t398: correct two false statements in FleetConfigValidateAllTest

Measured at review: reverting validateAll() to a hardcoded list of
today's six calls leaves the suite green (1491 tests, 0 failures). The
class javadoc claimed that mutation fails a test. It does not — claim 1
pins the generic helper on an unrelated class, claim 2 pins today's six,
and a hardcoded list satisfies both.

The interaction was the real hazard. The denominator assertion IS a
tripwire (declaring a seventh validator fails it), but its failure
message said the sweep reaches new validators 'by construction' and told
the author to just update the expected set. If the sweep were ever
replaced by a name list, the one assertion that fires would hand back a
false all-clear at the moment it fired.

Javadoc now states the measurement, and names the denominator test as
the actual guarantee. The assertion message now says to confirm
validateAll() still delegates to invokeAllValidators(this) BEFORE
updating the expected set.

fleetd/fleetd.example.yaml

@@ -877,3 +877,32 @@ guard:
 #   terminal: term_65619bd6174568
 #   pushReminders: 5
 #   pushBackoffMs: 15000
+
+# Central allow-list of models any profiles: entry may name. Nothing checked a profile's model:
+# value before this block existed — it was a free-form string handed straight to the backend
+# adapter, and a withdrawn or misspelled name failed silently instead of at config load (opencode
+# falls back to a default model rather than erroring on an unknown -m).
+#
+# Absent, or present with an empty allow:, is OFF: no profile's model: is checked, exactly like
+# before this block existed. fleetd.yaml is gitignored on every host, so an upgrade must not force
+# every operator to enumerate their models before the daemon will start.
+#
+# The list is the authority; profiles: is checked against it, never the reverse — adding or
+# editing a profiles: entry cannot, by itself, widen what is permitted here.
+#
+# Enforcement is at CONFIG LOAD only (a bad model: fails the daemon at startup, naming both the
+# model and the profile). There is no spawn-time enforcement, no runtime on/off switch, and no
+# interaction with BackendQuarantine — those are separate, later units.
+#
+#   allow → the permitted models. Each entry is its own block (not a bare string) so a later unit
+#           can add an on/off state or a load limit per model without changing this shape.
+#     model → the model id exactly as a profiles: entry's model: field would write it. One flat,
+#             opaque-string namespace: a bare Claude id (claude-sonnet-5) and an opencode
+#             provider-prefixed id (openai/gpt-5.6-terra) both fit here unchanged — the check is a
+#             plain string match, never a parse of the provider prefix or a branch on kind:.
+# models:
+#   allow:
+#     - model: claude-sonnet-5
+#     - model: claude-opus-5
+#     - model: openai/gpt-5.6-terra
+#     - model: amazon.nova-pro-v1:0

fleetd/src/main/java/dev/ltms/fleet/Fleetd.java

@@ -144,19 +144,16 @@ public final class Fleetd {
         SubscriptionGuard guard = new SubscriptionGuard(cfg.guard().hostSet());
         guard.assertPrimaryClean(System.getenv());
 
-        // CB-501: refuse to start if the bind is wider than the auth mode can defend. Under
-        // loopback-trust, "not a known worker" means "the primary" — sound only because the OS
-        // refuses remote connections to a loopback socket. This throws rather than warns so the
-        // dangerous configuration cannot be reached by ignoring a log line.
-        cfg.validateAuthExposure();
-        cfg.validateLeadTabPrefixes();
-        // CB-542: a subscription:true profile whose env: reseats ANTHROPIC_BASE_URL/AUTH_TOKEN would
-        // reach an unguarded endpoint (the launcher skips SubscriptionGuard for it). Refuse at load.
-        cfg.validateSubscriptionProfiles();
-        cfg.validateCharters();
-        // CB-548: every architect slot must name a configured workers: profile — the strong-model
-        // backend the future spawn lifecycle would read. A stale reference dies here, not later.
-        cfg.validateMembers();
+        // Every FleetConfig.validateXxx() the operator's config can fail — CB-501's auth-exposure
+        // check, CB-531's lead-tab-prefix check, CB-542's subscription-profile check, the charter
+        // and member-slot checks, and the "central allow-list of usable models" check — must run
+        // here, at load, before anything below opens a socket or spawns a member. fleetd ticket
+        // "central allow-list of usable models" follow-up: mutation testing found six individual
+        // calls here with nothing proving any of them still ran (deleting one left the full suite
+        // green). validateAll() replaces them with the one call that FleetConfigValidateAllTest
+        // and the Fleetd-startup tests actually pin — see FleetConfig#validateAll's javadoc for
+        // why a name-by-name list here would have the same defect it replaces.
+        cfg.validateAll();
 
         Path socket = cfg.herdrSocket() != null && !cfg.herdrSocket().isBlank()
                 ? Path.of(cfg.herdrSocket())
@@ -584,8 +581,12 @@ public final class Fleetd {
         if (cfg.health() != null && cfg.health().isEnabled()) {
             // CB-580: a member found GONE/NEVER_READY must fail whatever ticket is waiting on it,
             // through the same idempotent target-wide operation CB-516 already uses on release.
+            // fleetd #386: System::nanoTime freezes across a macOS sleep, so the stall check also
+            // gets a wall-clock source to detect and correct for that freeze. Every other decision
+            // in FleetHealthMonitor stays on the monotonic clock, unchanged.
             healthMonitor = new FleetHealthMonitor(agents, sessions::roster, messages, healthScheduler,
-                    System::nanoTime, cfg.health().intervalOrDefault(),
+                    System::nanoTime, () -> TimeUnit.MILLISECONDS.toNanos(System.currentTimeMillis()),
+                    cfg.health().intervalOrDefault(),
                     cfg.health().workingSuspectAfterOrDefault(), messages::abandon);
             String coverage = FleetHealthMonitor.coverage(true,
                     cfg.health().notifications() != null && cfg.health().notifications().configured());

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

@@ -43,6 +43,12 @@ import java.util.function.Supplier;
  *       running daemon keeps whatever this was at startup regardless of a later edit),
  *       {@code spawnReadyTimeoutMs} / {@code spawnReadyPollMs}, {@code quarantineCooldownSeconds}
  *       (CB-578 stage B — baked once into the {@code BackendQuarantine} built at startup),
+ *       {@code models:} (fleetd ticket "central allow-list of usable models" — {@link
+ *       FleetConfig#validateModels()} re-runs against the fresh config in {@link #reload()}
+ *       (via {@link FleetConfig#validateAll()}), so a
+ *       models.allow: edit that would refuse to boot still refuses the reload; a change that
+ *       passes has nothing built at startup to rebuild, so it is reported deferred rather than
+ *       silently accepted with no report at all),
  *       {@code guard:}, {@code worktreeRoot:}, {@code worktreeGroup:} and {@code memberSkills:}
  *       (all three of the latter baked once into the {@code GitWorktrees} built at
  *       {@code Fleetd.java:251} and never rebuilt — fleetd #323 instance 2 found
@@ -135,8 +141,9 @@ import java.util.function.Supplier;
  * </ul>
  *
  * <p><strong>The denominator, measured on 2026-09-04 (fleetd #330; recounted for fleetd #333);
- * recounted again for fleetd #362, and again after {@code idleSleepGuard:} was added.</strong>
- * {@code FleetConfig} has 24 top-level record components: 5 cold, 13 deferred, 3 split, 3
+ * recounted again for fleetd #362, again after {@code idleSleepGuard:} was added, and again after
+ * {@code models:} was added.</strong>
+ * {@code FleetConfig} has 25 top-level record components: 5 cold, 14 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()}
@@ -219,7 +226,7 @@ public final class ConfigRef implements Supplier<FleetConfig> {
     static final Set<String> DEFERRED_KEYS = Set.of(
             "guard", "worktreeRoot", "worktreeGroup", "memberSkills", "primary", "configReload",
             "leadHeartbeat", "lifecycle", "spawnReadyTimeoutMs", "spawnReadyPollMs",
-            "quarantineCooldownSeconds", "profiles", "idleSleepGuard");
+            "quarantineCooldownSeconds", "profiles", "idleSleepGuard", "models");
 
     private final Path path;
     private final AtomicReference<FleetConfig> current;
@@ -322,12 +329,12 @@ public final class ConfigRef implements Supplier<FleetConfig> {
             fresh = FleetConfig.load(path);
             // The same gate startup runs. A config that would have refused to boot must not be able
             // to slip in through a reload — that is how a daemon ends up in a state it could never
-            // have started in, which is the hardest kind to debug.
-            fresh.validateAuthExposure();
-            fresh.validateLeadTabPrefixes();
-            fresh.validateSubscriptionProfiles();
-            fresh.validateCharters();
-            fresh.validateMembers();
+            // have started in, which is the hardest kind to debug. fleetd ticket "central allow-list
+            // of usable models" follow-up: this used to be six individual validateXxx() calls, and
+            // mutation testing found two of the six unpinned here even though startup pinned nothing
+            // at all — see FleetConfig#validateAll's javadoc for why the fix is one reflective call,
+            // not a longer hand-maintained list.
+            fresh.validateAll();
         } catch (RuntimeException e) {
             String msg = e.getMessage() == null ? e.toString() : e.getMessage();
             log.warn("config reload from {} refused, keeping the running config: {}", path, msg);
@@ -439,6 +446,15 @@ public final class ConfigRef implements Supplier<FleetConfig> {
         if (!Objects.equals(old.idleSleepGuard(), fresh.idleSleepGuard())) {
             changed.add("idleSleepGuard");
         }
+        // fleetd ticket "central allow-list of usable models": validateModels() runs again in
+        // reload() above (via validateAll()), so a bad edit is already refused as cold-adjacent
+        // (the whole reload is refused via the catch block, never partially applied). A GOOD edit
+        // to the allow-list
+        // itself has nothing built at startup to rebuild — it only ever mattered to the validation
+        // call that already ran — so report it deferred rather than silently swallowing the change.
+        if (!Objects.equals(old.models(), fresh.models())) {
+            changed.add("models");
+        }
         if (!Objects.equals(old.spawnReadyTimeoutMs(), fresh.spawnReadyTimeoutMs())
                 || !Objects.equals(old.spawnReadyPollMs(), fresh.spawnReadyPollMs())) {
             changed.add("spawnReady*");

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

@@ -16,11 +16,15 @@ import org.slf4j.LoggerFactory;
 
 import java.io.IOException;
 import java.io.UncheckedIOException;
+import java.lang.reflect.InvocationTargetException;
+import java.lang.reflect.Method;
+import java.lang.reflect.Modifier;
 import java.nio.file.Files;
 import java.nio.file.Path;
 import java.util.ArrayList;
 import java.util.Arrays;
 import java.util.Collections;
+import java.util.Comparator;
 import java.util.HashSet;
 import java.util.LinkedHashMap;
 import java.util.List;
@@ -124,6 +128,13 @@ import java.util.regex.PatternSyntaxException;
  *                      under a member's long turn. {@code null} (the block omitted) behaves the
  *                      same as an explicit {@code enabled: true}; set {@code enabled: false} to
  *                      turn it off. See {@link dev.ltms.fleet.power.IdleSleepGuard}.
+ * @param models        central allow-list of models any {@code profiles:} entry may name. {@code
+ *                      null} or an empty {@code allow:} ⇒ off: {@link #validateModels()} checks
+ *                      nothing and every existing config keeps working exactly as it does today.
+ *                      When non-empty, a profile whose {@code model:} is not one of {@link
+ *                      Models#ids()} fails config load, naming both the model and the profile —
+ *                      see {@link #validateModels()}. This block only decides what may be
+ *                      CONFIGURED; nothing here enforces it at spawn time. See {@link Models}.
  */
 @JsonIgnoreProperties(ignoreUnknown = true)
 public record FleetConfig(
@@ -150,7 +161,22 @@ public record FleetConfig(
         String worktreeGroup,
         String memberLoginShell,
         String memberSkills,
-        IdleSleepGuard idleSleepGuard) {
+        IdleSleepGuard idleSleepGuard,
+        Models models) {
+
+    /** Back-compat form before the {@code models:} block was added. */
+    public FleetConfig(Bind bind, String herdrSocket, String memberHerdrSocket, Map<String, Profile> profiles,
+                         Guard guard, String worktreeRoot, Lifecycle lifecycle, Integer spawnReadyTimeoutMs,
+                         Integer spawnReadyPollMs, Broker broker, Primary primary, Fleet fleet,
+                         LeadHeartbeat leadHeartbeat, Health health, String placement, Auth auth,
+                         ConfigReload configReload, Integer quarantineCooldownSeconds,
+                         MemberCredentials memberCredentials, Coordinator coordinator, String worktreeGroup,
+                         String memberLoginShell, String memberSkills, IdleSleepGuard idleSleepGuard) {
+        this(bind, herdrSocket, memberHerdrSocket, profiles, guard, worktreeRoot, lifecycle, spawnReadyTimeoutMs,
+                spawnReadyPollMs, broker, primary, fleet, leadHeartbeat, health, placement, auth,
+                configReload, quarantineCooldownSeconds, memberCredentials, coordinator, worktreeGroup,
+                memberLoginShell, memberSkills, idleSleepGuard, null);
+    }
 
