * Keys read BOTH off the startup snapshot and live off {@link #get()} at different sites, so
+     * neither the hot, deferred nor cold class fits them as a whole — see the class doc's Split
+     * bullet (fleetd #330). A changed split key is accepted ({@link Outcome#applied()} stays
+     * {@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", "fleet");
+
     private final Path path;
     private final AtomicReference<FleetConfig> current;
 
@@ -140,25 +220,37 @@ public final class ConfigRef implements Supplier<FleetConfig> {
     /**
      * What a reload attempt did.
      *
+     * <p>{@code split} is a separate field from {@code deferred} rather than a differently-worded
+     * entry inside it, because the two carry different guarantees for any caller that branches on
+     * them rather than just printing {@link #summary()}: every {@code deferred} entry means "this
+     * key's whole change waits for a restart", while every {@code split} entry means "part of this
+     * key's change already applied, and the message says which part" — collapsing them would force
+     * a caller to re-parse the message to tell those apart. See the class doc's Split bullet
+     * (fleetd #330) for why the key needs this at all.
+     *
      * @param applied  true when the new config is now live
      * @param coldKeys cold keys whose value changed, which is why an unapplied reload was refused
-     * @param deferred keys that changed and were accepted, but whose effect waits for a restart
+     * @param deferred keys that changed and were accepted, but whose effect waits entirely on a
+     *                 restart
+     * @param split    split keys that changed and were accepted, each named with which half of it
+     *                 is already live and which half waits for a restart
      * @param error    the parse or validation failure that refused the reload, else {@code null}
      */
     public record Outcome(boolean applied, List<String> coldKeys, List<String> deferred,
-                          String error) {
+                          List<String> split, String error) {
 
         public Outcome {
             coldKeys = List.copyOf(coldKeys);
             deferred = List.copyOf(deferred);
+            split = List.copyOf(split);
         }
 
         static Outcome refusedCold(List<String> keys) {
-            return new Outcome(false, keys, List.of(), null);
+            return new Outcome(false, keys, List.of(), List.of(), null);
         }
 
         static Outcome failed(String error) {
-            return new Outcome(false, List.of(), List.of(), error);
+            return new Outcome(false, List.of(), List.of(), List.of(), error);
         }
 
         /** A one-line summary for the operator — the reason, not just the verdict. */
@@ -170,11 +262,18 @@ public final class ConfigRef implements Supplier<FleetConfig> {
                 return "config reload refused — these keys cannot change under a running daemon: "
                         + String.join(", ", coldKeys) + ". Restart fleetd to apply them.";
             }
-            if (!deferred.isEmpty()) {
-                return "config reloaded; these changes need a restart to take effect: "
-                        + String.join(", ", deferred);
+            if (deferred.isEmpty() && split.isEmpty()) {
+                return "config reloaded";
             }
-            return "config reloaded";
+            StringBuilder out = new StringBuilder("config reloaded");
+            if (!deferred.isEmpty()) {
+                out.append("; these changes need a restart to take effect: ")
+                        .append(String.join(", ", deferred));
+            }
+            if (!split.isEmpty()) {
+                out.append("; partially live — ").append(String.join(" | ", split));
+            }
+            return out.toString();
         }
     }
 
@@ -215,14 +314,21 @@ public final class ConfigRef implements Supplier<FleetConfig> {
         }
 
         List<String> deferred = changedDeferredKeys(old, fresh);
+        List<String> split = changedSplitKeys(old, fresh);
         current.set(fresh);
-        Outcome out = new Outcome(true, List.of(), deferred, null);
+        Outcome out = new Outcome(true, List.of(), deferred, split, null);
         log.info(out.summary());
         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");
@@ -325,6 +431,73 @@ public final class ConfigRef implements Supplier<FleetConfig> {
         return changed;
     }
 
