fleetd#326: classify primary and configReload as deferred top-level keys

ConfigRef.changedDeferredKeys only classified seven top-level FleetConfig
keys (#323 fixed the profile side). Two more keys are read only off the
startup snapshot and were missing:

- primary: Fleetd.java:506/519/520 feed PrimaryRegistry and ReplyPushLoop
  at construction; neither is rebuilt on reload.
- configReload: Fleetd.java:679-680 decide once at startup whether to
  build a ConfigWatcher at all, and with what interval; the watcher that
  would apply a later change is itself built once, so it is deferred
  (not cold — no already-open resource goes inconsistent, a running
  watcher just keeps its original settings).

health and coordinator are deliberately left unclassified: both are read
both off the startup snapshot AND live off the config supplier at a
second call site, so no single bucket is correct for either — see the
PR body for the options writeup and the coordinator.uriEnv exposure
question the issue asked to be answered.

Each fix is proven with a failing-first test in ConfigRefTest and a
revert-quote-restore mutation check (see PR body for the transcripts).

.claude/skills/hunter/SKILL.md

@@ -0,0 +1,102 @@
+---
+name: hunter
+description: Defect-hunt procedure for a fleetd worker — sweep an assigned package for real bugs and report several ranked findings without fixing anything. Load this when the lead asks you to hunt or audit a scope rather than review one diff. Do NOT load `reviewer` for this; the two want different output.
+---
+
+# Hunter worker — procedure
+
+The turn contract (one `fleet_reply`, `fleet_ask` for the lead's decisions, honest reporting,
+never merge) is in **`CLAUDE.md` → Bridge communication → Worker** and already applies.
+
+**This skill is not `reviewer`.** `reviewer` judges one diff and reports the *single* most
+important issue in about 90 words. A hunt sweeps a whole package and reports *several* findings
+in a long structured form. Loading both gives you two contradictory output contracts, and the
+usual result is a worker that writes a good report into its terminal and ends the turn without
+sending it. Load exactly one.
+
+## 0. Read this before you read code: how the report gets home
+
+Your terminal reaches nobody. The lead sees **only** the text inside your `fleet_reply` call.
+
+A long report is exactly the case where this goes wrong, so plan for it:
+
+- **Write the report into the `fleet_reply` argument itself.** Do not compose it in your terminal
+  and then summarise it into the call.
+- If the report is long, **send it anyway** — one `fleet_reply` with everything.
+- If you end the turn without replying, the bridge scrapes your pane instead. That scrape carries
+  at most the last 4000 characters, and on a hunt it usually captures the tail of the lead's own
+  brief rather than your findings. The lead then has nothing and has to ask you again.
+
+## 1. Change nothing
+
+A hunt is read-only. Do not edit a production file, do not "quickly fix" what you find, and do
+not run a formatter. You may run the build and tests to *check* a claim, and you should say so
+when you did.
+
+## 2. Read the whole scope first
+
+Read every file in the assigned package before you judge any of it. A defect that a caller
+elsewhere in the same package makes unreachable is not a defect, and you cannot know that from
+one file.
+
+Stay inside the scope. If a defect there depends on a class outside it, read that class to
+confirm — but the defect itself must live in the scope you were given.
+
+## 3. The bar — this matters more than the count
+
+**Name the path into the bad state.** Say which caller, in which state, reaches it. A defect on
+paper is not a reachable defect. If you cannot name that path, keep the finding but mark it
+`unproven` and say exactly what you could not check. Do not drop it, and do not dress it up.
+
+**Say which direction the harm goes.** Data loss, privilege escalation and silent wrong answers
+are worth reporting even when the window is narrow. A finding whose worst outcome is a worse log
+line is not worth a block.
+
+Two workers once ran the same scope: the one that applied the direction-of-harm filter found ten
+real defects, the one that did not found none. Fewer findings the lead can act on beat many the
+lead has to triage.
+
+## 4. Shapes that have produced real merged fixes here
+
+Read for these first:
+
+1. **A one-way gate.** A guard added after an incident closes only the direction that incident
+   came from. Do not only ask what closes the gate — ask **which states still open it**.
+2. **A value read once, then used later to authorise something destructive**, after something
+   else has had a chance to change it.
+3. **A failure downgraded to a value that looks like a legitimate result** — `-1`, `null`, an
+   empty list, `false` — which a caller then trusts.
+4. **A lock held for one half of a read-modify-write and not the other**, or two collections
+   updated under different locks.
+5. **A comment or javadoc stating an invariant the code no longer keeps.** Comments are
+   load-bearing in this repo; a stale one has already caused a bug.
+
+## 5. What you cannot check, and must not claim you did
+
+- `fleetd/fleetd.yaml` is gitignored and **absent from your worktree**. You cannot read it. If a
+  finding depends on live configuration, name the key and say you could not check it.
+- `.mcp.json`, `opencode.json` and `.autoenv` in your worktree are neutralised stubs, not the
+  repo's real files.
+- The `wiki/` submodule pointer is months old. Do not cite it.
+
+Reporting a fact you took from the lead's brief as something you measured yourself is a false
+report, even when the fact is correct. Say where each fact came from.
+
+## 6. The report — what goes in `fleet_reply`
+
+One block per finding, most severe first:
+
+```
+FINDING N — <one line>
+file:line
+Path in:        <which caller, in which state, reaches this>
+Direction:      <data loss | escalation | silent wrong answer | outage | ...>
+Window/trigger: <when it actually happens>
+Confidence:     <confirmed by reading | unproven — say what you could not check>
+Why nothing else catches it: <the guard or test you checked, and why it misses>
+```
+
+End with one line naming every file you read, so the lead knows the denominator.
+
+**Nothing clears the bar?** Reply `NO FINDINGS`, name the files you read, and say what you ruled
+out. A clean sweep is a valid result; an invented defect is worse than none.

.claude/skills/reviewer/SKILL.md

@@ -10,6 +10,11 @@ never merge) is in **`CLAUDE.md` → Bridge communication → Worker** and alrea
 skill is only the *review procedure*: how to work the scope, and the exact shape of what you
 send back.
 
+**Wrong skill for a sweep.** This one reviews *one* diff or scope and reports the *single* most
+important issue. If the lead asked you to hunt or audit a whole package for several defects, load
+`hunter` instead and ignore this file — the two want different output, and following both is how a
+worker ends its turn with a good report that never gets sent.
+
 ## 1. Read the whole scope before you judge
 
 The delegation names your scope — a file, a diff, a PR, a function. **Read all of it first.**

CLAUDE.md

@@ -200,8 +200,13 @@ must obey belongs in the charter, not here.
   adapter, with a message naming the credential and the remaining seconds ("cooling off after
   repeated backend errors") — distinct wording from a quarantine refusal, so don't conflate the
   two when reading a spawn failure.
-- **Skills available to delegate:** `implementer` (worktree → commit → push → own PR) and
-  `reviewer` (scoped review → one structured finding). Name one in every delegation.
+- **Skills available to delegate:** `implementer` (worktree → commit → push → own PR),
+  `reviewer` (one diff → one structured finding) and `hunter` (sweep a package → several ranked
+  findings, change nothing). Name exactly one in every delegation. **`reviewer` and `hunter` are
+  not interchangeable** — `reviewer` caps the answer at one finding in about 90 words, so naming
+  it for a multi-finding sweep hands the worker two contradictory output contracts. That has
+  already cost three workers' turns: each wrote a good report to its terminal and ended the turn
+  with no `fleet_reply`, and the scrape returned the tail of the brief instead.
 - **Primary-side skills** (not delegation playbooks — a worker cannot use them):
   `port-to-opencode` (make an OpenCode session a participant in this workspace) and
   `fleets-status` (report every fleet that shares one LavinMQ instance).