     /** Back-compat form before the {@code idleSleepGuard:} block was added. */
     public FleetConfig(Bind bind, String herdrSocket, String memberHerdrSocket, Map<String, Profile> profiles,
@@ -1318,6 +1344,72 @@ public record FleetConfig(
         }
     }
 
+    /**
+     * Central allow-list of models any {@code profiles:} entry may name (fleetd ticket: "a central
+     * allow-list of usable models"). Nothing before this block checked a profile's {@code model:}
+     * against anything — it was a free-form string handed straight to the backend adapter, and a
+     * withdrawn or misspelled name failed silently (opencode falls back to a default model rather
+     * than erroring on an unknown {@code -m}) rather than at config load, where a mistake is cheap.
+     *
+     * <p><b>Absent or empty {@code allow:} is "off"</b>, on purpose: this is a large deployment
+     * with gitignored {@code fleetd.yaml} on more than one host, and a change that forced every
+     * operator to enumerate their models before the daemon would start would break every one of
+     * them on upgrade. See {@link FleetConfig#validateModels()}, which is where the allow-list is
+     * actually enforced, at config load.
+     *
+     * <p><b>The list is the authority; a {@code profiles:} entry is checked against it, never the
+     * other way around.</b> Adding or editing a {@code profiles:} entry cannot, by itself, widen
+     * the set of permitted models — only editing {@code models.allow:} itself can. This is the
+     * invariant the ticket asked for: the two blocks are validated in one direction only.
+     *
+     * <p><b>Out of scope here, deliberately:</b> nothing in this block is read at spawn time —
+     * enforcing it against a live spawn, an on/off runtime switch, and any interaction with {@code
+     * BackendQuarantine} are separate units. This block is config-load validation only.
+     *
+     * @param allow the permitted models, each its own {@link ModelEntry} rather than a bare
+     *              string — see that record's javadoc for why. {@code null}/empty ⇒ the block is
+     *              treated as absent: {@link FleetConfig#validateModels()} checks nothing.
+     */
+    @JsonIgnoreProperties(ignoreUnknown = true)
+    public record Models(List<ModelEntry> allow) {
+
+        public Models {
+            allow = (allow == null) ? List.of() : List.copyOf(allow);
+        }
+
+        /**
+         * One permitted model, named as a record rather than a bare string on purpose: a later unit
+         * needs to hang an on/off state and a load-limit state off each entry, and a bare {@code
+         * List<String>} cannot grow those fields without changing the YAML shape underneath every
+         * operator who already wrote one. {@link #model()} is intentionally a single flat,
+         * opaque-string namespace — a bare Claude id ({@code claude-sonnet-5}) and an opencode
+         * provider-prefixed id ({@code openai/gpt-5.6-terra}) both fit it unchanged, because
+         * {@link FleetConfig#validateModels()} only ever compares a profile's {@code model:} value
+         * against this string for exact equality; it never parses a provider prefix or branches on
+         * a profile's {@code kind:}.
+         *
+         * @param model the model id exactly as a {@code profiles:} entry's {@code model:} field
+         *              would name it
+         */
+        @JsonIgnoreProperties(ignoreUnknown = true)
+        public record ModelEntry(String model) {
+            public ModelEntry {
+                model = (model == null || model.isBlank()) ? null : model.trim();
+            }
+        }
+
+        /** {@link #allow}'s model ids, as a set for membership checks. Blank/null entries are dropped. */
+        public Set<String> ids() {
+            Set<String> ids = new java.util.LinkedHashSet<>();
+            for (ModelEntry e : allow) {
+                if (e != null && e.model() != null) {
+                    ids.add(e.model());
+                }
+            }
+            return Collections.unmodifiableSet(ids);
+        }
+    }
+
     /**
      * The terminal → lead-name map seeded from the legacy singular {@code primary:} pin (CB-530).
      *
@@ -1569,7 +1661,7 @@ public record FleetConfig(
             "lifecycle", "spawnReadyTimeoutMs", "spawnReadyPollMs", "broker", "primary", "fleet",
             "leadHeartbeat", "health", "placement", "auth", "configReload", "quarantineCooldownSeconds",
             "memberCredentials", "coordinator", "worktreeGroup", "memberLoginShell", "memberSkills",
-            "idleSleepGuard");
+            "idleSleepGuard", "models");
 
     /** Load and validate config from {@code path}. */
     public static FleetConfig load(Path path) {
@@ -2254,10 +2346,15 @@ public record FleetConfig(
         // enabled: true — see its javadoc), so defaulting the block here would change nothing a
         // reader observes and would only obscure that "block omitted" and "block present and
         // enabled" are deliberately the same outcome.
+        // models is left as-is, like broker/primary/coordinator above: null/empty is "off", and an
+        // absent block must validate nothing (see Models's javadoc) — defaulting it here to an
+        // empty Models would be a no-op for validateModels() either way, since an empty allow-list
+        // already means "check nothing", so there is nothing to gain and one more null check to
+        // avoid by leaving it exactly as configured.
         return new FleetConfig(b, herdrSocket, memberHerdrSocket, profiles, g, worktreeRoot, l, timeout, pollMs,
                 broker, primary, f, leadHeartbeat, health, placementOrDefault, a, configReload,
                 quarantineCooldown, mc, coordinator, worktreeGroup, memberLoginShell, memberSkills,
-                idleSleepGuard);
+                idleSleepGuard, models);
     }
 
     /**
@@ -2477,6 +2574,118 @@ public record FleetConfig(
         }
     }
 
+    /**
+     * Reject a {@code profiles:} entry whose {@code model:} is not on the configured {@link
+     * #models} allow-list.
+     *
+     * <p>Absent or empty {@code models.allow:} validates nothing — see {@link Models}'s javadoc:
+     * an existing config with no such block must keep working exactly as it does today. Once the
+     * operator declares at least one entry, every profile's {@code model:} (when set — a profile
+     * may legitimately leave it {@code null}, e.g. a {@code subscription: true} profile relying on
+     * the account's own default) must equal one of {@link Models#ids()} exactly. The comparison is
+     * a flat string match: a bare Claude id and an opencode {@code provider/model} id are both
+     * just opaque strings here, so nothing here needs to know which {@code kind:} a profile runs.
+     *
+     * <p>The check runs one direction only, by construction: it reads {@link #profiles} and
+     * {@link #models}, and only ever adds to {@code bad} when a profile's model is missing from
+     * the allow-list. Nothing here can be satisfied by widening a {@code profiles:} entry — only
+     * editing {@code models.allow:} itself changes what passes. That is the invariant the ticket
+     * asked for: the list is the authority, profiles are checked against it.
+     *
+     * @throws IllegalStateException when any profile names a model outside the configured
+     *                               allow-list, naming both the model and the profile that wanted it
+     */
+    public void validateModels() {
+        if (models == null || models.allow().isEmpty()) {
+            return;
+        }
+        Set<String> allowed = models.ids();
+        List<String> bad = new ArrayList<>();
+        profiles.forEach((name, p) -> {
+            String model = p.model();
+            if (model != null && !model.isBlank() && !allowed.contains(model)) {
+                bad.add("profile '" + name + "' names model '" + model + "', which is not in "
+                        + "models.allow: (have: " + allowed + ").");
+            }
+        });
+        if (!bad.isEmpty()) {
+            throw new IllegalStateException("refusing to start: " + String.join(" ", bad));
+        }
+    }
+
+    /**
+     * Runs every validator this class declares — found by reflection, not by name.
+     *
+     * <p>fleetd ticket "central allow-list of usable models", follow-up: mutation testing found
+     * that although each of the six validators above was well pinned on its own, nothing proved
+     * either real caller ({@code Fleetd.main} and {@link ConfigRef#reload()}) still
+     * invoked it — deleting a call site left the full suite green. The fix is not a seventh test
+     * per caller; a hand-maintained list of six names here would have the exact same defect its
+     * own javadoc would warn against: the seventh validator someone adds next month has no reason
+     * to be added to it. So this method does not name any validator. It sweeps {@link
+     * #getClass()}'s own public, no-argument, {@code void} methods whose name starts with {@code
+     * "validate"} (excluding itself) and invokes every one it finds, via {@link
+     * #invokeAllValidators}. A new {@code validateXxx()} method is therefore wired into both
+     * callers the moment it is written — there is no second step to forget, and so no state in
+     * which it silently never runs.
+     *
+     * <p>{@code Fleetd.main} and {@link ConfigRef#reload()} each call this one method instead of
+     * the six individually — see the comments at those two call sites for why
+     * each must run it.
+     *
+     * <p>Methods run in a fixed (alphabetical) order, so a config with more than one violation
+     * always names the same one first, on every run.
+     *
+     * @throws IllegalStateException (or whatever unchecked exception a validator itself throws),
+     *                                propagated unchanged from the first validator, in that order,
+     *                                that finds a problem
+     */
+    public void validateAll() {
+        invokeAllValidators(this);
+    }
+
+    /**
+     * The reflective sweep behind {@link #validateAll()}, kept as its own method — taking any
+     * {@code target}, not just {@code this} — so a test can prove the MECHANISM is generic (it
+     * would sweep a seventh {@code validateXxx()} method added to any class, not just something
+     * special-cased to today's six on {@link FleetConfig}) without needing to add a real, unwanted
+     * seventh validator to this class just to exercise that claim. See {@code
+     * FleetConfigValidateAllTest} for that proof.
+     *
+     * @param target an object whose public, no-argument, {@code void} methods named {@code
+     *               validateXxx} (any name starting with {@code "validate"}, excluding {@code
+     *               validateAll} itself) should all run, in alphabetical-by-name order
+     */
+    static void invokeAllValidators(Object target) {
+        List<Method> methods = new ArrayList<>();
+        for (Method m : target.getClass().getMethods()) {
+            if (Modifier.isPublic(m.getModifiers())
+                    && m.getParameterCount() == 0
+                    && m.getReturnType() == void.class
+                    && m.getName().startsWith("validate")
+                    && !m.getName().equals("validateAll")) {
+                methods.add(m);
+            }
+        }
+        methods.sort(Comparator.comparing(Method::getName));
+        for (Method m : methods) {
+            try {
+                m.invoke(target);
+            } catch (InvocationTargetException e) {
+                Throwable cause = e.getCause();
+                if (cause instanceof RuntimeException re) {
+                    throw re;
+                }
+                if (cause instanceof Error err) {
+                    throw err;
+                }
+                throw new IllegalStateException("validator " + m.getName() + " failed", cause);
+            } catch (IllegalAccessException e) {
+                throw new IllegalStateException("cannot invoke validator " + m.getName(), e);
+            }
+        }
+    }
+
     /** True for the loopback addresses and the unspecified-but-local forms we treat as same-host. */
     private static boolean isLoopbackBind(String host) {
         if (host == null || host.isBlank()) {

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

@@ -38,15 +38,46 @@ public final class FleetHealthMonitor {
      */
     static final long ASK_LAPSE_RECHECK_DELAY_SECONDS = 120;
 