+    /**
+     * Split keys whose value differs between the running config and the candidate — see the class
+     * 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.
+     */
+    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 "
+                    + "frozen at startup and needs a restart; the coverage status fleet_profiles "
+                    + "reports is read live and already applied");
+        }
+        if (!Objects.equals(old.coordinator(), fresh.coordinator())) {
+            changed.add("coordinator: the LeadMailbox connection (uri, uriEnv, selfId, prefetch) is "
+                    + "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 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

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

@@ -1665,30 +1665,59 @@ 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 +1845,21 @@ 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 +1872,18 @@ public abstract class HerdrPeerLauncher implements PeerLauncher {
 
     /** The deny-by-default (and allow-list non-zsh fallback) WARN — unchanged byte-for-byte by #192. */
     private void warnGapUnprotected(List<String> gap) {
-        if (unprotectedGapLogged.compareAndSet(false, true)) {
+        // fleetd #341: same "only the names never warned before" filter as the sibling branch in
+        // logCredentialGap above — see unprotectedGapNamesWarned's javadoc. Both branches share
+        // this one guard because both report the exact same fact (a name reaching a member pane
+        // unprotected) at the exact same severity; keying it by name is what lets a later spawn's
+        // DIFFERENT name still get its own WARN after an earlier spawn's already fired.
+        List<String> newlyUnprotected = gap.stream().filter(unprotectedGapNamesWarned::add).toList();
+        if (!newlyUnprotected.isEmpty()) {
             log.warn("memberCredentials gap: {} credential-shaped env var name(s) are on neither "
                             + "known: nor allow: — every member pane inherits them UNBLOCKED — {}. "
                             + "Add each to memberCredentials.known (blocked by default) or .allow "
                             + "(if a member legitimately needs it).",
-                    gap.size(), gap);
+                    newlyUnprotected.size(), newlyUnprotected);
         }
     }
 

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

@@ -1037,9 +1037,22 @@ public final class MessageService {
                 // 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. When task is null (this was never an async
-                // ticket — a blocking fleet_ask's answer() call has no Task at all), there is nothing
-                // to complete, matching the old lookup-miss behaviour.
+                // 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. A genuine async ticket also lands here
+                // with task == null, because ask()'s timeout path runs clearAsyncQuestion(turnId,
+                // true) — which drops the asyncTasksByTurn entry — in its catch block, while
+                // rendezvous.closeAsk(turnId) runs later, in its finally. Between those two the ask
+                // is still answerable but the map entry is already gone, so the lookup at :991
+                // returns null and this ticket is never completed. Measured on 2026-09-04: a probe
+                // firing only that first half before answer() runs 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; that cannot change the outcome, because askTimedOut
+                // is read only by askAnsweredAsyncTasks, and reply() never reaches it while this
+                // method's own waiter is live. So #329 narrows this window rather than closing it.
+                // Open as fleetd #334 — do not read this guard as complete.
                 if (result.outcome() != Outcome.QUESTION && task != null) {
                     finishAsyncTask(task, result);
                 }

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

@@ -570,6 +570,228 @@ class ConfigRefTest {
         assertEquals(30, ref.get().configReload().intervalSeconds());
     }
 