fleetd/src/main/java/dev/ltms/fleet/auth/CallerResolver.java

@@ -236,7 +236,15 @@ public final class CallerResolver {
         // loopback-trust: same-host callers that are not workers are the primary. A non-loopback
         // caller is anonymous even here — and startup refuses that combination anyway
         // (FleetConfig.validateAuthExposure), so this is defence in depth, not the control.
-        return isLoopback(remoteAddr) ? Principal.primary(c.pid()) : Principal.anonymous();
+        //
+        // fleetd #317: "not a worker" must not be conflated with "identity unresolved". The real
+        // primary is a real process — its pid resolves (c.resolved()), it just owns no herdr pane.
+        // A caller whose peer-PID lookup failed (LsofPeerPidLookup's -1 sentinel — on any failure,
+        // silently including "lsof found no match") has no such pid, and PaneLocator's own javadoc
+        // already names what happens if that case is handed the primary role: a worker→primary
+        // escalation. So an unresolved caller is refused (ANONYMOUS — the same clean, already-tested
+        // "authenticated as nothing" outcome used everywhere else in this method), never promoted.
+        return isLoopback(remoteAddr) && c.resolved() ? Principal.primary(c.pid()) : Principal.anonymous();
     }
 
     private boolean presentedTokenMatches(String authorizationHeader) {

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

@@ -39,12 +39,28 @@ import java.util.function.Supplier;
  *       keeps the old value until a restart: {@code lifecycle:}, {@code leadHeartbeat:},
  *       {@code spawnReadyTimeoutMs} / {@code spawnReadyPollMs}, {@code quarantineCooldownSeconds}
  *       (CB-578 stage B — baked once into the {@code BackendQuarantine} built at startup),
- *       {@code guard:}, {@code worktreeRoot:}, adding or removing a profile (a new backend needs its own launcher,
+ *       {@code guard:}, {@code worktreeRoot:} and {@code worktreeGroup:} (both baked once into the
+ *       {@code GitWorktrees} built at {@code Fleetd.java:251} and never rebuilt — fleetd #323
+ *       instance 2 found {@code worktreeGroup} missing from this list and from
+ *       {@link #changedDeferredKeys}), {@code primary:} (fleetd #326 — {@code Fleetd.java:506, 519,
+ *       520} read {@code cfg.primary()} only off the startup snapshot to build {@code
+ *       PrimaryRegistry} and size {@code ReplyPushLoop}'s reminder cap/backoff, and neither is
+ *       rebuilt on reload; a lead whose pinned terminal changed under a running daemon stays
+ *       unresolved as primary until a restart), {@code configReload:} (fleetd #326 — {@code
+ *       Fleetd.java:679-680} read it only at startup to decide whether to build a {@code
+ *       ConfigWatcher} at all and with what interval; the watcher that would apply a later change is
+ *       itself built once, so a running watcher keeps polling on its original enabled flag and
+ *       interval regardless of what a reload changes it to, the same shape as {@code lifecycle} —
+ *       not cold, because no already-open resource goes inconsistent with the new value, the watcher
+ *       (if any) simply keeps its old settings), adding or removing a profile (a new backend needs its own launcher,
  *       which is constructed once), <em>and an existing profile's launch settings</em> —
  *       {@code model}, {@code baseUrl}, {@code argv}, {@code env}, {@code mcpUrl},
  *       {@code exhaustedPattern} (CB-578 stage A — compiled once into {@code Fleetd.main}'s
  *       pattern map at startup), {@code errorPattern} (fleetd #201 Unit 5 — compiled once into
- *       {@code Fleetd.main}'s backend-error pattern map at startup, the same way), and the rest.
+ *       {@code Fleetd.main}'s backend-error pattern map at startup, the same way),
+ *       {@code ideProjectDir} / {@code ideOpenCommand} / {@code autoCompactWindow} (fleetd #323
+ *       instance 1 — all three are read at spawn off the same frozen profile map and were missing
+ *       from {@link #sameLaunchSettings}), and the rest of {@link #sameLaunchSettings}.
  *       {@code credentialId} (CB-578 stage B) is NOT on
  *       this list — it is read live off the config supplier at every quarantine check and
  *       exhaustion event, exactly like {@code weight} / {@code maxLoad}, so it is hot instead.
@@ -221,6 +237,27 @@ public final class ConfigRef implements Supplier<FleetConfig> {
         if (!Objects.equals(old.worktreeRoot(), fresh.worktreeRoot())) {
             changed.add("worktreeRoot");
         }
+        // Baked into the same GitWorktrees as worktreeRoot (Fleetd.java:251) and never rebuilt
+        // either — see the class doc. Missing this check was fleetd #323 instance 2: a reload
+        // that changed only worktreeGroup reported "config reloaded" with nothing deferred, and
+        // newly provisioned worktrees kept the old sharing behaviour.
+        if (!Objects.equals(old.worktreeGroup(), fresh.worktreeGroup())) {
+            changed.add("worktreeGroup");
+        }
+        // fleetd #326: Fleetd.java:506, 519, 520 read cfg.primary() only off the startup snapshot
+        // (PrimaryRegistry's pinned terminal, ReplyPushLoop's reminder cap and backoff) — neither is
+        // rebuilt on reload, so a changed pin needs a restart before a lead resolves as primary again.
+        if (!Objects.equals(old.primary(), fresh.primary())) {
+            changed.add("primary");
+        }
+        // fleetd #326: Fleetd.java:679-680 read cfg.configReload() only at startup to decide whether
+        // to build a ConfigWatcher at all and with what interval — the watcher that would apply a
+        // later change is itself built once, so a running watcher keeps its original enabled flag and
+        // interval regardless of what a reload changes it to. Not cold: no already-open resource goes
+        // inconsistent with the new value, a watcher (if any) simply keeps polling on the old settings.
+        if (!Objects.equals(old.configReload(), fresh.configReload())) {
+            changed.add("configReload");
+        }
         if (!Objects.equals(old.spawnReadyTimeoutMs(), fresh.spawnReadyTimeoutMs())
                 || !Objects.equals(old.spawnReadyPollMs(), fresh.spawnReadyPollMs())) {
             changed.add("spawnReady*");
@@ -265,14 +302,34 @@ public final class ConfigRef implements Supplier<FleetConfig> {
     }
 