+    /**
+     * fleetd #386: {@code System.nanoTime()} (or whatever {@link #clock} is) does not advance while
+     * macOS sleeps, so a raw {@code nowNanos - lastActivityAtNanos} comparison freezes with the
+     * host and can never cross {@link #workingSuspectAfterNanos}. This is a second, wall-clock
+     * source used ONLY inside the stall check ({@link #stallElapsedNanos}) to detect and correct
+     * for that freeze. Nothing else in this class reads it — every other decision (readiness grace,
+     * the fault classification itself) stays exactly on {@link #clock}, as the ticket requires.
+     */
+    private static final LongSupplier DEFAULT_REALTIME_CLOCK =
+            () -> TimeUnit.MILLISECONDS.toNanos(System.currentTimeMillis());
+
     private final AgentControl agents;
     private final Supplier<List<MemberSession>> roster;
     private final MessageService messages;
     private final ScheduledExecutorService scheduler;
     private final LongSupplier clock;
+    private final LongSupplier realtimeClock;
     private final long intervalSeconds;
+    private final long tickIntervalNanos;
     private final long workingSuspectAfterNanos;
     private final BiConsumer<String, String> failTarget;
     private final Map<String, HealthPrior> priors = new HashMap<>();
+    /**
+     * fleetd #386 clock-drift bookkeeping. {@code haveClockBaseline}/{@code lastTickMonoNanos}/
+     * {@code lastTickRealNanos} track the previous tick's pair of readings so each new tick can
+     * measure how far the two clocks moved apart since then. {@code accumulatedDriftNanos} is the
+     * running total of every such divergence observed since this monitor started (never decreases —
+     * the monotonic clock can only lag real time, never lead it). {@code busyDriftBaselineNanos}/
+     * {@code busyBaselineActivityNanos} record, per target, the value of {@code accumulatedDriftNanos}
+     * at the moment this monitor first saw that target's CURRENT {@code lastActivityAtNanos} while
+     * BUSY — so {@link #stallElapsedNanos} adds back only the drift observed DURING this BUSY span,
+     * never drift from a sleep that happened before the member went busy. All five fields are touched
+     * only from {@code tick()}, like {@link #priors}.
+     */
+    private boolean haveClockBaseline = false;
+    private long lastTickMonoNanos;
+    private long lastTickRealNanos;
+    private long accumulatedDriftNanos = 0;
+    private final Map<String, Long> busyDriftBaselineNanos = new HashMap<>();
+    private final Map<String, Long> busyBaselineActivityNanos = new HashMap<>();
     /**
      * The live classification per member, and the only one of this class's three maps that more
      * than one scheduler task touches. {@code tick} writes it (and prunes it to the roster);
@@ -88,12 +119,29 @@ public final class FleetHealthMonitor {
     public FleetHealthMonitor(AgentControl agents, Supplier<List<MemberSession>> roster, MessageService messages,
                               ScheduledExecutorService scheduler, LongSupplier clock, long intervalSeconds,
                               long workingSuspectAfterSeconds, BiConsumer<String, String> failTarget) {
+        this(agents, roster, messages, scheduler, clock, DEFAULT_REALTIME_CLOCK, intervalSeconds,
+                workingSuspectAfterSeconds, failTarget);
+    }
+
+    /**
+     * @param realtimeClock fleetd #386: a wall-clock nanosecond source (e.g.
+     *                       {@code System.currentTimeMillis()} converted to nanos) that keeps
+     *                       advancing while {@code clock} is frozen by a host sleep. Used only to
+     *                       correct the stall check — see the class-level javadoc on the
+     *                       clock-drift fields.
+     */
+    public FleetHealthMonitor(AgentControl agents, Supplier<List<MemberSession>> roster, MessageService messages,
+                              ScheduledExecutorService scheduler, LongSupplier clock, LongSupplier realtimeClock,
+                              long intervalSeconds, long workingSuspectAfterSeconds,
+                              BiConsumer<String, String> failTarget) {
         this.agents = agents;
         this.roster = roster;
         this.messages = messages;
         this.scheduler = scheduler;
         this.clock = clock;
+        this.realtimeClock = Objects.requireNonNull(realtimeClock, "realtimeClock");
         this.intervalSeconds = intervalSeconds;
+        this.tickIntervalNanos = TimeUnit.SECONDS.toNanos(intervalSeconds);
         this.workingSuspectAfterNanos = TimeUnit.SECONDS.toNanos(workingSuspectAfterSeconds);
         this.failTarget = Objects.requireNonNull(failTarget, "failTarget");
     }
@@ -123,6 +171,7 @@ public final class FleetHealthMonitor {
             for (Agent agent : agentsNow) live.put(agent.terminalId(), agent);
             HashSet<String> current = new HashSet<>();
             long nowNanos = clock.getAsLong();
+            long driftBeforeThisTick = observeClockDrift(nowNanos);
             for (MemberSession session : rosterNow) {
                 current.add(session.terminalId());
                 Agent agent = live.get(session.terminalId());
@@ -133,7 +182,7 @@ public final class FleetHealthMonitor {
                         && session.state() != MemberSession.State.SPAWNING;
                 boolean readinessGraceElapsed = nowNanos - session.spawnedAtNanos() >= READINESS_GRACE_NANOS;
                 boolean stalled = session.state() == MemberSession.State.BUSY
-                        && nowNanos - session.lastActivityAtNanos() >= workingSuspectAfterNanos;
+                        && stallElapsedNanos(session, nowNanos, driftBeforeThisTick) >= workingSuspectAfterNanos;
                 // CB-643: the three message-layer facts CB-640 published. Read them here rather than
                 // leaving them false — that constant is what made 8 of the 9 fault states dead.
                 boolean queuedDelivery = messages.hasQueuedDelivery(session.terminalId());
@@ -151,6 +200,8 @@ public final class FleetHealthMonitor {
             priors.keySet().retainAll(current);
             states.keySet().retainAll(current);
             orphanStreaks.keySet().retainAll(current);
+            busyDriftBaselineNanos.keySet().retainAll(current);
+            busyBaselineActivityNanos.keySet().retainAll(current);
         } catch (Throwable error) {
             // Any unclassified collection failure must never kill the monitor's only scheduler task.
             log.warn("fleet health collection failed; will retry next tick", error);
@@ -175,6 +226,65 @@ public final class FleetHealthMonitor {
         return streak >= ORPHAN_CONFIRM_TICKS;
     }
 
+    /**
+     * fleetd #386: compare this tick's monotonic and real-time readings against the previous
+     * tick's, and fold any positive divergence into {@link #accumulatedDriftNanos} (a ratchet — it
+     * never decreases, since the monotonic clock can only fall behind real time, never ahead of
+     * it). Logs once, at WARN, when that single tick's divergence exceeds one full tick interval —
+     * the signature of a host that slept between the two ticks (a tick literally cannot run while
+     * the process itself is suspended, so the whole sleep duration lands inside one tick's gap).
+     *
+     * @return {@link #accumulatedDriftNanos} as it stood BEFORE this tick's divergence was folded
+     *         in — the baseline {@link #stallElapsedNanos} needs when a target is observed BUSY
+     *         for the first time this tick, so a sleep that happened before this member went busy
+     *         is not attributed to it.
+     */
+    private long observeClockDrift(long nowNanos) {
+        long nowRealNanos = realtimeClock.getAsLong();
+        long driftBeforeThisTick = accumulatedDriftNanos;
+        if (haveClockBaseline) {
+            long monoDelta = nowNanos - lastTickMonoNanos;
+            long realDelta = nowRealNanos - lastTickRealNanos;
+            long tickDrift = realDelta - monoDelta;
+            if (tickDrift > tickIntervalNanos) {
+                log.warn("fleet health: the monotonic clock did not advance for about {}s that the "
+                        + "real clock did since the last tick (host likely slept); the stall "
+                        + "detector could not see that time", TimeUnit.NANOSECONDS.toSeconds(tickDrift));
+            }
+            if (tickDrift > 0) {
+                accumulatedDriftNanos = driftBeforeThisTick + tickDrift;
+            }
+        }
+        lastTickMonoNanos = nowNanos;
+        lastTickRealNanos = nowRealNanos;
+        haveClockBaseline = true;
+        return driftBeforeThisTick;
+    }
+
+    /**
+     * fleetd #386: {@code nowNanos - lastActivityAtNanos} alone freezes across a host sleep, since
+     * both come from the monotonic {@link #clock}. This adds back the real-time drift observed
+     * since this BUSY span started — not the monitor's whole lifetime, so a sleep that happened
+     * before this member went busy never leaks into its stall reading (see the class-level javadoc
+     * on the drift fields). The baseline resets whenever {@code lastActivityAtNanos} changes (a new
+     * turn) or the member is not currently BUSY.
+     */
+    private long stallElapsedNanos(MemberSession session, long nowNanos, long driftBeforeThisTick) {
+        String target = session.terminalId();
+        if (session.state() != MemberSession.State.BUSY) {
+            busyDriftBaselineNanos.remove(target);
+            busyBaselineActivityNanos.remove(target);
+            return nowNanos - session.lastActivityAtNanos();
+        }
+        Long baselineActivity = busyBaselineActivityNanos.get(target);
+        if (baselineActivity == null || baselineActivity != session.lastActivityAtNanos()) {
+            busyBaselineActivityNanos.put(target, session.lastActivityAtNanos());
+            busyDriftBaselineNanos.put(target, driftBeforeThisTick);
+        }
+        long driftSinceBusyStart = accumulatedDriftNanos - busyDriftBaselineNanos.get(target);
+        return (nowNanos - session.lastActivityAtNanos()) + driftSinceBusyStart;
+    }
+
     void reportTransition(String target, HealthState next) {
         HealthState previous = states.put(target, next);
         if (previous == next) return;