+    /**
+     * fleetd #330: {@code health:} is read both ways — {@code Fleetd.java:556-563} builds the
+     * monitor off the startup snapshot and never rebuilds it, but {@code Fleetd.java:648-650} reads
+     * {@code config.get().health()} live on every {@code fleet_profiles} call. A changed value is
+     * neither purely hot nor purely deferred, so it gets its own {@code split} report naming both
+     * halves rather than a bare "config reloaded" (which would hide the frozen half) or a plain
+     * {@code deferred} entry (which would hide that the coverage string already applied).
+     */
+    @Test
+    void changingHealthIsReportedAsSplit(@TempDir Path dir) throws Exception {
+        Path f = dir.resolve("fleetd.yaml");
+        Files.writeString(f, yaml("""
+                health:
+                  enabled: true
+                  intervalSeconds: 30
+                """));
+        ConfigRef ref = refFor(f);
+
+        Files.writeString(f, yaml("""
+                health:
+                  enabled: true
+                  intervalSeconds: 90
+                """));
+        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("health:"), 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 monitor itself is what waits for a restart.
+        assertEquals(90, ref.get().health().intervalSeconds());
+    }
+
+    /**
+     * fleetd #330: {@code coordinator:} is the other split key — {@code Fleetd.java:502} opens the
+     * {@code LeadMailbox} off the startup snapshot and never reopens it, but
+     * {@code HerdrPeerLauncher.java:1530} reads {@code config.get().coordinator()} live on every
+     * spawn to keep the broker URI env-var name out of a member's environment.
+     */
+    @Test
+    void changingCoordinatorIsReportedAsSplit(@TempDir Path dir) throws Exception {
+        Path f = dir.resolve("fleetd.yaml");
+        Files.writeString(f, yaml("""
+                coordinator:
+                  selfId: mac-a
+                """));
+        ConfigRef ref = refFor(f);
+
+        Files.writeString(f, yaml("""
+                coordinator:
+                  selfId: mac-b
+                """));
+        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("coordinator:"), out.split().toString());
+        assertTrue(out.split().getFirst().contains("restart"), out.split().toString());
+        assertTrue(out.split().getFirst().contains("live"), out.split().toString());
+        // The snapshot still carries the new value — the LeadMailbox connection is what waits for a
+        // restart; selfId names this daemon's own inbox queue and a peer cannot discover a rename.
+        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
+     * live half of a split key genuinely took effect, so refusing would throw that away.
+     */
+    @Test
+    void aSplitChangeDoesNotRefuseTheReload(@TempDir Path dir) throws Exception {
+        Path f = dir.resolve("fleetd.yaml");
+        Files.writeString(f, yaml("coordinator:\n  selfId: mac-a\n"));
+        ConfigRef ref = refFor(f);
+
+        Files.writeString(f, yaml("coordinator:\n  selfId: mac-b\n"));
+        ConfigRef.Outcome out = ref.reload();
+
+        assertTrue(out.applied());
+        assertTrue(out.error() == null);
+        assertTrue(out.coldKeys().isEmpty());
+    }
+
+    /**
+     * Both split keys can change in one reload — the report names both, and a caller reading
+     * {@code split} does not have to guess which half of which key already applied.
+     */
+    @Test
+    void changingBothSplitKeysReportsBoth(@TempDir Path dir) throws Exception {
+        Path f = dir.resolve("fleetd.yaml");
+        Files.writeString(f, yaml("""
+                health:
+                  enabled: true
+                coordinator:
+                  selfId: mac-a
+                """));
+        ConfigRef ref = refFor(f);
+
+        Files.writeString(f, yaml("""
+                health:
+                  enabled: false
+                coordinator:
+                  selfId: mac-b
+                """));
+        ConfigRef.Outcome out = ref.reload();
+
+        assertTrue(out.applied());
+        assertEquals(2, out.split().size(), out.split().toString());
+        assertTrue(out.split().stream().anyMatch(s -> s.startsWith("health:")), out.split().toString());
+        assertTrue(out.split().stream().anyMatch(s -> s.startsWith("coordinator:")), out.split().toString());
+    }
+
+    /**
+     * A split key and a deferred key changing in the same reload must both show up, each in its own
+     * list — proving the two fields do not step on each other and {@link ConfigRef.Outcome#summary()}
+     * reports both halves of the message.
+     */
+    @Test
+    void aSplitChangeAndADeferredChangeCoexist(@TempDir Path dir) throws Exception {
+        Path f = dir.resolve("fleetd.yaml");
+        Files.writeString(f, yaml("""
+                coordinator:
+                  selfId: mac-a
+                lifecycle:
+                  drainTimeoutSeconds: 30
+                """));
+        ConfigRef ref = refFor(f);
+
+        Files.writeString(f, yaml("""
+                coordinator:
+                  selfId: mac-b
+                lifecycle:
+                  drainTimeoutSeconds: 60
+                """));
+        ConfigRef.Outcome out = ref.reload();
+
+        assertTrue(out.applied());
+        assertEquals(java.util.List.of("lifecycle"), out.deferred());
+        assertEquals(1, out.split().size(), out.split().toString());
+        assertTrue(out.split().getFirst().startsWith("coordinator:"), out.split().toString());
+        assertTrue(out.summary().contains("need a restart") || out.summary().contains("needs a restart"),
+                out.summary());
+        assertTrue(out.summary().contains("partially live"), out.summary());
+    }
+
     @Test
     void aFixedRefHasNoFileAndRefusesToReload() {
         FleetConfig cfg = new FleetConfig(null, null, null, null, null, null,