     /**
-     * Whether two versions of a profile would launch a peer identically. Compares every component
-     * the launcher reads at spawn; {@code weight}, {@code maxLoad} and {@code credentialId} are
-     * excluded because those are read live (by the placement policy and, for credentialId, by
-     * {@code CompositePeerLauncher}/the CB-578 stage B exhaustion sink) and really do take effect on
-     * the next spawn.
+     * {@link FleetConfig.Profile} record components deliberately left out of
+     * {@link #sameLaunchSettings} because they are read <em>live</em>, not baked in at spawn — see
+     * the class doc's <em>Hot</em> bullet. {@code weight} and {@code maxLoad} are read live by the
+     * placement policy on every spawn; {@code credentialId} is read live by
+     * {@code CompositePeerLauncher} and the CB-578 stage B exhaustion sink. Nothing else is
+     * excluded — see {@code sameLaunchSettingsComparesEveryProfileComponentOrExcludesIt} in
+     * {@code ConfigRefProfileCoverageTest}, which enumerates every {@code Profile} record component
+     * by reflection and fails the build if one is neither compared below nor named here.
      */
-    private static boolean sameLaunchSettings(FleetConfig.Profile a, FleetConfig.Profile b) {
-        return Objects.equals(a.baseUrl(), b.baseUrl())
+    static final Set<String> LAUNCH_SETTINGS_EXCLUDED = Set.of("weight", "maxLoad", "credentialId");
+
+    /**
+     * Whether two versions of a profile would launch a peer identically.
+     *
+     * <p>This must compare every {@link FleetConfig.Profile} record component except the three in
+     * {@link #LAUNCH_SETTINGS_EXCLUDED}. That is not a claim this javadoc can make good on by
+     * itself — a javadoc saying "compares every component" is exactly what fleetd #323 found to be
+     * false for three fields (and a sibling method's field list, for a fourth). The actual
+     * guarantee comes from {@code ConfigRefProfileCoverageTest}: it enumerates every record
+     * component of {@code FleetConfig.Profile} by reflection, mutates each one not in
+     * {@code LAUNCH_SETTINGS_EXCLUDED} on a base profile, and asserts this method reports a
+     * difference — so a new component that is neither compared here nor added to
+     * {@code LAUNCH_SETTINGS_EXCLUDED} (with a reason) fails that test by name, rather than
+     * silently reporting "config reloaded" for a value the daemon never picked up.
+     */
+    static boolean sameLaunchSettings(FleetConfig.Profile a, FleetConfig.Profile b) {
+        return Objects.equals(a.profile(), b.profile())
+                && Objects.equals(a.baseUrl(), b.baseUrl())
                 && Objects.equals(a.model(), b.model())
                 && Objects.equals(a.configDir(), b.configDir())
                 && Objects.equals(a.tokenEnv(), b.tokenEnv())
@@ -298,6 +355,16 @@ public final class ConfigRef implements Supplier<FleetConfig> {
                 // fleetd #201 Unit 5: errorPattern is compiled once into Fleetd.main's backend-error
                 // pattern map at startup (see BackendErrorPatternLookup wiring), the same way
                 // exhaustedPattern is — a reload never re-reads it either.
-                && Objects.equals(a.errorPattern(), b.errorPattern());
+                && Objects.equals(a.errorPattern(), b.errorPattern())
+                // fleetd #323 instance 1: ideProjectDir and ideOpenCommand are read at spawn off the
+                // same frozen profile map as ideMcpUrl above (ClaudeCodeLauncher.java:267/269,
+                // OpenCodeLauncher.java:474/480/486) and were missing from this comparison.
+                && Objects.equals(a.ideProjectDir(), b.ideProjectDir())
+                && Objects.equals(a.ideOpenCommand(), b.ideOpenCommand())
+                // fleetd #323 instance 1: autoCompactWindow is read at spawn the same way
+                // (ClaudeCodeLauncher.java:926, OpenCodeLauncher.java:650). Comparing it here only
+                // makes the reload REPORT that a restart is needed — it deliberately does not make
+                // autoCompactWindow take effect live, which is a separate, larger change.
+                && Objects.equals(a.autoCompactWindow(), b.autoCompactWindow());
     }
 }

fleetd/src/main/java/dev/ltms/fleet/mcp/ConnectionIdentity.java

@@ -36,6 +36,25 @@ public final class ConnectionIdentity {
      * primary / an off-host client) and its {@code pid} (or {@code -1} if not resolvable).
      */
     public record Caller(String terminal, long pid) {
+
+        /**
+         * Whether the OS peer-PID lookup actually succeeded — {@code false} means {@code pid} is
+         * the {@code -1} sentinel, not a real process id, so this caller's identity could not be
+         * established at all. That is a different fact from a real pid that simply owns no worker
+         * pane (the primary's own connection): the primary is {@code resolved()} and has a
+         * {@code null terminal}; an unresolvable caller is {@code !resolved()} and also has a
+         * {@code null terminal}. The two look identical through {@link #terminal} alone, which is
+         * exactly how fleetd #317 happened — a failed {@code lsof} lookup and a genuine primary both
+         * fell through to {@code Principal.primary(...)}.
+         *
+         * <p>Centralised here, next to the sentinel it tests, for the same reason
+         * {@link ConnectionIdentity#isLoopback} is centralised rather than left for each caller to
+         * reimplement: a raw {@code pid > 0} check duplicated at every call site is precisely the
+         * "one rule, two copies" shape that let #305 drift.
+         */
+        public boolean resolved() {
+            return pid > 0;
+        }
     }
 
     /** Resolve the caller's terminal and PID from one peer-PID lookup. */
@@ -71,13 +90,27 @@ public final class ConnectionIdentity {
      * {@code 127.0.0.1:8765} with a source address of {@code 127.0.0.2} — measured on the Linux
      * fleet host, where binding that source succeeds.
      *
-     * <p><strong>Being strict here does not make the daemon safer; it makes it unsafe.</strong>
-     * That reads backwards, so it is worth stating plainly. This predicate does not decide whether
-     * a caller is trusted — it decides whether the caller's identity is <em>resolved at all</em>.
-     * Returning false means {@link #resolve} answers "no terminal", and downstream a caller with no
-     * terminal is treated as the primary under loopback-trust. So every address excluded here is an
-     * address on which a worker silently becomes the lead. Widening a check normally weakens it;
-     * widening this one is what closes the hole.
+     * <p><strong>What excluding an address costs, stated as it is today.</strong> This paragraph
+     * used to say that narrowing this range turned a worker into the lead, and that widening the
+     * check was what closed the hole. That was true only while there were <em>two</em> definitions
+     * that disagreed: {@code ConnectionIdentity} skipped the identity lookup for {@code 127.0.0.2}
+     * while {@code CallerResolver} read the same address as loopback and granted the primary role.
+     * #305 removed the second copy, and with one shared definition the old sentence no longer holds.
+     *
+     * <p>Measured on 2026-09-04 by narrowing this method back to exactly {@code 127.0.0.1} and
+     * running {@code CallerResolverTest} and {@code ConnectionIdentityTest}: a caller from
+     * {@code 127.0.0.2} then resolves to {@code ANONYMOUS}, not {@code PRIMARY} — for a worker
+     * ({@code aWorkerOnAnyLoopbackSourceAddressIsStillAWorkerNotThePrimary}) and for a non-worker
+     * ({@code aNonWorkerOnAnyLoopbackSourceAddressIsStillThePrimary}) alike. Excluding an address
+     * now <em>refuses</em> its caller; it does not promote one.
+     *
+     * <p>So keep the whole range, but for the plain reason: a genuine worker or primary that
+     * connects from {@code 127.0.0.2} must be identifiable at all, and narrowing this predicate
+     * locks it out. That is an outage, and an outage is the direction to fail in — which is exactly
+     * why the range must not be narrowed casually and also why doing so is no longer a security
+     * hole. This predicate still does not decide whether a caller is trusted; it decides whether the
+     * caller's identity is <em>resolved at all</em>. What makes an unresolved caller safe is
+     * {@link Caller#resolved()} (#317), not this method.
      */
     public static boolean isLoopback(String addr) {
         if (addr == null) {