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

@@ -13,6 +13,7 @@ import java.nio.file.attribute.PosixFilePermissions;
 import java.time.Duration;
 import java.time.Instant;
 import java.util.ArrayList;
+import java.util.HashSet;
 import java.util.List;
 import java.util.Set;
 import java.util.stream.Stream;
@@ -28,24 +29,46 @@ import java.util.stream.Stream;
  * control lacked: herdr applies that overlay BEFORE the shell starts, so any sourced file can undo
  * it — and did.
  *
- * <p><b>Which file is last depends on the platform, so the scrub runs from two of them.</b> zsh
+ * <p><b>zsh reads its four startup files under three different conditions, so no single file is
+ * guaranteed to run — the scrub has to cover the gap between them, not just the platforms.</b> zsh
  * reads {@code .zshenv} always, {@code .zprofile} and {@code .zlogin} only for a LOGIN shell, and
- * {@code .zshrc} only for an INTERACTIVE one. herdr does not open the same kind of shell
- * everywhere — measured on herdr 0.8.0: macOS panes run {@code -zsh} (login, so {@code .zlogin}
- * runs), Linux panes run a plain {@code /usr/bin/zsh} (interactive but NOT login, so
- * {@code .zlogin} never runs at all). A scrub in {@code .zlogin} alone is therefore a control that
- * silently does nothing on Linux — the exact failure this class exists to remove, one platform
- * over.
+ * {@code .zshrc} only for an INTERACTIVE one. A pane shell that is at least one of login or
+ * interactive is covered by sourcing the scrub from {@code .zshrc} and {@code .zlogin} (below), but
+ * a pane shell that is NEITHER reads only {@code .zshenv} and stops — fleetd #388, measured: a herdr
+ * pane can be neither login nor interactive, and such a pane read {@code .zshenv}, never reached
+ * {@code scrub.zsh}, and left no report at all. A bare {@code argv[0]} of {@code /usr/bin/zsh}
+ * proves the shell is NOT a login shell; it says nothing about whether it is interactive, so it
+ * must never be read as "therefore interactive" — that wrong inference is what let #388 ship.
  *
- * <p>So both {@code .zshrc} and {@code .zlogin} source the same generated {@code scrub.zsh} after
- * sourcing their {@code $HOME} counterpart. On Linux only the first fires; on macOS both do, and
- * the second pass is deliberate rather than merely harmless — it re-scrubs anything the operator's
- * own {@code ~/.zlogin} exported after {@code .zshrc} had finished. Re-running is idempotent: a
- * name already blank is blanked again, and the report is rewritten with the same counts.
+ * <p>So {@code .zshenv} carries a THIRD pass, guarded by the exact condition that defines the gap:
+ * {@code [[ ! -o login && ! -o interactive ]]}. That guard is why this pass cannot double-scrub a
+ * pane that {@code .zshrc} or {@code .zlogin} will also cover — one of {@code -o login}/
+ * {@code -o interactive} is always true there, so the {@code .zshenv} pass never fires for them, and
+ * their own unconditional sourcing is untouched. The guard also carries a sentinel
+ * ({@value #SCRUB_SENTINEL}) so it fires once per PANE and not once per PROCESS: {@code .zshenv} is
+ * read by every zsh a member's own tooling forks (a plain {@code zsh -c '...'} for a single
+ * command is itself neither login nor interactive), and those children inherit variables their
+ * parent deliberately set for them (git hooks get {@code GIT_DIR}, a venv gets
+ * {@code VIRTUAL_ENV}, a build tool gets {@code NODE_OPTIONS} or {@code JAVA_TOOL_OPTIONS}).
+ * Re-scrubbing every such child would blank all of that, and would also make the pane's own
+ * {@code scrub-report.txt} — rewritten on every pass — describe whichever child exited last
+ * instead of the pane. The sentinel is exported only AFTER {@code scrub.zsh} runs, so the pass
+ * that sets it never sees it and cannot blank it; it must also be on the scrub's own allow-list
+ * (see {@link #generate(Path, Set)}) so a later pass, in the same pane, cannot blank it back to
+ * empty — an exported-but-empty sentinel reads as unset to the {@code -z} guard and would silently
+ * re-enable scrubbing for every subsequent child of that pane.
+ *
+ * <p>So all three of {@code .zshenv} (gap only, guarded), {@code .zshrc}, and {@code .zlogin}
+ * source the same generated {@code scrub.zsh} after sourcing their {@code $HOME} counterpart. A
+ * login-and-interactive pane runs the {@code .zshrc} and {@code .zlogin} passes, and the second is
+ * deliberate rather than merely harmless — it re-scrubs anything the operator's own
+ * {@code ~/.zlogin} exported after {@code .zshrc} had finished. A pane that is neither runs only the
+ * {@code .zshenv} pass. Re-running is idempotent: a name already blank is blanked again, and the
+ * report is rewritten with the same counts.
  *
  * <p>Each generated file sources its {@code $HOME} counterpart FIRST, so {@code PATH} and every
- * toolchain binary still resolve exactly as the operator configured them; only afterwards does
- * {@code .zlogin} run the scrub: every EXPORTED variable not on the derived allow-list is re-exported
+ * toolchain binary still resolve exactly as the operator configured them; only afterwards does the
+ * scrub run: every EXPORTED variable not on the derived allow-list is re-exported
  * blank. Blank, not credential-shaped-pattern-filtered: a pattern list ({@code *TOKEN*}, …) is an
  * enumeration and misses what it did not think of — a username is the other half of a credential and
  * is shaped like none. Credential-SHAPED names among the blanked set go to the WARN log only,
@@ -63,7 +86,11 @@ public final class EnvAllowListScrub {
     /** Name of the report file written into the generated directory by the scrub itself. */
     static final String REPORT_FILE = "scrub-report.txt";
 
-    /** The scrub body, generated once and sourced from both {@code .zshrc} and {@code .zlogin}. */
+    /**
+     * The scrub body, generated once and sourced from {@code .zshrc} and {@code .zlogin}
+     * unconditionally, and from {@code .zshenv} when the pane shell is neither login nor
+     * interactive (fleetd #388) — see the class javadoc.
+     */
     static final String SCRUB_FILE = "scrub.zsh";
 
     /** Prefix of every generated directory — also what {@link #reapOrphans} matches on. */
@@ -79,6 +106,28 @@ public final class EnvAllowListScrub {
     private static final String SOURCE_SCRUB =
             "source \"$ZDOTDIR/" + SCRUB_FILE + "\"\n";
 
+    /**
+     * fleetd #388: marks a pane, not a process, as already scrubbed. Set only by the guarded
+     * {@code .zshenv} pass (see {@link #NEITHER_LOGIN_NOR_INTERACTIVE_SCRUB}) after
+     * {@code scrub.zsh} has run, so it must also be folded into that pass's own allow-list — see
+     * the class javadoc's "must also be on the scrub's own allow-list" paragraph.
+     */
+    static final String SCRUB_SENTINEL = "_CB633_SCRUBBED";
+
+    /**
+     * Appended to {@code .zshenv}, after its {@code $HOME} source: the third pass, guarded on the
+     * exact condition that defines the gap {@code .zshrc}/{@code .zlogin} do not cover — a shell
+     * that is neither login nor interactive. The sentinel export happens only once the scrub has
+     * already run, and only for as long as the current pane's environment has not been rebuilt from
+     * scratch (a fresh {@code env -i} child would not inherit it — that is out of scope here, since
+     * such a child is no longer running under the pane's own environment at all).
+     */
+    private static final String NEITHER_LOGIN_NOR_INTERACTIVE_SCRUB =
+            "if [[ ! -o login && ! -o interactive && -z \"${" + SCRUB_SENTINEL + ":-}\" ]]; then\n"
+            + "  " + SOURCE_SCRUB
+            + "  export " + SCRUB_SENTINEL + "=1\n"
+            + "fi\n";
+
     private EnvAllowListScrub() {
     }
 
@@ -105,11 +154,16 @@ public final class EnvAllowListScrub {
             reapOrphans(parentDir);
             Path dir = Files.createTempDirectory(parentDir, DIR_PREFIX);
             dir.toFile().deleteOnExit();
-            // The report is written by zsh, after these hooks are registered, so register its path
-            // too — otherwise the directory is non-empty at JVM exit and cannot be removed at all.
-            dir.resolve(REPORT_FILE).toFile().deleteOnExit();
-            write(dir, SCRUB_FILE, scrubScript(allowedNames));
-            write(dir, ".zshenv", homeSourcingFile(".zshenv"));
+            // zsh truncates this pre-created receipt after these hooks are registered. Register its
+            // path too — otherwise the directory is non-empty at JVM exit and cannot be removed.
+            Files.createFile(dir.resolve(REPORT_FILE)).toFile().deleteOnExit();
+            // fleetd #388: scrub.zsh's OWN allow-list must also keep SCRUB_SENTINEL, or a later
+            // pass in the same pane blanks it back to empty and the .zshenv guard below thinks it
+            // was never scrubbed — see the class javadoc.
+            Set<String> namesForScrubScript = new HashSet<>(allowedNames);
+            namesForScrubScript.add(SCRUB_SENTINEL);
+            write(dir, SCRUB_FILE, scrubScript(namesForScrubScript));
+            write(dir, ".zshenv", homeSourcingFile(".zshenv") + NEITHER_LOGIN_NOR_INTERACTIVE_SCRUB);
             write(dir, ".zprofile", homeSourcingFile(".zprofile"));
             write(dir, ".zshrc", homeSourcingFile(".zshrc") + SOURCE_SCRUB);
             write(dir, ".zlogin", homeSourcingFile(".zlogin") + SOURCE_SCRUB);
@@ -147,12 +201,11 @@ public final class EnvAllowListScrub {
      * into it: owner keeps full access, {@code group} gets traverse+read on the directory ({@code
      * rwxr-x---}, so a member process — a login shell reading it via {@code ZDOTDIR}, or another
      * process simply opening a file under it — running under that group can find and read the
-     * files) and read-only on each file ({@code rw-r-----}) — deliberately no group WRITE anywhere,
-     * since a member never needs to add or change what fleetd generated. (For the ZDOTDIR scrub
-     * specifically, this also means the scrub script's own report write inside the pane fails
-     * closed rather than open — see {@code scrub.zsh}'s trailing {@code 2>/dev/null} — which {@link
-     * dev.ltms.fleet.member.HerdrPeerLauncher#releaseZdotdir} already treats as "cannot be
-     * confirmed to have run" rather than success.)
+     * files) and read-only on each file ({@code rw-r-----}), except the pre-created ZDOTDIR
+     * {@code scrub-report.txt}. That receipt gets group write ({@code rw-rw----}), so
+     * {@code scrub.zsh} can truncate and write it without granting group write on the directory.
+     * If its optional permission change fails, the member cannot write a receipt and the launcher
+     * keeps its existing WARN rather than failing the spawn.
      *
      * <p>Package-private and named generically on purpose: fleetd #213 built this for the ZDOTDIR
      * scrub directory, and fleetd #219 reuses it verbatim for {@link
@@ -167,7 +220,15 @@ public final class EnvAllowListScrub {
             setGroupAndPermissions(dir, principal, "rwxr-x---");
             try (Stream<Path> entries = Files.list(dir)) {
                 for (Path file : entries.toList()) {
-                    setGroupAndPermissions(file, principal, "rw-r-----");
+                    if (REPORT_FILE.equals(file.getFileName().toString())) {
+                        try {
+                            setGroupAndPermissions(file, principal, "rw-rw----");
+                        } catch (IOException | UnsupportedOperationException ignored) {
+                            // The receipt is optional. Its absence keeps the existing WARN path.
+                        }
+                    } else {
+                        setGroupAndPermissions(file, principal, "rw-r-----");
+                    }
                 }
             }
         } catch (IOException e) {
@@ -245,11 +306,13 @@ public final class EnvAllowListScrub {
         }
         return """
                 # generated by fleetd (CB-633 memberCredentials policy=allow-list) — do not edit.
-                # Sourced from .zshrc and again from .zlogin, each time AFTER that file has sourced
-                # its $HOME counterpart — so this runs after everything the operator sourced, on a
-                # login shell (macOS panes) and on a plain interactive one (Linux panes) alike.
-                # Running twice is idempotent and deliberate: the second pass catches anything
-                # ~/.zlogin exported after ~/.zshrc had finished.
+                # Sourced unconditionally from .zshrc and again from .zlogin, each time AFTER that
+                # file has sourced its $HOME counterpart — so this runs after everything the
+                # operator sourced, on any pane that is login and/or interactive. Running twice is
+                # idempotent and deliberate: the second pass catches anything ~/.zlogin exported
+                # after ~/.zshrc had finished. Also sourced, once, from a guarded pass in .zshenv
+                # (fleetd #388) when the pane shell is NEITHER login nor interactive — the one gap
+                # those two files do not cover.
 
                 typeset -A _cb633_allowed
                 for _cb633_n in %s; do _cb633_allowed[$_cb633_n]=1; done