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

@@ -0,0 +1,171 @@
+package dev.ltms.fleet.config;
+
+import org.junit.jupiter.api.Test;
+
+import java.lang.reflect.RecordComponent;
+import java.util.LinkedHashSet;
+import java.util.List;
+import java.util.Set;
+import java.util.TreeSet;
+
+import static org.junit.jupiter.api.Assertions.assertEquals;
+import static org.junit.jupiter.api.Assertions.assertTrue;
+
+/**
+ * fleetd #330: a new top-level {@link FleetConfig} record component must be triaged into a reload
+ * class before it ships, or it repeats fleetd #323 ({@code worktreeGroup} missing from {@code
+ * ConfigRef.changedDeferredKeys}) and fleetd #326 ({@code primary}/{@code configReload} missing the
+ * same way) — a key silently absent from {@link ConfigRef}'s reload machinery, so a reload changing
+ * only that key reports a bare "config reloaded" for a change the running daemon never picked up.
+ *
+ * <p>This is the top-level counterpart of {@link ConfigRefProfileCoverageTest}: instead of
+ * enumerating {@link FleetConfig.Profile}'s record components, it enumerates {@link FleetConfig}'s
+ * own — {@code bind}, {@code health}, {@code memberCredentials}, and so on — and requires each to
+ * fall into exactly one of four homes: {@link ConfigRef#COLD_KEYS}, {@link #DEFERRED_TOP_LEVEL_KEYS}
+ * (compared in {@code ConfigRef.changedDeferredKeys}), {@link ConfigRef#SPLIT_KEYS}, or
+ * {@link #HOT_EXCLUDED_TOP_LEVEL_KEYS} (the escape hatch: read live off the config supplier, so no
+ * reload bookkeeping is needed for it at all).
+ *
+ * <h2>What this checker can and cannot prove</h2>
+ * It proves the record's <em>shape</em> is fully triaged: every one of {@code FleetConfig}'s
+ * components sits in exactly one of the four sets, none sits in two, and the escape hatch
+ * ({@link #HOT_EXCLUDED_TOP_LEVEL_KEYS}) cannot silently grow without a visible diff to this file.
+ * That is what "a new component cannot be added without someone triaging it" means in practice.
+ *
+ * <p>It CANNOT prove that any of the citations are <em>true</em>. "Compared in {@code
+ * changedDeferredKeys}" and "read live off {@code config.get()}" are facts about {@code
+ * ConfigRef.java}, {@code Fleetd.java} and {@code HerdrPeerLauncher.java} that a reflection-only
+ * test over {@code FleetConfig}'s shape has no way to inspect — this test would pass identically
+ * whether or not the cited line still does what the comment next to it says. Trust the citation
+ * because a person read the source (the exact call sites are named next to each set below), not
+ * because this test is green. {@link ConfigRefTest} is what behaviourally proves the deferred and
+ * split keys it covers actually get reported; {@link ConfigRefProfileCoverageTest} does the same,
+ * behaviourally, for {@code FleetConfig.Profile}'s own fields.
+ */
+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.
+     */
+    private static final Set<String> DEFERRED_TOP_LEVEL_KEYS = Set.of(
+            "guard", "worktreeRoot", "worktreeGroup", "primary", "configReload",
+            "leadHeartbeat", "lifecycle", "spawnReadyTimeoutMs", "spawnReadyPollMs",
+            "quarantineCooldownSeconds", "profiles");
+
+    /**
+     * The escape hatch: top-level components with no reload bookkeeping at all, because every read
+     * of them goes live through {@link ConfigRef#get()} rather than off a startup snapshot. A
+     * component belongs here ONLY if that is true — never because adding it here makes this test
+     * pass. fleetd #323 is the cautionary tale for exactly this pattern: the identical hatch on
+     * {@code ConfigRef.LAUNCH_SETTINGS_EXCLUDED} let two profile fields be silently re-broken with
+     * the whole suite green, and it was only caught by mutating the checker itself (see this class's
+     * own mutation test below, and {@code ConfigRefProfileCoverageTest}'s equivalent).
+     *
+     * <ul>
+     *   <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 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", "memberCredentials", "memberLoginShell");