fleetd/src/main/java/dev/ltms/fleet/mcp/LsofPeerPidLookup.java

@@ -41,6 +41,14 @@ public final class LsofPeerPidLookup implements PeerPidLookup {
             if (!p.waitFor(2, TimeUnit.SECONDS)) {
                 p.destroyForcibly();
             }
+            if (found < 0) {
+                // fleetd #317: this is the silent path — lsof ran clean and simply reported no
+                // matching process (e.g. queried before the OS socket table settles). Previously
+                // this logged nothing at all, which is exactly why the escalation went unnoticed;
+                // the exception path below already logs. A caller now refused because of this is
+                // still refused (never promoted) — this line only makes the refusal diagnosable.
+                log.debug("lsof peer-pid lookup for port {} found no matching process", port);
+            }
             return found;
         } catch (Exception e) {
             log.debug("lsof peer-pid lookup for port {} failed: {}", port, e.getMessage());

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

@@ -385,9 +385,12 @@ public final class CompositePeerLauncher implements PeerLauncher {
             }
         }
 
+        // unreachable.size() counts DISTINCT profiles, not attempts (a HashSet dedupes a profile
+        // added twice) — say "distinct" so the count matches the sentence and the profile list that
+        // follows, rather than reading as a count of attempts made (fleetd #315).
         throw new PeerUnreachableException(
                 "no reachable worker profile available after trying " + unreachable.size()
-                        + " candidate(s): " + String.join(", ", unreachable));
+                        + " distinct candidate(s): " + String.join(", ", unreachable));
     }
 
     /**

fleetd/src/main/java/dev/ltms/fleet/placement/FixedPlacementPolicy.java

@@ -5,10 +5,14 @@ import java.util.List;
 
 /**
  * Backward-compatible placement: an unqualified spawn always resolves to the configured default
- * profile, exactly as {@code CompositePeerLauncher} did before CB-518. This ignores caps and
- * reachability so that a pre-existing config behaves identically after upgrade.
+ * profile, exactly as {@code CompositePeerLauncher} did before CB-518. This ignores caps
+ * ({@code maxLoad}) so that a pre-existing config behaves identically after upgrade — capacity
+ * gating for automatic placement is deliberately out of scope for {@code fixed}, exactly as it
+ * always has been. Reachability is a narrower exception (fleetd #315, below): a profile is never
+ * checked for reachability up front, only skipped once it has already failed in <em>this same</em>
+ * spawn call's retry loop — see the unreachable case below.
  *
- * <p>Three exceptions walk past the default instead of returning it unconditionally:
+ * <p>Four exceptions walk past the default instead of returning it unconditionally:
  * <ul>
  *   <li>Quarantine (CB-578 stage B): a quarantined default is a credential that just refused on
  *       a usage limit, not a transient capacity or reachability concern.
@@ -16,13 +20,21 @@ import java.util.List;
  *       ({@code BackendOutagePolicy}) — a separate, shorter-lived source from quarantine. When a
  *       profile is both quarantined and cooling off, only the quarantine reason is reported
  *       (exhaustion takes priority), matching {@code CompositePeerLauncher}'s explicit-spawn order.
+ *   <li>Unreachable (fleetd #315): {@code CompositePeerLauncher.spawn} retries a failed candidate
+ *       on the next one and rebuilds the {@link PlacementContext} so {@code ctx.unreachable()}
+ *       names every profile that already failed with {@code PeerUnreachableException} in this same
+ *       call. Without this check {@code select} kept handing back the same dead default forever —
+ *       the retry loop's own comment says "so the policy excludes this profile", and this is what
+ *       makes that true for {@code fixed} too, matching {@code weighted}/{@code round-robin}
+ *       (both filter on {@code ctx.unreachable()} via {@link PlacementPolicyUtil#available}).
  *   <li>Weight 0 (CB-554): {@code fixed} is still automatic selection, so a profile the operator
  *       marked "never auto-select me" ({@code weight <= 0}) must be skipped here exactly as
  *       {@code weighted}/{@code round-robin} skip it — an explicit {@code fleet_spawn} naming
  *       the profile is unaffected, only this automatic fallback walk.
  * </ul>
- * A fleet where nothing is ever quarantined, cooling off, or weight-0 never exercises any of these
- * paths, so today's behaviour is unchanged.
+ * A fleet where nothing is ever quarantined, cooling off, unreachable, or weight-0 never exercises
+ * any of these paths, so today's behaviour is unchanged — in particular, the very first selection
+ * of a spawn call always sees an empty {@code unreachable} set, so the first choice is untouched.
  */
 final class FixedPlacementPolicy implements PlacementPolicy {
 
@@ -30,12 +42,12 @@ final class FixedPlacementPolicy implements PlacementPolicy {
     public PlacementCandidate select(PlacementContext ctx) {
         String d = ctx.defaultProfile();
         if (d != null && !d.isBlank() && !ctx.quarantined().contains(d) && !ctx.coolingOff().contains(d)
-                && !weightExcluded(ctx, d)) {
+                && !ctx.unreachable().contains(d) && !weightExcluded(ctx, d)) {
             return new PlacementCandidate(d, null, 1.0f, null);
         }
         for (PlacementCandidate c : ctx.candidates()) {
             if (!ctx.quarantined().contains(c.profile()) && !ctx.coolingOff().contains(c.profile())
-                    && !c.excluded()) {
+                    && !ctx.unreachable().contains(c.profile()) && !c.excluded()) {
                 return new PlacementCandidate(c.profile(), null, c.weight(), c.maxLoad());
             }
         }
@@ -44,8 +56,9 @@ final class FixedPlacementPolicy implements PlacementPolicy {
             // Exhaustion quarantine takes priority: reported only when quarantine is absent, so the
             // message never claims "cooling off" for a profile that is really backend-exhausted.
             boolean dCoolingOff = !dQuarantined && ctx.coolingOff().contains(d);
+            boolean dUnreachable = ctx.unreachable().contains(d);
             boolean dWeightExcluded = weightExcluded(ctx, d);
-            if (dQuarantined || dCoolingOff || dWeightExcluded) {
+            if (dQuarantined || dCoolingOff || dUnreachable || dWeightExcluded) {
                 List<String> reasons = new ArrayList<>();
                 if (dQuarantined) {
                     reasons.add("is quarantined (backend exhausted)");
@@ -53,6 +66,9 @@ final class FixedPlacementPolicy implements PlacementPolicy {
                 if (dCoolingOff) {
                     reasons.add("is cooling off after repeated backend errors");
                 }
+                if (dUnreachable) {
+                    reasons.add("is unreachable");
+                }
                 if (dWeightExcluded) {
                     reasons.add("has weight 0 (excluded from automatic selection)");
                 }
@@ -62,7 +78,7 @@ final class FixedPlacementPolicy implements PlacementPolicy {
         }
         if (!ctx.candidates().isEmpty()) {
             throw new PlacementException("all worker profiles are excluded from automatic "
-                    + "selection (quarantined, cooling off, or weight-0)");
+                    + "selection (quarantined, cooling off, unreachable, or weight-0)");
         }
         throw new PlacementException("no worker profiles configured");
     }