fleetd/src/test/java/dev/ltms/fleet/FleetdStartupValidationTest.java

@@ -0,0 +1,132 @@
+package dev.ltms.fleet;
+
+import org.junit.jupiter.api.Test;
+import org.junit.jupiter.api.io.TempDir;
+
+import java.nio.file.Files;
+import java.nio.file.Path;
+
+import static org.junit.jupiter.api.Assertions.assertThrows;
+import static org.junit.jupiter.api.Assertions.assertTrue;
+
+/**
+ * Proves the one thing no test proved before this ticket's follow-up: that {@code Fleetd.main}
+ * ITSELF — not a copy of its logic, not the validator called directly — still refuses to start on
+ * a bad config. Mutation testing found that deleting {@code cfg.validateAll();} (née six
+ * individual {@code cfg.validateXxx();} calls) from {@code Fleetd.main} left the full 1478-test
+ * suite green; every existing test called a validator directly and none exercised {@code
+ * Fleetd.main} as the caller. See {@code FleetConfigValidateAllTest} for why the fix collapses
+ * those six calls into one reflective {@link FleetConfig#validateAll()}, and {@code
+ * ConfigRefTest} for the equivalent proof on the {@link
+ * dev.ltms.fleet.config.ConfigRef#reload()} path.
+ *
+ * <p>This is deliberately the real {@code static void main(String[] args)} — package-private, so
+ * only a test in this package can call it, which is exactly what makes this proof strong: it is
+ * not a helper extracted for testability, it is the literal method {@code java -jar fleetd.jar}
+ * invokes. Every fixture below is otherwise-valid and fails exactly one validator, and — because
+ * {@link FleetConfig#validateAll()} runs immediately after {@code SubscriptionGuard.
+ * assertPrimaryClean}, before {@code main} opens the herdr socket, binds Javalin, or touches
+ * anything else with a real side effect — calling {@code Fleetd.main} with one of these configs is
+ * safe: it is guaranteed to throw before reaching any of that, precisely because the config is
+ * deliberately invalid.
+ */
+class FleetdStartupValidationTest {
+
+    private static void assertMainRefuses(Path dir, String fileName, String yaml,
+                                           String mustContain) throws Exception {
+        Path f = dir.resolve(fileName);
+        Files.writeString(f, yaml);
+        IllegalStateException e = assertThrows(IllegalStateException.class,
+                () -> Fleetd.main(new String[]{f.toString()}),
+                fileName + ": Fleetd.main must refuse this config before doing anything else");
+        assertTrue(e.getMessage().contains(mustContain),
+                fileName + ": expected message to contain \"" + mustContain + "\" but was: "
+                        + e.getMessage());
+    }
+
+    @Test
+    void mainRefusesANonLoopbackBindWithoutTokenMode(@TempDir Path dir) throws Exception {
+        assertMainRefuses(dir, "auth-exposure.yaml", """
+                bind:
+                  host: 0.0.0.0
+                  port: 8765
+                """, "auth.mode: token");
+    }
+
+    @Test
+    void mainRefusesALeadTabPrefixCollision(@TempDir Path dir) throws Exception {
+        assertMainRefuses(dir, "lead-tab-prefixes.yaml", """
+                bind:
+                  host: 127.0.0.1
+                  port: 8765
+                fleet:
+                  tabLabel: "lead: {role} {profile}"
+                  leaders:
+                    opus:
+                      tab: "lead: opus"
+                """, "fleet.tabLabel");
+    }
+
+    @Test
+    void mainRefusesASubscriptionProfileThatReseatsAnthropicBaseUrl(@TempDir Path dir)
+            throws Exception {
+        assertMainRefuses(dir, "subscription-profiles.yaml", """
+                bind:
+                  host: 127.0.0.1
+                  port: 8765
+                profiles:
+                  sonnet:
+                    subscription: true
+                    argv: ["ccs", "sonnet"]
+                    env:
+                      ANTHROPIC_BASE_URL: http://anything-not-on-the-allowlist
+                """, "ANTHROPIC_BASE_URL");
+    }
+
+    @Test
+    void mainRefusesAnUnknownCharterKey(@TempDir Path dir) throws Exception {
+        assertMainRefuses(dir, "charters.yaml", """
+                bind:
+                  host: 127.0.0.1
+                  port: 8765
+                fleet:
+                  charters:
+                    architetc: text
+                """, "architetc");
+    }
+
+    @Test
+    void mainRefusesAnArchitectSlotNamingAnUnconfiguredProfile(@TempDir Path dir) throws Exception {
+        assertMainRefuses(dir, "members.yaml", """
+                bind:
+                  host: 127.0.0.1
+                  port: 8765
+                profiles:
+                  gx10:
+                    baseUrl: http://gx10.gw:8000
+                fleet:
+                  architects:
+                    lead-designer:
+                      profile: sonnet
+                """, "lead-designer");
+    }
+
+    @Test
+    void mainRefusesAProfileNamingAModelOutsideTheAllowList(@TempDir Path dir) throws Exception {
+        assertMainRefuses(dir, "models.yaml", """
+                bind:
+                  host: 127.0.0.1
+                  port: 8765
+                profiles:
+                  sonnet:
+                    baseUrl: http://gx10.gw:8000
+                    model: claude-sonnet-5
+                  rogue:
+                    baseUrl: http://gx11.gw:8000
+                    model: claude-opus-9000
+                models:
+                  allow:
+                    - model: claude-sonnet-5
+                """, "rogue");
+    }
+}

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

@@ -109,6 +109,8 @@ class ConfigRefTopLevelReportingCoverageTest {
         v.put("memberLoginShell", null);
         v.put("memberSkills", "/skills/a");
         v.put("idleSleepGuard", new FleetConfig.IdleSleepGuard(true));
+        v.put("models", new FleetConfig.Models(
+                List.of(new FleetConfig.Models.ModelEntry("model-a"))));
         assertNamesMatchComponents(v);
         return v;
     }
@@ -151,6 +153,8 @@ class ConfigRefTopLevelReportingCoverageTest {
         v.put("memberLoginShell", null);
         v.put("memberSkills", "/skills/b");
         v.put("idleSleepGuard", new FleetConfig.IdleSleepGuard(false));
+        v.put("models", new FleetConfig.Models(
+                List.of(new FleetConfig.Models.ModelEntry("model-b"))));
         assertNamesMatchComponents(v);
         return v;
     }

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

@@ -2683,4 +2683,222 @@ class FleetConfigTest {
         assertTrue(cfg.idleSleepGuard().isEnabled(),
                 "unlike ConfigReload/Health, this block defaults to ON even when present but empty");
     }
+
+    // --- models: central allow-list of usable models -----------------------------------------
+
+    /**
+     * An absent {@code models:} block is today's behaviour exactly: no profile's {@code model:} is
+     * checked against anything, whatever it says. This is the "existing config keeps working"
+     * invariant — an operator on a gitignored {@code fleetd.yaml} that predates this feature must
+     * not be broken by upgrading the daemon.
+     */
+    @Test
+    void absentModelsBlockValidatesNothing(@TempDir Path dir) throws Exception {
+        Path f = dir.resolve("fleetd.yaml");
+        Files.writeString(f, """
+                bind:
+                  port: 8080
+                profiles:
+                  gx10:
+                    baseUrl: http://gx10.gw:8000
+                    model: totally-unheard-of-model-xyz
+                """);
+
+        FleetConfig cfg = FleetConfig.load(f);
+        assertDoesNotThrow(cfg::validateModels);
+    }
+
+    /**
+     * {@code models:} present but with an empty (or absent) {@code allow:} must behave exactly like
+     * the block being absent — an operator adding the block for the first time with nothing in it
+     * yet must not be surprised by every profile suddenly refusing to start.
+     */
+    @Test
+    void modelsBlockPresentButEmptyValidatesNothing(@TempDir Path dir) throws Exception {
+        Path f = dir.resolve("fleetd.yaml");
+        Files.writeString(f, """
+                bind:
+                  port: 8080
+                profiles:
+                  gx10:
+                    baseUrl: http://gx10.gw:8000
+                    model: whatever-the-operator-typed
+                models:
+                  allow: []
+                """);
+
+        FleetConfig cfg = FleetConfig.load(f);
+        assertDoesNotThrow(cfg::validateModels);
+    }
+
+    /**
+     * The core of the ticket: a profile naming a model outside the configured allow-list fails
+     * config load, naming both the model and the profile that wanted it.
+     */
+    @Test
+    void aProfileNamingAModelOutsideTheAllowListRefusesToStart(@TempDir Path dir) throws Exception {
+        Path f = dir.resolve("fleetd.yaml");
+        Files.writeString(f, """
+                bind:
+                  port: 8080
+                profiles:
+                  sonnet:
+                    baseUrl: http://gx10.gw:8000
+                    model: claude-sonnet-5
+                  rogue:
+                    baseUrl: http://gx11.gw:8000
+                    model: claude-opus-9000
+                models:
+                  allow:
+                    - model: claude-sonnet-5
+                    - model: claude-haiku-5
+                """);
+        FleetConfig cfg = FleetConfig.load(f);
+
+        IllegalStateException e = assertThrows(IllegalStateException.class, cfg::validateModels);
+        assertTrue(e.getMessage().contains("rogue"), "the refusal names the offending profile");
+        assertTrue(e.getMessage().contains("claude-opus-9000"), "the refusal names the offending model");
+        assertFalse(e.getMessage().contains("'sonnet'"),
+                "the profile whose model IS allowed must not be reported");
+    }
+
+    /** A profile whose {@code model:} is on the allow-list loads fine. */
+    @Test
+    void aProfileNamingAModelOnTheAllowListLoadsFine(@TempDir Path dir) throws Exception {
+        Path f = dir.resolve("fleetd.yaml");
+        Files.writeString(f, """
+                bind:
+                  port: 8080
+                profiles:
+                  sonnet:
+                    baseUrl: http://gx10.gw:8000
+                    model: claude-sonnet-5
+                models:
+                  allow:
+                    - model: claude-sonnet-5
+                """);
+
+        FleetConfig cfg = FleetConfig.load(f);
+        assertDoesNotThrow(cfg::validateModels);
+    }
+
+    /**
+     * A profile that never sets {@code model:} (a {@code subscription: true} profile relying on the
+     * account's own default is the live example) must not be refused just because an allow-list is
+     * active — there is nothing to check it against.
+     */
+    @Test
+    void aProfileWithNoModelPassesEvenWithAnActiveAllowList(@TempDir Path dir) throws Exception {
+        Path f = dir.resolve("fleetd.yaml");
+        Files.writeString(f, """
+                bind:
+                  port: 8080
+                profiles:
+                  opus:
+                    subscription: true
+                models:
+                  allow:
+                    - model: claude-sonnet-5
+                """);
+
+        FleetConfig cfg = FleetConfig.load(f);
+        assertDoesNotThrow(cfg::validateModels);
+    }
+
+    /**
+     * Reproduces the live config shape this ticket measured: a mix of {@code amazon-bedrock},
+     * {@code opencode} and {@code openai}-backed profiles, some naming a bare Claude id and some a
+     * provider-prefixed opencode id, in ONE allow-list. Both forms are just opaque strings compared
+     * for exact equality — proves the shape decision holds against the shape actually seen live,
+     * not just against a synthetic single-provider example.
+     */
+    @Test
+    void aBareClaudeIdAndAnOpencodeProviderPrefixedIdBothFitOneAllowList(@TempDir Path dir) throws Exception {
+        Path f = dir.resolve("fleetd.yaml");
+        Files.writeString(f, """
+                bind:
+                  port: 8080
+                profiles:
+                  sonnet:
+                    baseUrl: http://gx10.gw:8000
+                    model: claude-sonnet-5
+                  terra:
+                    kind: opencode
+                    model: openai/gpt-5.6-terra
+                  nova:
+                    kind: opencode
+                    model: amazon-bedrock/amazon.nova-pro-v1:0
+                models:
+                  allow:
+                    - model: claude-sonnet-5
+                    - model: openai/gpt-5.6-terra
+                    - model: amazon-bedrock/amazon.nova-pro-v1:0
+                """);
+
+        FleetConfig cfg = FleetConfig.load(f);
+        assertDoesNotThrow(cfg::validateModels);
+    }
+
+    /** The same live shape, but one opencode profile's model is missing from the allow-list. */
+    @Test
+    void anUnlistedOpencodeProviderPrefixedModelRefusesToStart(@TempDir Path dir) throws Exception {
+        Path f = dir.resolve("fleetd.yaml");
+        Files.writeString(f, """
+                bind:
+                  port: 8080
+                profiles:
+                  sonnet:
+                    baseUrl: http://gx10.gw:8000
+                    model: claude-sonnet-5
+                  terra:
+                    kind: opencode
+                    model: openai/gpt-5.6-terra-withdrawn
+                models:
+                  allow:
+                    - model: claude-sonnet-5
+                    - model: openai/gpt-5.6-terra
+                """);
+        FleetConfig cfg = FleetConfig.load(f);
+
+        IllegalStateException e = assertThrows(IllegalStateException.class, cfg::validateModels);
+        assertTrue(e.getMessage().contains("terra"), "the refusal names the offending profile");
+        assertTrue(e.getMessage().contains("openai/gpt-5.6-terra-withdrawn"),
+                "the refusal names the offending model, with its provider prefix intact");
+    }
+
+    /**
+     * Editing a {@code profiles:} entry alone must never be able to widen what is permitted — only
+     * editing {@code models.allow:} itself can. This is the invariant the ticket states explicitly;
+     * this test pins it by giving a profile a plausible-looking model that was never added to the
+     * allow-list and confirming it is still refused.
+     */
+    @Test
+    void addingAProfileCannotWidenTheAllowListByItself(@TempDir Path dir) throws Exception {
+        Path f = dir.resolve("fleetd.yaml");
+        Files.writeString(f, """
+                bind:
+                  port: 8080
+                profiles:
+                  sonnet:
+                    baseUrl: http://gx10.gw:8000
+                    model: claude-sonnet-5
+                  brand-new:
+                    baseUrl: http://gx12.gw:8000
+                    model: claude-sonnet-6-preview
+                models:
+                  allow:
+                    - model: claude-sonnet-5
+                """);
+        FleetConfig cfg = FleetConfig.load(f);
+
+        IllegalStateException e = assertThrows(IllegalStateException.class, cfg::validateModels);
+        assertTrue(e.getMessage().contains("brand-new"));
+        assertTrue(e.getMessage().contains("claude-sonnet-6-preview"));
+    }
+
+    /** {@code models} is a brand-new top-level key and must be recognized, not WARN-ed as unknown. */
+    @Test
+    void modelsIsAKnownTopLevelKey() {
+        assertTrue(FleetConfig.KNOWN_TOP_LEVEL_KEYS.contains("models"));
+    }
 }