+
+    @Test
+    void everyTopLevelComponentIsAccountedForInExactlyOneClass() {
+        Set<String> allNames = new TreeSet<>();
+        for (RecordComponent rc : COMPONENTS) {
+            allNames.add(rc.getName());
+        }
+
+        Set<String> cold = ConfigRef.COLD_KEYS;
+        Set<String> split = ConfigRef.SPLIT_KEYS;
+        Set<String> deferred = DEFERRED_TOP_LEVEL_KEYS;
+        Set<String> hot = HOT_EXCLUDED_TOP_LEVEL_KEYS;
+
+        // Typo guard on each set — the same check ConfigRefProfileCoverageTest runs on
+        // LAUNCH_SETTINGS_EXCLUDED. A name that does not exist on FleetConfig is a silent no-op.
+        assertTrue(allNames.containsAll(cold),
+                "ConfigRef.COLD_KEYS names a component that does not exist on FleetConfig: " + cold);
+        assertTrue(allNames.containsAll(split),
+                "ConfigRef.SPLIT_KEYS names a component that does not exist on FleetConfig: " + split);
+        assertTrue(allNames.containsAll(deferred),
+                "DEFERRED_TOP_LEVEL_KEYS names a component that does not exist on FleetConfig: " + deferred);
+        assertTrue(allNames.containsAll(hot),
+                "HOT_EXCLUDED_TOP_LEVEL_KEYS names a component that does not exist on FleetConfig: " + hot);
+
+        // 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", "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 "
+                        + "happening again: account for it in ConfigRef.changedDeferredKeys (or "
+                        + "COLD_KEYS/SPLIT_KEYS) instead. If it really is read live, name where and "
+                        + "update this assertion and the field javadoc together.");
+
+        // No component may sit in two buckets at once — the denominator check below could not catch
+        // that on its own (two buckets double-booking one key still sums to the right total if
+        // another key is simultaneously missing), so check every pair directly and name the culprit.
+        record Bucket(String name, Set<String> keys) {}
+        List<Bucket> buckets = List.of(
+                new Bucket("COLD_KEYS", cold), new Bucket("SPLIT_KEYS", split),
+                new Bucket("DEFERRED_TOP_LEVEL_KEYS", deferred), new Bucket("HOT_EXCLUDED_TOP_LEVEL_KEYS", hot));
+        for (int i = 0; i < buckets.size(); i++) {
+            for (int j = i + 1; j < buckets.size(); j++) {
+                Set<String> overlap = new LinkedHashSet<>(buckets.get(i).keys());
+                overlap.retainAll(buckets.get(j).keys());
+                assertEquals(Set.of(), overlap, "a component is in both " + buckets.get(i).name()
+                        + " and " + buckets.get(j).name() + ": " + overlap);
+            }
+        }
+
+        Set<String> union = new TreeSet<>();
+        union.addAll(cold);
+        union.addAll(split);
+        union.addAll(deferred);
+        union.addAll(hot);
+
+        System.out.printf(
+                "FleetConfig top-level coverage — %d components total: %d cold %s, %d deferred %s, "
+                        + "%d split %s, %d hot-excluded %s%n",
+                allNames.size(), cold.size(), cold, deferred.size(), deferred, split.size(), split,
+                hot.size(), hot);
+
+        Set<String> missing = new TreeSet<>(allNames);
+        missing.removeAll(union);
+        assertEquals(Set.of(), missing,
+                "these FleetConfig components are in none of COLD_KEYS, DEFERRED_TOP_LEVEL_KEYS, "
+                        + "SPLIT_KEYS or HOT_EXCLUDED_TOP_LEVEL_KEYS — triage each one into whichever "
+                        + "actually describes it: " + missing);
+        assertEquals(allNames.size(), cold.size() + deferred.size() + split.size() + hot.size(),
+                "counts don't sum to the component total even though every component was found in "
+                        + "the union — " + allNames.size() + " components, " + cold.size()
+                        + " cold + " + deferred.size() + " deferred + " + split.size() + " split + "
+                        + hot.size() + " hot-excluded");
+    }
+}