fleetd/src/main/java/dev/ltms/fleet/session/SessionManager.java

@@ -386,6 +386,31 @@ public final class SessionManager implements TurnListener {
         // from the registry with no pane stop is an orphaned pane — a live terminal burning a fleet
         // slot that no longer appears in the roster and can never be reclaimed.
         launcher.stop(paneId);
+        if (removed != null && !preserveWorktree && removed.worktree() != null) {
+            // fleetd #316: the `dirty` read above ran while the worker could still write to this
+            // worktree, so a stale `false` must not be trusted to authorise the --force removal
+            // below. Re-read the worktree's state one more time, right here — immediately before
+            // the one step that would destroy it, and only on the path that is actually about to
+            // do that (invariant 4: no second unconditional `git status` on a release that already
+            // decided to preserve). By now `launcher.stop` has returned, so this read reflects
+            // whatever the worker managed to write up to and including its teardown, not whatever
+            // it had written at release-start time.
+            if (dirtyImmediatelyBeforeRemoval(removed)) {
+                preserveWorktree = true;
+                // The pre-stop snapshot above never ran for this session (the pre-stop read said
+                // clean), so this is the only chance to get the newly-discovered work into
+                // refs/wip/* rather than leaving the on-disk preserve as the sole copy. Best-effort,
+                // like every other snapshot attempt — trySnapshot logs and swallows its own failure.
+                String lateSnapshotRef = trySnapshot(removed, cause);
+                log.warn("release {} preserves worktree {} for pane={} terminal={}: it reported "
+                        + "clean before the pane stopped but dirty immediately before removal — the "
+                        + "worker wrote to it during teardown, and --force removing it now would "
+                        + "have destroyed that work{}",
+                        cause, removed.worktree(), paneId, removed.terminalId(),
+                        lateSnapshotRef == null ? "" : " (snapshotted to refs/wip/" + removed.branch()
+                                + " commit=" + lateSnapshotRef + ")");
+            }
+        }
         if (removed != null && !preserveWorktree && removed.worktree() != null) {
             // fleetd #283: this is the one cleanup step in this method that used to be bare. By the
             // time it runs, the registry entry, the retained handle, and the pane are all already
@@ -404,6 +429,24 @@ public final class SessionManager implements TurnListener {
         }
     }
 
+    /**
+     * fleetd #316: the read that actually authorises {@code worktrees.remove}, taken with the
+     * worker's pane already stopped. Fails toward preserving (returns {@code true}) on any
+     * exception — the same rule the pre-stop check applies (CB-581): once we can no longer tell
+     * whether the worktree is dirty, preserving costs disk while deleting on a guess can destroy
+     * work that has no other copy.
+     */
+    private boolean dirtyImmediatelyBeforeRemoval(MemberSession removed) {
+        try {
+            return worktrees.hasUncommitted(removed.worktree());
+        } catch (RuntimeException e) {
+            log.warn("release could not re-check worktree {} for pane={} terminal={} immediately "
+                    + "before removal; preserving it rather than risk destroying unsaved work: {}",
+                    removed.worktree(), removed.paneId(), removed.terminalId(), e.toString());
+            return true;
+        }
+    }
+
     /**
      * Best-effort snapshot of a dirty worktree into {@code refs/wip/<branch>} (CB-578 stage C). A
      * failure here must never escalate: the caller has already decided to preserve the worktree

fleetd/src/test/java/dev/ltms/fleet/auth/CallerResolverTest.java

@@ -103,6 +103,54 @@ class CallerResolverTest {
         assertEquals(Role.PRIMARY, p.role(), "the historical behaviour, now an explicit choice");
     }
 
+    // ── fleetd #317: an unresolvable caller must never be promoted to the primary ──────────────────
+    // #305 closed the trigger where a resolved pid matched no pane *and* had no ancestry walk to
+    // save it. This is the other trigger PaneLocator's javadoc names: the pid never resolves at
+    // all — LsofPeerPidLookup returns -1 on any failure, including (silently) "lsof found no
+    // match" — so there is no candidate pid for an ancestry walk to even attempt.
+
+    /**
+     * The failing-without-the-fix case. Before #317's fix, {@code c.terminal() == null} was the
+     * only test in the loopback-trust fallback, and an unresolved pid produces exactly that same
+     * {@code null} terminal as a genuine primary — so this caller was handed
+     * {@code Principal.primary(...)}, a real worker's failed lookup becoming indistinguishable from
+     * the lead.
+     */
+    @Test
+    void aFailedPeerPidLookupIsRefusedNotPromotedToPrimary() {
+        ConnectionIdentity unresolved = new ConnectionIdentity(new PaneLocator(herdr), _ -> -1);
+        Principal p = new CallerResolver(unresolved).resolve("127.0.0.1", 55555, null);
+
+        assertEquals(Role.ANONYMOUS, p.role(),
+                "an unresolvable caller must never be silently promoted to the primary");
+    }
+
+    /**
+     * The companion invariant #317 must not break: a caller whose lookup genuinely succeeded, and
+     * who simply owns no herdr pane — the real primary's own connection — is still the primary.
+     * This is {@link #loopbackTrustTreatsANonWorkerLoopbackCallerAsThePrimary} pinned again here,
+     * named for #317 and placed next to the test it must be distinguished from: same {@code null}
+     * terminal, opposite verdict, because {@code Caller.resolved()} tells them apart.
+     */
+    @Test
+    void aRealPidThatOwnsNoPaneIsStillThePrimaryNotRefused() {
+        Principal p = new CallerResolver(nonWorkerIdentity()).resolve("127.0.0.1", 55555, null);
+
+        assertEquals(Role.PRIMARY, p.role());
+    }
+
+    /** #317 point 4: token mode never consults {@code c.pid()}, so a failed lookup must not change it. */
+    @Test
+    void tokenModeIsUndisturbedByAnUnresolvedLookup() {
+        ConnectionIdentity unresolved = new ConnectionIdentity(new PaneLocator(herdr), _ -> -1);
+        CallerResolver r = new CallerResolver(unresolved, true, "s3cret");
+
+        assertEquals(Role.ANONYMOUS, r.resolve("127.0.0.1", 55555, null).role(),
+                "no credential is still just ANONYMOUS, as before #317 — unchanged by the lookup failing");
+        assertEquals(Role.PRIMARY, r.resolve("127.0.0.1", 55555, "Bearer s3cret").role(),
+                "a valid token still authenticates the primary even though the peer-pid lookup failed");
+    }
+
     @Test
     void tokenModeRefusesANonWorkerCallerThatPresentsNoToken() {
         Principal p = new CallerResolver(nonWorkerIdentity(), true, "s3cret")