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

@@ -0,0 +1,358 @@
+package dev.ltms.fleet.config;
+
+import org.junit.jupiter.api.Test;
+import org.junit.jupiter.api.io.TempDir;
+
+import java.lang.reflect.Method;
+import java.nio.file.Files;
+import java.nio.file.Path;
+import java.util.ArrayList;
+import java.util.List;
+import java.util.Set;
+import java.util.TreeSet;
+
+import static org.junit.jupiter.api.Assertions.assertDoesNotThrow;
+import static org.junit.jupiter.api.Assertions.assertEquals;
+import static org.junit.jupiter.api.Assertions.assertThrows;
+import static org.junit.jupiter.api.Assertions.assertTrue;
+
+/**
+ * The gap this class exists to close: mutation testing on the fleetd ticket "central allow-list
+ * of usable models" found that although {@link FleetConfig#validateModels()}'s own logic was well
+ * pinned, nothing proved either real caller ({@code Fleetd.main} and {@link ConfigRef#reload()})
+ * still invoked it — deleting the call site left the full suite green (1478/0/0/0). A follow-up
+ * measurement (same technique — remove one call site, run the suite, not read the code) found the
+ * SAME gap for all five of {@link FleetConfig}'s other validators at startup, and for four of the
+ * six inside {@link ConfigRef#reload()}. This is a class of gap, not one line's mistake: every one
+ * of those thirteen tests called the validator itself directly, never the real caller that was
+ * supposed to.
+ *
+ * <p>The fix replaces the six individual {@code cfg.validateXxx()} calls at each of the two real
+ * call sites with one {@link FleetConfig#validateAll()}, which reaches every validator by
+ * reflection rather than by a hand-maintained list of names. A hand-maintained list of six names
+ * would have exactly the defect it replaces: the seventh validator someone adds next month has no
+ * reason to be added to it, and nothing would say so. This class proves TWO separate claims, and
+ * keeps them separate on purpose:
+ *
+ * <ol>
+ *   <li>{@link #theSweepMechanismIsGenericNotHardcodedToFleetConfigsSixNames()} and its neighbours
+ *       prove the reflective sweep itself ({@link FleetConfig#invokeAllValidators}) is a general
+ *       mechanism — it runs whatever public, no-arg, void {@code validateXxx()} methods a class
+ *       happens to declare today, including a class with more of them than {@link FleetConfig}
+ *       has right now. This is the proof that a future, real seventh validator on {@link
+ *       FleetConfig} would be swept automatically, without needing to add a real (unwanted)
+ *       seventh validator just to exercise the claim.</li>
+ *   <li>{@link #validateAllReachesEveryOneOfTodaysSixValidators()} proves {@link
+ *       FleetConfig#validateAll()} itself is wired to that same generic mechanism and genuinely
+ *       reaches each of today's six real validators — reusing the exact minimal failing
+ *       configurations {@code FleetConfigTest} already established for each one directly, so a
+ *       single call to {@code validateAll()} is shown to reproduce every one of those six
+ *       failures.</li>
+ * </ol>
+ *
+ * <p>Together with the direct-{@code Fleetd.main}-invocation tests in {@code
+ * FleetdStartupValidationTest} (which prove the real startup call site still calls {@code
+ * validateAll()}) and the {@code ConfigRefTest} reload tests (which prove the same for {@link
+ * ConfigRef#reload()}), removing {@code cfg.validateAll();} from either real call site now fails
+ * a test in this module.
+ *
+ * <p><b>What is NOT pinned, measured rather than assumed.</b> Reverting {@link
+ * FleetConfig#validateAll()} to a hardcoded list of today's six method calls leaves the whole
+ * suite green (measured at review: 1491 tests, 0 failures). Nothing ties {@code validateAll()} to
+ * the generic sweep — claim 1 proves {@link FleetConfig#invokeAllValidators} is generic, and claim
+ * 2 proves {@code validateAll()} reaches today's six, and a hardcoded list satisfies both. So the
+ * reflective sweep is a convenience, not the guarantee. The guarantee is {@link
+ * #fleetConfigDeclaresExactlyTheseSixValidatorsToday()}: it fails the moment a seventh validator
+ * is declared, which forces whoever adds it to look at this file.
+ */
+class FleetConfigValidateAllTest {
+
+    // ── Claim 1: the reflective sweep is a general mechanism, not six names in disguise ──────────
+
+    /**
+     * A throwaway fixture class, unrelated to {@link FleetConfig} in every way except shape: three
+     * public, no-arg, void methods named {@code validateXxx}. Proves the sweep works on ANY class
+     * with this shape, not on something special-cased to {@link FleetConfig}.
+     */
+    static class ThreeValidators {
+        final List<String> ran = new ArrayList<>();
+
+        public void validateAlpha() {
+            ran.add("validateAlpha");
+        }
+
+        public void validateBeta() {
+            ran.add("validateBeta");
+        }
+
+        public void validateGamma() {
+            ran.add("validateGamma");
+        }
+    }
+
+    @Test
+    void theSweepMechanismIsGenericNotHardcodedToFleetConfigsSixNames() {
+        ThreeValidators target = new ThreeValidators();
+        FleetConfig.invokeAllValidators(target);
+        assertEquals(List.of("validateAlpha", "validateBeta", "validateGamma"), target.ran,
+                "every validateXxx() method on this unrelated class must run, in alphabetical "
+                        + "order — the sweep reads the class's own shape, not a name FleetConfig "
+                        + "happens to know about");
+    }
+
+    /**
+     * The core of the "self-maintaining" requirement: the exact same class shape as {@link
+     * ThreeValidators}, plus one more method — standing in for "a developer adds a validator next
+     * month". Nothing about the sweep changes to pick it up; the new method is invoked purely
+     * because it exists and matches the shape. This is what makes adding a seventh real validator
+     * to {@link FleetConfig} safe without touching {@link FleetConfig#validateAll()} or either
+     * call site — there is no "wire it in" step left to forget.
+     */
+    static class FourValidators {
+        final List<String> ran = new ArrayList<>();
+
+        public void validateAlpha() {
+            ran.add("validateAlpha");
+        }
+
+        public void validateBeta() {
+            ran.add("validateBeta");
+        }
+
+        public void validateGamma() {
+            ran.add("validateGamma");
+        }
+
+        public void validateDelta() {
+            ran.add("validateDelta");
+        }
+    }
+
+    @Test
+    void addingAFourthValidatorMethodGetsSweptWithNoOtherChange() {
+        FourValidators target = new FourValidators();
+        FleetConfig.invokeAllValidators(target);
+        assertEquals(List.of("validateAlpha", "validateBeta", "validateDelta", "validateGamma"),
+                sorted(target.ran),
+                "the fourth method must be reached automatically — proving a class can grow the "
+                        + "set of things it validates with no change to the sweep itself");
+    }
+
+    private static List<String> sorted(List<String> in) {
+        List<String> copy = new ArrayList<>(in);
+        copy.sort(String::compareTo);
+        return copy;
+    }
+
+    @Test
+    void aFailingValidatorStopsTheSweepAndPropagatesUnchanged() {
+        class OneFails {
+            @SuppressWarnings("unused")
+            public void validateOk() {
+                // passes
+            }
+
+            public void validateBoom() {
+                throw new IllegalStateException("refusing to start: boom");
+            }
+        }
+        IllegalStateException e = assertThrows(IllegalStateException.class,
+                () -> FleetConfig.invokeAllValidators(new OneFails()));
+        assertEquals("refusing to start: boom", e.getMessage(),
+                "the real exception must propagate unchanged, not be wrapped or swallowed");
+    }
+
+    /**
+     * Every rule the sweep's filter applies, proven independently: only public, no-arg, void
+     * methods whose name starts with {@code "validate"} run, {@code validateAll} itself is
+     * excluded (so a class that declares one of its own — as {@link FleetConfig} does — cannot
+     * recurse into itself), and a same-shaped-but-wrongly-named or wrongly-shaped method never
+     * runs. A reader who "simplifies" the filter in {@link FleetConfig#invokeAllValidators} in a
+     * way that widens or narrows it breaks one of these.
+     */
+    static class FilterEdgeCases {
+        final List<String> ran = new ArrayList<>();
+
+        public void validateReal() {
+            ran.add("validateReal");
+        }
+
+        /** Wrong name — must not run. */
+        public void checkSomething() {
+            ran.add("checkSomething");
+        }
+
+        /** Wrong shape — takes an argument. */
+        public void validateWithArg(String ignored) {
+            ran.add("validateWithArg");
+        }
+
+        /** Wrong shape — returns something. */
+        public boolean validateReturnsBoolean() {
+            ran.add("validateReturnsBoolean");
+            return true;
+        }
+
+        /** Excluded by name on purpose, so the sweep cannot call itself. */
+        public void validateAll() {
+            ran.add("validateAll");
+        }
+    }
+
+    @Test
+    void onlyPublicNoArgVoidValidateNamedMethodsRun() {
+        FilterEdgeCases target = new FilterEdgeCases();
+        FleetConfig.invokeAllValidators(target);
+        assertEquals(List.of("validateReal"), target.ran,
+                "checkSomething (wrong name), validateWithArg (wrong shape), "
+                        + "validateReturnsBoolean (wrong shape), and validateAll (excluded by "
+                        + "name) must all be skipped");
+    }
+
+    // ── Claim 2: FleetConfig.validateAll() is wired to that mechanism and reaches all six today ──
+
+    /**
+     * Reflectively enumerates {@link FleetConfig}'s own public, no-arg, void {@code validateXxx()}
+     * methods (excluding {@code validateAll} itself) — the exact same filter {@link
+     * FleetConfig#invokeAllValidators} applies. This is not the mechanism proof (that is claim 1,
+     * above, on an unrelated class) — it is a visible denominator: today there are six, named
+     * here, so a reader adding a seventh sees this assertion name the new count rather than a
+     * silent pass at the old one.
+     */
+    @Test
+    void fleetConfigDeclaresExactlyTheseSixValidatorsToday() {
+        Set<String> names = new TreeSet<>();
+        for (Method m : FleetConfig.class.getMethods()) {
+            if (java.lang.reflect.Modifier.isPublic(m.getModifiers())
+                    && m.getParameterCount() == 0
+                    && m.getReturnType() == void.class
+                    && m.getName().startsWith("validate")
+                    && !m.getName().equals("validateAll")) {
+                names.add(m.getName());
+            }
+        }
+        assertEquals(new TreeSet<>(Set.of("validateAuthExposure", "validateLeadTabPrefixes",
+                "validateSubscriptionProfiles", "validateCharters", "validateMembers",
+                "validateModels")), names,
+                "FleetConfig's public validate*() methods changed. Do TWO things, in this "
+                        + "order. First confirm validateAll() still delegates to "
+                        + "invokeAllValidators(this) — a hardcoded list there passes every other "
+                        + "test in this class, so this assertion is the only place that will ever "
+                        + "make you check. Only then update the expected set to match.");
+    }
+
+    /** A minimal, otherwise-valid file — same shape FleetConfigTest and ConfigRefTest use. */
+    private static String minimalValidYaml() {
+        return """
+                bind:
+                  host: 127.0.0.1
+                  port: 8765
+                profiles:
+                  sonnet:
+                    baseUrl: http://gx00.gw:8000
+                    model: sonnet
+                """;
+    }
+
+    @Test
+    void aFullyValidConfigPassesValidateAll(@TempDir Path dir) throws Exception {
+        Path f = dir.resolve("fleetd.yaml");
+        Files.writeString(f, minimalValidYaml());
+        assertDoesNotThrow(() -> FleetConfig.load(f).validateAll());
+    }
+
+    /**
+     * The heart of claim 2: for each of today's six real validators, a minimal file that fails
+     * ONLY that one — the exact fixtures {@code FleetConfigTest} uses to test each validator
+     * directly — must also fail through {@link FleetConfig#validateAll()}. If a future edit to
+     * {@code validateAll()} silently dropped one validator from the sweep (e.g. a typo'd name
+     * filter), exactly one of these six would start passing when it must not.
+     */
+    @Test
+    void validateAllReachesEveryOneOfTodaysSixValidators(@TempDir Path dir) throws Exception {
+        // validateAuthExposure: a non-loopback bind without token mode.
+        assertValidateAllRefuses(dir, "auth-exposure.yaml", """
+                bind:
+                  host: 0.0.0.0
+                  port: 8765
+                """, "auth.mode: token");
+
+        // validateLeadTabPrefixes: a fleet-wide tabLabel that starts with a lead's own tabPrefix.
+        assertValidateAllRefuses(dir, "lead-tab-prefixes.yaml", """
+                bind:
+                  host: 127.0.0.1
+                  port: 8765
+                fleet:
+                  tabLabel: "lead: {role} {profile}"
+                  leaders:
+                    opus:
+                      tab: "lead: opus"
+                """, "fleet.tabLabel");
+
+        // validateSubscriptionProfiles: subscription: true with env: reseating ANTHROPIC_BASE_URL.
+        assertValidateAllRefuses(dir, "subscription-profiles.yaml", """
+                bind:
+                  host: 127.0.0.1
+                  port: 8765
+                profiles:
+                  sonnet:
+                    subscription: true
+                    argv: ["ccs", "sonnet"]
+                    env:
+                      ANTHROPIC_BASE_URL: http://anything-not-on-the-allowlist
+                """, "ANTHROPIC_BASE_URL");
+
+        // validateCharters: a charter key that is not a role wire name.
+        assertValidateAllRefuses(dir, "charters.yaml", """
+                bind:
+                  host: 127.0.0.1
+                  port: 8765
+                fleet:
+                  charters:
+                    architetc: text
+                """, "architetc");
+
+        // validateMembers: an architect slot naming an unconfigured profile.
+        assertValidateAllRefuses(dir, "members.yaml", """
+                bind:
+                  host: 127.0.0.1
+                  port: 8765
+                profiles:
+                  gx10:
+                    baseUrl: http://gx10.gw:8000
+                fleet:
+                  architects:
+                    lead-designer:
+                      profile: sonnet
+                """, "lead-designer");
+
+        // validateModels: a profile naming a model outside the configured allow-list.
+        assertValidateAllRefuses(dir, "models.yaml", """
+                bind:
+                  host: 127.0.0.1
+                  port: 8765
+                profiles:
+                  sonnet:
+                    baseUrl: http://gx10.gw:8000
+                    model: claude-sonnet-5
+                  rogue:
+                    baseUrl: http://gx11.gw:8000
+                    model: claude-opus-9000
+                models:
+                  allow:
+                    - model: claude-sonnet-5
+                """, "rogue");
+    }
+
+    private static void assertValidateAllRefuses(Path dir, String fileName, String yaml,
+                                                  String mustContain) throws Exception {
+        Path f = dir.resolve(fileName);
+        Files.writeString(f, yaml);
+        FleetConfig cfg = FleetConfig.load(f);
+        IllegalStateException e = assertThrows(IllegalStateException.class, cfg::validateAll,
+                fileName + ": validateAll() must refuse this config");
+        assertTrue(e.getMessage().contains(mustContain),
+                fileName + ": expected message to contain \"" + mustContain + "\" but was: "
+                        + e.getMessage());
+    }
+}