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

@@ -0,0 +1,219 @@
+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}, {@code
+ * DEFERRED_TOP_LEVEL_KEYS}, {@link ConfigRef#SPLIT_KEYS} or the hot-excluded set. It does
+ * <strong>not</strong> prove that a key's membership in {@code COLD_KEYS} or {@code SPLIT_KEYS}
+ * corresponds to any actual comparison in {@link ConfigRef}: a key can sit in either set with no
+ * branch in {@code changedColdKeys}/{@code changedSplitKeys} checking it, and both
+ * {@link ConfigRefTopLevelCoverageTest} and the "kept in step" {@code assert} inside each of those
+ * methods stay green, because neither one reads the method body — the coverage test only reads
+ * set membership, and the assert only checks that reported entries are a SUBSET of the set, never
+ * that every set member produced a reported entry.
+ *
+ * <p>Measured directly, live, while fixing fleetd #333: dropping the {@code coordinator} branch out
+ * of {@code ConfigRef.changedSplitKeys} while leaving {@code "coordinator"} in
+ * {@link ConfigRef#SPLIT_KEYS} left {@link ConfigRefTopLevelCoverageTest} and the in-method assert
+ * both green — only a hand-written behavioural case in {@link ConfigRefTest} caught it, because it
+ * happened to name that exact key. This is the {@link ConfigRefProfileCoverageTest} mechanism one
+ * level up, generalised over every {@code COLD_KEYS}/{@code SPLIT_KEYS} member rather than one
+ * hand-picked field: enumerate {@link FleetConfig}'s own record components by reflection, build "a
+ * config where only {@code <key>} differs" for each cold/split key, and call the real
+ * {@link ConfigRef#changedColdKeys}/{@link ConfigRef#changedSplitKeys} methods (made
+ * package-private for exactly this, the same reason {@link ConfigRef#sameLaunchSettings} already
+ * is) to prove each one is actually reported — not assumed from a set literal.
+ *
+ * <h2>What this deliberately does NOT cover</h2>
+ * {@code DEFERRED_TOP_LEVEL_KEYS} is not exercised here. That bucket carries the identical
+ * one-way risk in principle — a key added to it with no matching branch in
+ * {@code changedDeferredKeys} would pass {@link ConfigRefTopLevelCoverageTest} exactly the way
+ * {@code coordinator} passed it above — but this test stays narrow to {@code COLD_KEYS} and
+ * {@code SPLIT_KEYS} for two reasons. First, that is where fleetd #333 actually found and measured
+ * the gap (F1 was a live instance of it). Second, most of {@code DEFERRED_TOP_LEVEL_KEYS} already
+ * carries an individual behavioural test in {@link ConfigRefTest} naming it by key — {@code
+ * worktreeGroup}, {@code primary}, {@code configReload}, {@code profiles}' launch settings /
+ * weight-maxLoad / exhaustedPattern / errorPattern / ideProjectDir — which is the same protection
+ * this class gives {@code COLD_KEYS}/{@code SPLIT_KEYS}, just written by hand per key instead of
+ * generated by reflection over the whole set. {@code guard}, {@code lifecycle},
+ * {@code leadHeartbeat}, {@code spawnReadyTimeoutMs}/{@code spawnReadyPollMs} and
+ * {@code quarantineCooldownSeconds} do NOT have a dedicated behavioural test naming them, so the
+ * one-way gap fleetd's own memory notes ("pre-existing on COLD_KEYS and on the test's own
+ * DEFERRED_TOP_LEVEL_KEYS") is real and not fully closed by this class — extending this mechanism's
+ * {@code BASE}/{@code ALT} map to cover every top-level component and adding a third
+ * {@code everyDeferredKeyIsActuallyReportedByChangedDeferredKeys} test is the natural next step, left