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

@@ -0,0 +1,202 @@
+package dev.ltms.fleet.config;
+
+import org.junit.jupiter.api.Test;
+
+import java.lang.reflect.Constructor;
+import java.lang.reflect.RecordComponent;
+import java.util.Arrays;
+import java.util.LinkedHashMap;
+import java.util.List;
+import java.util.Map;
+import java.util.Set;
+import java.util.TreeSet;
+
+import static org.junit.jupiter.api.Assertions.assertEquals;
+import static org.junit.jupiter.api.Assertions.assertTrue;
+
+/**
+ * fleetd #323: {@code ConfigRef.sameLaunchSettings} javadoc used to claim it "compares every
+ * component the launcher reads at spawn". It did not — {@code ideProjectDir}, {@code
+ * ideOpenCommand} and {@code autoCompactWindow} were all baked in at daemon startup (see
+ * {@code ClaudeCodeLauncher}/{@code OpenCodeLauncher}) and missing from the comparison, so a reload
+ * that changed only one of them reported "config reloaded" and the running daemon kept the old
+ * value.
+ *
+ * <p>This class is the mechanism the issue asked for: it enumerates every record component of
+ * {@link FleetConfig.Profile} by reflection and proves — by actually mutating a base profile one
+ * field at a time and calling the real method — that each component is either compared by
+ * {@link ConfigRef#sameLaunchSettings} or named in {@link ConfigRef#LAUNCH_SETTINGS_EXCLUDED} with
+ * a reason. A new profile field that is neither fails this test by name, not a hand-maintained list
+ * going stale.
+ */
+class ConfigRefProfileCoverageTest {
+
+    private static final RecordComponent[] COMPONENTS = FleetConfig.Profile.class.getRecordComponents();
+
+    /**
+     * One valid, non-blank value per record component — "the a value". None of these trip any
+     * defaulting/normalization in {@code Profile}'s compact constructor (see {@code
+     * FleetConfig.java}), so what goes in is what {@code sameLaunchSettings} sees back out.
+     */
+    private static final Map<String, Object> BASE = baseValues();
+
+    /** The same shape, each value distinct from {@link #BASE} — "the b value". */
+    private static final Map<String, Object> ALT = altValues();
+
+    private static Map<String, Object> baseValues() {
+        Map<String, Object> v = new LinkedHashMap<>();
+        v.put("profile", "sonnet");
+        v.put("baseUrl", "http://gx00.gw:8000");
+        v.put("model", "sonnet");
+        v.put("configDir", "/config/a");
+        v.put("tokenEnv", "TOKEN_A");
+        v.put("argv", List.of("claude", "--flag-a"));
+        v.put("placement", "tab");
+        v.put("workspace", "workspace-a");
+        v.put("tabLabel", "label-a");
+        v.put("mcpUrl", "http://mcp-a");
+        v.put("cwd", "/cwd/a");
+        v.put("parityOverlay", List.of(".env", ".env.a"));
+        v.put("gitTokenEnv", "GIT_TOKEN_A");
+        v.put("gitHostEnv", "GITEA_HOST_A");
+        v.put("kind", "claude-code");
+        v.put("env", Map.of("K", "A"));
+        v.put("weight", 1.0f);
+        v.put("maxLoad", 5);
+        v.put("subscription", Boolean.TRUE);
+        v.put("exhaustedPattern", "usage limit a");
+        v.put("credentialId", "cred-a");
+        v.put("ideMcpUrl", "http://ide-mcp-a");
+        v.put("ideProjectDir", "modules/a");
+        v.put("ideOpenCommand", "open-cmd-a {dir}");
+        v.put("autoCompactWindow", 150000);
+        v.put("errorPattern", "error a");
+        assertNamesMatchComponents(v);
+        return v;
+    }
+
+    private static Map<String, Object> altValues() {
+        Map<String, Object> v = new LinkedHashMap<>();
+        v.put("profile", "sonnet-b");
+        v.put("baseUrl", "http://gx01.gw:8000");
+        v.put("model", "haiku");
+        v.put("configDir", "/config/b");
+        v.put("tokenEnv", "TOKEN_B");
+        v.put("argv", List.of("claude", "--flag-b"));
+        v.put("placement", "weighted");
+        v.put("workspace", "workspace-b");
+        v.put("tabLabel", "label-b");
+        v.put("mcpUrl", "http://mcp-b");
+        v.put("cwd", "/cwd/b");
+        v.put("parityOverlay", List.of(".env", ".env.b"));
+        v.put("gitTokenEnv", "GIT_TOKEN_B");
+        v.put("gitHostEnv", "GITEA_HOST_B");
+        v.put("kind", "opencode");
+        v.put("env", Map.of("K", "B"));
+        v.put("weight", 2.0f);
+        v.put("maxLoad", 9);
+        v.put("subscription", Boolean.FALSE);
+        v.put("exhaustedPattern", "usage limit b");
+        v.put("credentialId", "cred-b");
+        v.put("ideMcpUrl", "http://ide-mcp-b");
+        v.put("ideProjectDir", "modules/b");
+        v.put("ideOpenCommand", "open-cmd-b {dir}");
+        v.put("autoCompactWindow", 250000);
+        v.put("errorPattern", "error b");
+        assertNamesMatchComponents(v);
+        return v;
+    }
+
+    private static void assertNamesMatchComponents(Map<String, Object> values) {
+        Set<String> componentNames = new TreeSet<>();
+        for (RecordComponent rc : COMPONENTS) {
+            componentNames.add(rc.getName());
+        }
+        assertEquals(componentNames, new TreeSet<>(values.keySet()),
+                "this test's value map has drifted from FleetConfig.Profile's actual components — "
+                        + "update BASE/ALT alongside the record");
+    }
+
+    private static FleetConfig.Profile profileOf(Map<String, Object> values) throws ReflectiveOperationException {
+        Class<?>[] types = Arrays.stream(COMPONENTS).map(RecordComponent::getType).toArray(Class<?>[]::new);
+        Object[] args = Arrays.stream(COMPONENTS)
+                .map(rc -> values.get(rc.getName()))
+                .toArray();
+        Constructor<FleetConfig.Profile> ctor = FleetConfig.Profile.class.getDeclaredConstructor(types);
+        return ctor.newInstance(args);
+    }
+
+    /** {@code BASE} with exactly one named component swapped for its {@code ALT} value. */
+    private static FleetConfig.Profile mutate(String componentName) throws ReflectiveOperationException {
+        Map<String, Object> values = new LinkedHashMap<>(BASE);
+        values.put(componentName, ALT.get(componentName));
+        return profileOf(values);
+    }
+
+    /**
+     * The mechanism fleetd #323 asked for: enumerate {@link FleetConfig.Profile}'s record
+     * components, mutate each non-excluded one, and prove {@code sameLaunchSettings} actually
+     * notices — not just that some hand-maintained list claims it does. Prints the denominator
+     * (total / compared / excluded) the issue required: a checker that cannot state its own
+     * denominator is the failure this repo keeps hitting.
+     */
+    @Test
+    void sameLaunchSettingsComparesEveryProfileComponentOrExcludesIt() throws ReflectiveOperationException {
+        int total = COMPONENTS.length;
+        Set<String> excluded = ConfigRef.LAUNCH_SETTINGS_EXCLUDED;
+
+        Set<String> allNames = new TreeSet<>();
+        for (RecordComponent rc : COMPONENTS) {
+            allNames.add(rc.getName());
+        }
+        assertTrue(allNames.containsAll(excluded),
+                "ConfigRef.LAUNCH_SETTINGS_EXCLUDED names a component that does not exist on "
+                        + "FleetConfig.Profile — check for a typo: " + excluded);
+
+        // The exclusion set is this mechanism's own escape hatch, so it has to be pinned too.
+        // Found by mutation while verifying fleetd #323: moving autoCompactWindow and
+        // ideOpenCommand OUT of the comparison and INTO the exclusion set left the whole suite
+        // green — the loop below simply skips them, and the denominator assertion still balances.
+        // That is exactly the lazy move a failing coverage test invites, and it silently restores
+        // the #323 bug. Only ideProjectDir and worktreeGroup were saved by a behavioural test in
+        // ConfigRefTest; the other two had none. So: growing this set now requires editing this
+        // line as well, which is a visible, deliberate diff rather than a quiet one.
+        assertEquals(Set.of("weight", "maxLoad", "credentialId"), excluded,
+                "ConfigRef.LAUNCH_SETTINGS_EXCLUDED changed. A component belongs in it ONLY if it "
+                        + "is read live off the config supplier, not baked into a launcher at "
+                        + "startup. If you are adding one to silence this test, that is fleetd #323 "
+                        + "happening again: compare it in sameLaunchSettings instead. If it really "
+                        + "is read live, name where it is read and update this assertion.");
+
+        FleetConfig.Profile base = profileOf(BASE);
+        List<String> uncovered = new java.util.ArrayList<>();
+        int compared = 0;
+        for (RecordComponent rc : COMPONENTS) {
+            String name = rc.getName();
+            if (excluded.contains(name)) {
+                continue;
+            }
+            FleetConfig.Profile mutated = mutate(name);
+            if (ConfigRef.sameLaunchSettings(base, mutated)) {
+                uncovered.add(name);
+            } else {
+                compared++;
+            }
+        }
+
+        System.out.printf(
+                "ConfigRef.sameLaunchSettings coverage — %d Profile components total, %d compared, "
+                        + "%d excluded (%s)%n",
+                total, compared, excluded.size(), excluded);
+
+        assertEquals(List.of(), uncovered,
+                "these FleetConfig.Profile components changed but ConfigRef.sameLaunchSettings "
+                        + "reported no difference — add each one to the comparison (it is read at "
+                        + "spawn and baked in until a restart) or to ConfigRef.LAUNCH_SETTINGS_EXCLUDED "
+                        + "with a reason it is genuinely read live: " + uncovered);
+        assertEquals(total, compared + excluded.size(),
+                "every FleetConfig.Profile record component must be either compared or excluded — "
+                        + total + " components, " + compared + " compared, " + excluded.size()
+                        + " excluded");
+    }
+}