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

@@ -97,6 +97,8 @@ class FleetConfigWithDefaultsPreservesEveryComponentTest {
         v.put("memberLoginShell", "/bin/zsh");
         v.put("memberSkills", "/skills/guard");
         v.put("idleSleepGuard", new FleetConfig.IdleSleepGuard(true));
+        v.put("models", new FleetConfig.Models(
+                List.of(new FleetConfig.Models.ModelEntry("model-guard"))));
         assertNamesMatchComponents(v);
         return v;
     }

fleetd/src/test/java/dev/ltms/fleet/health/FleetHealthMonitorTest.java

@@ -30,6 +30,7 @@ import java.util.concurrent.ScheduledFuture;
 import java.util.concurrent.ScheduledThreadPoolExecutor;
 import java.util.concurrent.TimeUnit;
 import java.util.concurrent.atomic.AtomicLong;
+import java.util.concurrent.atomic.AtomicReference;
 import java.util.function.BiConsumer;
 import java.util.function.LongSupplier;
 
@@ -206,6 +207,139 @@ class FleetHealthMonitorTest {
                 .workingSuspectAfterOrDefault());
     }
 
+    // --- fleetd #386: a stall detector whose only clock freezes with a sleeping host is worse
+    // than a silent one — it reports "quiet" for a member that was genuinely busy for hours.
+
+    @Test void monotonicClockFrozenPastThresholdOnRealClockStillReportsStallSuspected() {
+        Logger logger = (Logger) LoggerFactory.getLogger(FleetHealthMonitor.class);
+        ListAppender<ILoggingEvent> appender = new ListAppender<>();
+        appender.start();
+        logger.addAppender(appender);
+        try {
+            AtomicLong mono = new AtomicLong(0);
+            AtomicLong real = new AtomicLong(0);
+            FakeHerdr herdr = new FakeHerdr().withAgent("busy", "term_busy", "pane_busy", "tab_busy");
+            var scheduler = Executors.newSingleThreadScheduledExecutor();
+            FleetHealthMonitor monitor = monitorWithClocks(herdr,
+                    List.of(member("term_busy", MemberSession.State.BUSY, 0, 0)), scheduler,
+                    mono::get, real::get, 60, 600, (_, _) -> { });
+
+            monitor.tick(); // establishes the clock baseline; nothing has diverged yet
+            assertEquals(0, appender.list.stream().filter(event -> event.getFormattedMessage()
+                    .contains("state=STALL_SUSPECTED")).count());
+
+            // The host "sleeps": the monotonic clock stands completely still while the real clock
+            // keeps moving, past the 600s stall threshold.
+            real.set(TimeUnit.SECONDS.toNanos(700));
+            monitor.tick();
+            monitor.stop();
+
+            assertTrue(appender.list.stream().anyMatch(event -> event.getFormattedMessage()
+                            .contains("member=term_busy state=STALL_SUSPECTED")),
+                    "the real clock crossed the stall threshold even though the monotonic clock never moved");
+        } finally {
+            logger.detachAppender(appender);
+        }
+    }
+
+    @Test void clockDivergenceIsLoggedOnceNotOncePerTick() {
+        Logger logger = (Logger) LoggerFactory.getLogger(FleetHealthMonitor.class);
+        ListAppender<ILoggingEvent> appender = new ListAppender<>();
+        appender.start();
+        logger.addAppender(appender);
+        try {
+            AtomicLong mono = new AtomicLong(0);
+            AtomicLong real = new AtomicLong(0);
+            var scheduler = Executors.newSingleThreadScheduledExecutor();
+            FleetHealthMonitor monitor = monitorWithClocks(new FakeHerdr(), List.of(), scheduler,
+                    mono::get, real::get, 60, 600, (_, _) -> { });
+
+            monitor.tick(); // baseline: no divergence possible yet
+
+            // One sleep gap: the monotonic clock is frozen while the real clock jumps far past one
+            // tick interval (60s).
+            real.set(TimeUnit.SECONDS.toNanos(700));
+            monitor.tick();
+
+            // The host is awake again: both clocks advance together from here, so no more divergence.
+            mono.set(TimeUnit.SECONDS.toNanos(10));
+            real.set(TimeUnit.SECONDS.toNanos(710));
+            monitor.tick();
+            mono.set(TimeUnit.SECONDS.toNanos(20));
+            real.set(TimeUnit.SECONDS.toNanos(720));
+            monitor.tick();
+            monitor.stop();
+
+            assertEquals(1, appender.list.stream().filter(event -> event.getFormattedMessage()
+                            .contains("the monotonic clock did not advance"))
+                    .count(), "one sleep gap must produce exactly one divergence line, not one per tick");
+        } finally {
+            logger.detachAppender(appender);
+        }
+    }
+
+    /**
+     * fleetd #386 follow-up, added on merge. The fix carries a PER-MEMBER drift baseline, so drift
+     * from a sleep that happened BEFORE a member went busy is never charged to that member. The
+     * two tests shipped with the fix both start with the member already BUSY, so a single global
+     * baseline passes them — this one fails without the per-member map.
+     *
+     * <p>Order matters: the host sleeps while nothing is busy, and only then does a member take a
+     * turn. Its stall clock must start at zero.
+     */
+    @Test void driftFromASleepBeforeAMemberWentBusyIsNotChargedToThatMember() {
+        Logger logger = (Logger) LoggerFactory.getLogger(FleetHealthMonitor.class);
+        ListAppender<ILoggingEvent> appender = new ListAppender<>();
+        appender.start();
+        logger.addAppender(appender);
+        try {
+            AtomicLong mono = new AtomicLong(0);
+            AtomicLong real = new AtomicLong(0);
+            AtomicReference<List<MemberSession>> roster = new AtomicReference<>(List.of());
+            FakeHerdr herdr = new FakeHerdr().withAgent("busy", "term_busy", "pane_busy", "tab_busy");
+            var scheduler = Executors.newSingleThreadScheduledExecutor();
+            AgentControl agents = new AgentControl(herdr);
+            FleetHealthMonitor monitor = new FleetHealthMonitor(agents, roster::get,
+                    new MessageService(agents, new Injector(agents), new Rendezvous(), new InMemoryReplyInbox()),
+                    scheduler, mono::get, real::get, 60, 600, (_, _) -> { });
+
+            monitor.tick();                                   // baseline, no members yet
+
+            // The host sleeps for 700s with nobody busy: the monotonic clock stands still.
+            real.set(TimeUnit.SECONDS.toNanos(700));
+            monitor.tick();
+
+            // Awake again. Only NOW does a member start a turn, with a fresh activity stamp taken
+            // from the monotonic clock. Both clocks advance together from here.
+            mono.set(TimeUnit.SECONDS.toNanos(10));
+            real.set(TimeUnit.SECONDS.toNanos(710));
+            roster.set(List.of(member("term_busy", MemberSession.State.BUSY,
+                    0, TimeUnit.SECONDS.toNanos(10))));
+            monitor.tick();
+
+            mono.set(TimeUnit.SECONDS.toNanos(20));
+            real.set(TimeUnit.SECONDS.toNanos(720));
+            monitor.tick();
+            monitor.stop();
+
+            assertEquals(0, appender.list.stream().filter(event -> event.getFormattedMessage()
+                            .contains("member=term_busy state=STALL_SUSPECTED")).count(),
+                    "the member has been busy for 10s, not 710s — the earlier sleep is not its stall");
+        } finally {
+            logger.detachAppender(appender);
+        }
+    }
+
+    private static FleetHealthMonitor monitorWithClocks(FakeHerdr herdr, List<MemberSession> roster,
+            java.util.concurrent.ScheduledExecutorService scheduler, LongSupplier clock,
+            LongSupplier realtimeClock, long intervalSeconds, long workingSuspectAfterSeconds,
+            BiConsumer<String, String> failTarget) {
+        AgentControl agents = new AgentControl(herdr);
+        return new FleetHealthMonitor(agents, () -> roster,
+                new MessageService(agents, new Injector(agents), new Rendezvous(), new InMemoryReplyInbox()),
+                scheduler, clock, realtimeClock, intervalSeconds, workingSuspectAfterSeconds, failTarget);
+    }
+
     @Test void goneMemberRecoveryLogsOnceWithoutRefiringTargetFailure() {
         Logger logger = (Logger) LoggerFactory.getLogger(FleetHealthMonitor.class);
         Level previousLevel = logger.getLevel();