+ * for whoever next finds a deferred key with the same shape as this ticket's {@code fleet.leaders}.
+ */
+class ConfigRefTopLevelReportingCoverageTest {
+
+    private static final RecordComponent[] COMPONENTS = FleetConfig.class.getRecordComponents();
+
+    /**
+     * One valid value per top-level {@link FleetConfig} component — "the a value". Components not
+     * exercised by either test below ({@code profiles}, {@code guard}, {@code worktreeRoot}, …) are
+     * left {@code null}/empty; {@link FleetConfig}'s compact constructor only normalizes
+     * {@code profiles}, so every other field accepts {@code null} unmutated.
+     */
+    private static final Map<String, Object> BASE = baseValues();
+
+    /** The same shape, each value distinct from {@link #BASE} — "the b value" — for COLD_KEYS/SPLIT_KEYS only. */
+    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", null);
+        v.put("worktreeRoot", null);
+        v.put("lifecycle", null);
+        v.put("spawnReadyTimeoutMs", null);
+        v.put("spawnReadyPollMs", null);
+        v.put("broker", new FleetConfig.Broker("amqp://a", null, 1));
+        v.put("primary", null);
+        v.put("fleet", new FleetConfig.Fleet(
+                Map.of("opus", new FleetConfig.Leader("sonnet", "lead: opus-a", 1, null, 10,
+                        "claude", null, null, null)),
+                Map.of(), Map.of(), Map.of(), Map.of(), "{role}: {profile} #{n}"));
+        v.put("leadHeartbeat", null);
+        v.put("health", new FleetConfig.Health(true, 30, 600, null, null));
+        v.put("placement", null);
+        v.put("auth", new FleetConfig.Auth("loopback-trust", null));
+        v.put("configReload", null);
+        v.put("quarantineCooldownSeconds", null);
+        v.put("memberCredentials", null);
+        v.put("coordinator", new FleetConfig.Coordinator("amqp://coord-a", null, "self-a", 1));
+        v.put("worktreeGroup", null);
+        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");
+        v.put("profiles", Map.of());
+        v.put("guard", null);
+        v.put("worktreeRoot", null);
+        v.put("lifecycle", null);
+        v.put("spawnReadyTimeoutMs", null);
+        v.put("spawnReadyPollMs", null);
+        v.put("broker", new FleetConfig.Broker("amqp://b", null, 2));
+        v.put("primary", null);
+        // 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", null);
+        v.put("health", new FleetConfig.Health(false, 90, 900, null, null));
+        v.put("placement", null);
+        v.put("auth", new FleetConfig.Auth("token", "TOKEN_ENV"));
+        v.put("configReload", null);
+        v.put("quarantineCooldownSeconds", null);
+        v.put("memberCredentials", null);
+        v.put("coordinator", new FleetConfig.Coordinator("amqp://coord-b", null, "self-b", 2));
+        v.put("worktreeGroup", null);
+        v.put("memberLoginShell", null);
+        assertNamesMatchComponents(v);
+        return v;
+    }
+
+    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);
+    }
+}

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

@@ -23,6 +23,7 @@ import java.util.List;
 import java.util.Map;
 import java.util.Set;
 import java.util.concurrent.atomic.AtomicBoolean;
+import java.util.concurrent.atomic.AtomicReference;
 import java.util.function.Function;
 import java.util.function.Supplier;
 
@@ -370,6 +371,69 @@ 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 #185 stage 2: with {@code memberHerdrSocket:} configured, member panes run under a
      * different OS user — {@link HerdrPeerLauncher#hostEnvNames} describes fleetd's own process, not