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

@@ -434,6 +434,142 @@ class ConfigRefTest {
         assertEquals("provider 5xx", ref.get().profiles().get("sonnet").errorPattern());
     }
 
+    /**
+     * fleetd #323 instance 1: {@code ideProjectDir} is read at spawn off the frozen profile map
+     * (see {@code ClaudeCodeLauncher}/{@code OpenCodeLauncher}) exactly like {@code model}, but was
+     * missing from {@code sameLaunchSettings} — a reload changing only this field used to report a
+     * bare "config reloaded" and the running daemon kept launching with the old value.
+     */
+    @Test
+    void changingAProfilesIdeProjectDirIsReportedAsDeferred(@TempDir Path dir) throws Exception {
+        Path f = dir.resolve("fleetd.yaml");
+        Files.writeString(f, """
+                bind:
+                  host: 127.0.0.1
+                  port: 8765
+                herdrSocket: ~/.config/herdr/herdr.sock
+                profiles:
+                  sonnet:
+                    baseUrl: http://gx00.gw:8000
+                    model: sonnet
+                    ideProjectDir: fleetd
+                guard:
+                  offSubscriptionHosts:
+                    - gx00.gw
+                """);
+        ConfigRef ref = refFor(f);
+
+        Files.writeString(f, """
+                bind:
+                  host: 127.0.0.1
+                  port: 8765
+                herdrSocket: ~/.config/herdr/herdr.sock
+                profiles:
+                  sonnet:
+                    baseUrl: http://gx00.gw:8000
+                    model: sonnet
+                    ideProjectDir: fleetd-renamed
+                guard:
+                  offSubscriptionHosts:
+                    - gx00.gw
+                """);
+        ConfigRef.Outcome out = ref.reload();
+
+        assertTrue(out.applied());
+        assertEquals(1, out.deferred().size(), out.deferred().toString());
+        assertTrue(out.deferred().getFirst().contains("sonnet"), out.deferred().toString());
+        assertTrue(out.deferred().getFirst().contains("launch settings"), out.deferred().toString());
+        // The snapshot still carries the new value — a restart is what makes it take effect.
+        assertEquals("fleetd-renamed", ref.get().profiles().get("sonnet").ideProjectDir());
+    }
+
+    /**
+     * fleetd #323 instance 2: {@code worktreeGroup} is baked into the same {@code GitWorktrees}
+     * as {@code worktreeRoot} (Fleetd.java:251) and never rebuilt, but only {@code worktreeRoot}
+     * was on {@code changedDeferredKeys} — a reload changing only the group reported a bare
+     * "config reloaded" and newly provisioned worktrees kept the old sharing behaviour.
+     */
+    @Test
+    void changingWorktreeGroupIsReportedAsDeferred(@TempDir Path dir) throws Exception {
+        Path f = dir.resolve("fleetd.yaml");
+        Files.writeString(f, yaml("worktreeGroup: devgroup\n"));
+        ConfigRef ref = refFor(f);
+
+        Files.writeString(f, yaml("worktreeGroup: devgroup2\n"));
+        ConfigRef.Outcome out = ref.reload();
+
+        assertTrue(out.applied());
+        assertEquals(java.util.List.of("worktreeGroup"), out.deferred());
+        assertTrue(out.summary().contains("needs a restart") || out.summary().contains("need a restart"),
+                out.summary());
+        // The snapshot still carries the new value — a restart is what makes it take effect.
+        assertEquals("devgroup2", ref.get().worktreeGroup());
+    }
+
+    /**
+     * fleetd #326: {@code primary} is read only off the startup snapshot — {@code Fleetd.java:506,
+     * 519, 520} feed {@code PrimaryRegistry} and {@code ReplyPushLoop} at construction and neither is
+     * rebuilt on reload — but it was missing from {@link ConfigRef#changedDeferredKeys}, so a reload
+     * that only changed the pinned primary terminal reported a bare "config reloaded" while a lead
+     * whose tab no longer matched stayed demoted to worker.
+     */
+    @Test
+    void changingPrimaryIsReportedAsDeferred(@TempDir Path dir) throws Exception {
+        Path f = dir.resolve("fleetd.yaml");
+        Files.writeString(f, yaml("""
+                primary:
+                  terminal: term-a
+                """));
+        ConfigRef ref = refFor(f);
+
+        Files.writeString(f, yaml("""
+                primary:
+                  terminal: term-b
+                """));
+        ConfigRef.Outcome out = ref.reload();
+
+        assertTrue(out.applied());
+        assertEquals(java.util.List.of("primary"), out.deferred());
+        assertTrue(out.summary().contains("need") && out.summary().contains("restart"), out.summary());
+        // The snapshot still carries the new value — a restart is what makes it take effect.
+        assertEquals("term-b", ref.get().primary().terminal());
+    }
+
+    /**
+     * fleetd #326: {@code configReload} itself is read only at startup ({@code Fleetd.java:679-680})
+     * to decide whether to build a {@code ConfigWatcher} at all, and with what interval — the watcher
+     * that would apply a later change is itself built once, so it is deferred rather than cold (see
+     * {@link ConfigRef}'s class doc: cold means an already-open resource would go inconsistent with
+     * the new value, and there is no such resource here — a running watcher just keeps polling on its
+     * original enabled/interval until a restart, exactly like {@code lifecycle} or {@code guard}).
+     * Before this fix, turning reload off (or changing its interval) through a reload reported a bare
+     * "config reloaded" — the obvious joke the issue names.
+     */
+    @Test
+    void changingConfigReloadIsReportedAsDeferred(@TempDir Path dir) throws Exception {
+        Path f = dir.resolve("fleetd.yaml");
+        Files.writeString(f, yaml("""
+                configReload:
+                  enabled: true
+                  intervalSeconds: 10
+                """));
+        ConfigRef ref = refFor(f);
+
+        Files.writeString(f, yaml("""
+                configReload:
+                  enabled: false
+                  intervalSeconds: 30
+                """));
+        ConfigRef.Outcome out = ref.reload();
+
+        assertTrue(out.applied());
+        assertEquals(java.util.List.of("configReload"), out.deferred());
+        assertTrue(out.summary().contains("need") && out.summary().contains("restart"), out.summary());
+        // The snapshot still carries the new value — a restart is what makes it take effect.
+        assertFalse(ref.get().configReload().isEnabled());
+        assertEquals(30, ref.get().configReload().intervalSeconds());
+    }
+
     @Test
     void aFixedRefHasNoFileAndRefusesToReload() {
         FleetConfig cfg = new FleetConfig(null, null, null, null, null, null,