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

@@ -8,6 +8,7 @@ import java.nio.charset.StandardCharsets;
 import java.nio.file.Files;
 import java.nio.file.Path;
 import java.util.ArrayList;
+import java.util.HashMap;
 import java.util.HashSet;
 import java.util.List;
 import java.util.Map;
@@ -45,6 +46,15 @@ class EnvAllowListScrubTest {
     /** Env var names appearing in command output; anything else (prompts, wrapped lines) is noise. */
     private static final Pattern ENV_NAME = Pattern.compile("^([A-Za-z_][A-Za-z0-9_]*)$");
 
+    /**
+     * fleetd #388: the sentinel name the generated {@code .zshenv} guard uses, kept here as a
+     * literal rather than referencing {@link EnvAllowListScrub#SCRUB_SENTINEL} — the two tests that
+     * use it must still compile and run against the pre-fix production class (which has no such
+     * constant), so the revert-and-prove-it-fails step exercises a real assertion instead of a
+     * compilation error.
+     */
+    private static final String SCRUB_SENTINEL_NAME = "_CB633_SCRUBBED";
+
     /**
      * The equality test. Expected survivors = baseline exports ∩ allowed — i.e. every survivor is
      * allowed AND every allowed name that existed survives. The operator's own secret-store exports
@@ -108,6 +118,31 @@ class EnvAllowListScrubTest {
                 "allowed N of M with N <= M — the denominator is always reported");
     }
 
+    /** A group-shared ZDOTDIR still lets the member truncate and write its pre-created receipt. */
+    @Test
+    void groupSharedScrubWritesAndReadsItsReport(@TempDir Path tmp) throws Exception {
+        assumeTrue(Files.isExecutable(ZSH), "/bin/zsh not present — nothing to prove here");
+        Set<String> allowed = MemberEnvAllowList.derive(List.of());
+        Path zdotdir = EnvAllowListScrub.generate(tmp, allowed, currentUserGroup());
+
+        Map<String, String> cleanParent = Map.of(
+                "HOME", System.getProperty("user.home"),
+                "PATH", "/usr/bin:/bin",
+                "SHELL", "/bin/zsh");
+        exportedNamesFromCleanParent(cleanParent, zdotdir);
+
+        EnvAllowListScrub.ScrubReport report = EnvAllowListScrub.readReport(zdotdir);
+        assertNotNull(report, "a group-shared completed login shell must leave a report behind");
+        assertTrue(report.allowed() >= 0 && report.total() >= report.allowed(),
+                "allowed N of M with N <= M — the denominator is always reported");
+        assertEquals("rw-rw----", java.nio.file.attribute.PosixFilePermissions.toString(
+                        Files.getPosixFilePermissions(zdotdir.resolve(EnvAllowListScrub.REPORT_FILE))),
+                "the pre-created receipt must be group-writable");
+        assertEquals("rwxr-x---", java.nio.file.attribute.PosixFilePermissions.toString(
+                        Files.getPosixFilePermissions(zdotdir)),
+                "group sharing must not make the ZDOTDIR directory group-writable");
+    }
+
     /** Report parsing is lenient: absent file → null (no measurement), not an exception. */
     @Test
     void readReportReturnsNullForADirectoryWithoutOne(@TempDir Path dir) {
@@ -224,6 +259,140 @@ class EnvAllowListScrubTest {
                         + "shell. A difference here means the scrub is dead on Linux.");
     }
 
+    /**
+     * fleetd #388: the actual gap. zsh reads {@code .zshenv} always, {@code .zprofile}/
+     * {@code .zlogin} only for a LOGIN shell, and {@code .zshrc} only for an INTERACTIVE one — so a
+     * shell that is NEITHER (a bare {@code /bin/zsh} reading a script off a non-tty stdin, no
+     * {@code -l}, no {@code -i}) reads only {@code .zshenv} and stops. Before this fix, that shell
+     * never reached {@code scrub.zsh} at all: the decoy secret below would survive untouched. This
+     * test injects that decoy directly into the process environment (not via a sourced dotfile,
+     * since the whole point of the gap is that {@code .zshenv} is normally close to empty) so the
+     * test does not depend on any real {@code ~/.zshrc} content existing on the host.
+     */
+    @Test
+    void scrubRunsInAShellThatIsNeitherLoginNorInteractive(@TempDir Path tmp) throws Exception {
+        assumeTrue(Files.isExecutable(ZSH), "/bin/zsh not present — nothing to prove here");
+        Set<String> allowed = MemberEnvAllowList.derive(List.of());
+        Path zdotdir = EnvAllowListScrub.generate(tmp, allowed);
+
+        Map<String, String> cleanParent = new HashMap<>(Map.of(
+                "HOME", System.getProperty("user.home"),
+                "PATH", "/usr/bin:/bin",
+                "SHELL", "/bin/zsh",
+                "USER", System.getProperty("user.name", "nobody"),
+                "TMPDIR", tmp.toString()));
+        cleanParent.put("FLEETD_TEST_DECOY_SECRET", "x"); // not on any allow-list; must be blanked
+
+        List<String> neither = List.of(); // no -l, no -i; stdin is a pipe (never a tty) either way
+        Set<String> baseline = exportedNamesFromCleanParent(cleanParent, null, neither);
+        Set<String> scrubbed = exportedNamesFromCleanParent(cleanParent, zdotdir, neither);
+
+        Set<String> expected = new TreeSet<>();
+        for (String name : baseline) {
+            if (MemberEnvAllowList.keeps(allowed, name)) {
+                expected.add(name);
+            }
+        }
+        assertTrue(baseline.contains("FLEETD_TEST_DECOY_SECRET"),
+                "sanity: the decoy must actually reach the un-scrubbed baseline, or this test proves "
+                        + "nothing");
+        expected.add("ZDOTDIR"); // the harness set it and it is infrastructure, so it must survive
+        expected.add(SCRUB_SENTINEL_NAME); // set by the new .zshenv guard once scrubbed
+        assertEquals(expected, scrubbed,
+                "a pane shell that is NEITHER login nor interactive must still be scrubbed — its "
+                        + "surviving exported names must EQUAL baseline ∩ allow-list, plus the "
+                        + "sentinel the guard sets once it has run. FLEETD_TEST_DECOY_SECRET surviving "
+                        + "here means the gap is still open.");
+    }
+
+    /**
+     * fleetd #388 invariants 3 and 4, which a name-set equality cannot show: a member's own tooling
+     * forks plain, non-login, non-interactive zsh processes for a single command (the same shape as
+     * the pane shell itself), and such a child must (a) keep whatever its parent deliberately set
+     * for it, never (b) re-run the scrub and blank it, and never (c) overwrite the pane's own
+     * {@code scrub-report.txt} with a description of itself instead of the pane. All three can only
+     * be shown by actually running a child process from within the scrubbed pane shell.
+     *
+     * <p>The pane and the child both report presence via {@code ${NAME:+present}} — empty when a
+     * name is unset OR blanked (exported empty), {@code present} when it is set and non-empty. No
+     * value is ever printed, only these two shapes and the literal word {@code set}/{@code unset}
+     * for the sentinel.
+     */
+    @Test
+    void neitherShellChildKeepsParentVariablesAndReceiptStillDescribesThePane(@TempDir Path tmp) throws Exception {
+        assumeTrue(Files.isExecutable(ZSH), "/bin/zsh not present — nothing to prove here");
+        Set<String> allowed = MemberEnvAllowList.derive(List.of());
+        Path zdotdir = EnvAllowListScrub.generate(tmp, allowed);
+
+        Map<String, String> paneEnv = new HashMap<>(Map.of(
+                "HOME", System.getProperty("user.home"),
+                "PATH", "/usr/bin:/bin",
+                "SHELL", "/bin/zsh",
+                "USER", System.getProperty("user.name", "nobody"),
+                "TMPDIR", tmp.toString()));
+        paneEnv.put("FLEETD_TEST_DECOY_SECRET", "x"); // not allow-listed; the pane must blank it
+        paneEnv.put("ZDOTDIR", zdotdir.toAbsolutePath().toString());
+
+        // The pane's own script reports what IT sees, then forks a plain non-login, non-interactive
+        // child — the shape a member's own tooling uses — carrying a variable the "parent" (this
+        // pane) deliberately set for it, the way git sets GIT_DIR for a hook.
+        String outerScript = """
+                print -r -- "PANE_SENTINEL=${%1$s:+set}"
+                print -r -- "PANE_DECOY=${FLEETD_TEST_DECOY_SECRET:+present}"
+                FLEETD_TEST_TOOL_VAR=keep /bin/zsh <<'CHILD'
+                print -r -- "CHILD_LOGIN=$([[ -o login ]] && echo yes || echo no)"
+                print -r -- "CHILD_INTERACTIVE=$([[ -o interactive ]] && echo yes || echo no)"
+                print -r -- "CHILD_TOOL_VAR=${FLEETD_TEST_TOOL_VAR:+present}"
+                print -r -- "CHILD_DECOY=${FLEETD_TEST_DECOY_SECRET:+present}"
+                print -r -- "CHILD_SENTINEL=${%1$s:+set}"
+                CHILD
+                exit
+                """.formatted(SCRUB_SENTINEL_NAME);
+
+        ProcessBuilder pb = new ProcessBuilder("/bin/zsh"); // no -l, no -i: the pane's own shape
+        pb.environment().clear();
+        pb.environment().putAll(paneEnv);
+        pb.redirectError(ProcessBuilder.Redirect.DISCARD);
+        Process zsh = pb.start();
+        zsh.getOutputStream().write(outerScript.getBytes(StandardCharsets.UTF_8));
+        zsh.getOutputStream().flush();
+        zsh.getOutputStream().close();
+        String stdout = new String(zsh.getInputStream().readAllBytes(), StandardCharsets.UTF_8);
+        assertTrue(zsh.waitFor(60, java.util.concurrent.TimeUnit.SECONDS),
+                "the pane+child probe did not exit within 60s");
+        assertTrue(zsh.exitValue() == 0, "probe zsh exited non-zero: " + stdout);
+
+        Map<String, String> reported = new HashMap<>();
+        for (String line : stdout.split("\n")) {
+            int eq = line.indexOf('=');
+            if (eq > 0) {
+                reported.put(line.substring(0, eq).trim(), line.substring(eq + 1).trim());
+            }
+        }
+
+        assertEquals("set", reported.get("PANE_SENTINEL"),
+                "the pane itself is neither login nor interactive, so the .zshenv guard must have "
+                        + "run the scrub and exported the sentinel");
+        assertEquals("", reported.get("PANE_DECOY"),
+                "the pane must blank a non-allow-listed name — invariant 1");
+        assertEquals("no", reported.get("CHILD_LOGIN"), "sanity: the child must also be non-login");
+        assertEquals("no", reported.get("CHILD_INTERACTIVE"), "sanity: the child must also be non-interactive");
+        assertEquals("present", reported.get("CHILD_TOOL_VAR"),
+                "invariant 3: a variable the pane deliberately set for its child must survive — a "
+                        + "child that re-ran the scrub would have blanked it");
+        assertEquals("", reported.get("CHILD_DECOY"),
+                "a name already blanked by the pane must stay blanked in the child, never resurrected");
+        assertEquals("set", reported.get("CHILD_SENTINEL"),
+                "the child must inherit the sentinel from the pane's environment, or it would re-scrub");
+
+        EnvAllowListScrub.ScrubReport report = EnvAllowListScrub.readReport(zdotdir);
+        assertNotNull(report, "the pane's own scrub pass must leave a report behind");
+        assertTrue(report.blanked().stream().noneMatch(n -> n.startsWith("FLEETD_TEST_TOOL_VAR")),
+                "invariant 4: the receipt must still describe the PANE, not the child — a child that "
+                        + "re-ran the scrub would have rewritten this file to list its own "
+                        + "FLEETD_TEST_TOOL_VAR as blanked");
+    }
+
     /**
      * Run {@code /bin/zsh -l -i} from a clean parent and return the NAMES it has exported by prompt
      * time. With {@code zdotdir} non-null, {@code ZDOTDIR} points at a generated scrub directory, so