fleetd/src/test/java/dev/ltms/fleet/mcp/ConnectionIdentityTest.java

@@ -43,6 +43,22 @@ class ConnectionIdentityTest {
         assertNull(with(_ -> 999_999).callerTerminal("127.0.0.1", 55555));
     }
 
+    @Test
+    void callerIsUnresolvedWhenThePeerPidLookupFails() {
+        // fleetd #317: LsofPeerPidLookup returns -1 on any failure — a fork error, or (silently)
+        // simply no matching lsof line. Caller.resolved() is the one place that sentinel is tested.
+        ConnectionIdentity.Caller c = with(_ -> -1).resolve("127.0.0.1", 55555);
+        assertFalse(c.resolved(), "a -1 pid means the lookup failed, not that this pid owns no pane");
+    }
+
+    @Test
+    void callerIsResolvedWhenThePidIsRealEvenThoughItOwnsNoPane() {
+        // The primary's own connection: a real, lsof-found pid that just isn't a worker pane. This
+        // must read as "resolved" — the distinction #317 turns on.
+        ConnectionIdentity.Caller c = with(_ -> 999_999).resolve("127.0.0.1", 55555);
+        assertTrue(c.resolved());
+    }
+
     @Test
     void resolvesTheCallersPidAndCwd() {
         // CB-112: the primary maps to no pane, but its PID and cwd are still readable.

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

@@ -615,6 +615,67 @@ class CompositePeerLauncherTest {
         assertEquals(1, adapter.spawnCount("b"));
     }
 
+    /**
+     * fleetd #315: {@code CompositePeerLauncher.spawn} rebuilds the {@link PlacementContext} after
+     * every failed attempt "so the policy excludes this profile" (see the comment at the retry call
+     * site) — but {@code FixedPlacementPolicy} never read {@code ctx.unreachable()}, so under the
+     * default {@code fixed} placement every retry re-picked the same dead default and a second,
+     * healthy, configured profile was never tried. This is the same scenario as
+     * {@link #failoverRetriesNextCandidateWhenProfileIsUnreachable}, but pinned to {@code fixed()}
+     * instead of {@code weighted()} — the three existing failover tests all use {@code weighted()},
+     * which is exactly why nobody caught this: the retry loop's contract has no coverage under its
+     * own default policy.
+     */
+    @Test
+    void failoverRetriesNextCandidateUnderFixedPlacementWhenProfileIsUnreachable() {
+        FakeHerdr herdr = new FakeHerdr();
+        Map<String, FleetConfig.Profile> profiles = ordered(
+                "a", stubWorker("a"),
+                "b", stubWorker("b"));
+        StubLauncher adapter = new StubLauncher("claude", herdr, profiles, "a", Set.of("a"));
+        CompositePeerLauncher composite = new CompositePeerLauncher(
+                List.of(adapter), "a", profiles, PlacementPolicies.fixed(), _ -> 0);
+
+        PeerHandle h = composite.spawn(new SpawnRequest(null, null, null));
+        assertEquals("b", h.profile(),
+                "fixed placement must fail over from the unreachable default a to the healthy b");
+        assertEquals(1, adapter.spawnCount("a"), "a was tried once and failed");
+        assertEquals(1, adapter.spawnCount("b"), "b was tried once and succeeded");
+    }
+
+    /**
+     * fleetd #315: the same fix — {@code FixedPlacementPolicy} consulting {@code ctx.unreachable()}
+     * — also covers the wiring-bug branch in {@code CompositePeerLauncher.spawn}: a profile that
+     * placement is allowed to choose (it is in the configured candidate list) but that no delegate
+     * declares ({@code byProfile.get(chosen.profile()) == null}). That branch adds the profile to
+     * {@code unreachable} and {@code continue}s without ever calling a launcher, so before this fix
+     * {@code fixed} handed back the same adapterless profile on every remaining attempt too.
+     */
+    @Test
+    void failoverSkipsAConfiguredProfileNoAdapterDeclaresUnderFixedPlacement() {
+        FakeHerdr herdr = new FakeHerdr();
+        // Placement's candidate list has three profiles, in this order (LinkedHashMap preserves it,
+        // and the fixed default resolves to the first — see the `ordered` helper's own javadoc).
+        Map<String, FleetConfig.Profile> profiles = new LinkedHashMap<>();
+        profiles.put("c", stubWorker("c"));
+        profiles.put("a", stubWorker("a"));
+        profiles.put("b", stubWorker("b"));
+        // The adapter only declares a and b — c is a configured profile with no owning adapter,
+        // the "wiring bug" the comment in CompositePeerLauncher.spawn calls out.
+        Map<String, FleetConfig.Profile> adapterProfiles = new LinkedHashMap<>();
+        adapterProfiles.put("a", stubWorker("a"));
+        adapterProfiles.put("b", stubWorker("b"));
+        StubLauncher adapter = new StubLauncher("claude", herdr, adapterProfiles, "a", Set.of());
+        CompositePeerLauncher composite = new CompositePeerLauncher(
+                List.of(adapter), "a", profiles, PlacementPolicies.fixed(), _ -> 0);
+
+        PeerHandle h = composite.spawn(new SpawnRequest(null, null, null));
+        assertEquals("a", h.profile(),
+                "c has no adapter, so fixed placement must skip it and land on the next candidate, a");
+        assertEquals(0, adapter.spawnCount("c"), "c is never spawned — no adapter owns it");
+        assertEquals(1, adapter.spawnCount("a"));
+    }
+
     @Test
     void explicitSpawnAtMaxLoadThrowsPlacementExceptionNamingProfileLiveAndCap() {
         FakeHerdr herdr = new FakeHerdr();

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

@@ -86,17 +86,50 @@ class SessionManagerTest {
         private volatile RuntimeException hasUncommittedFailure;
         private volatile RuntimeException snapshotFailure;
         private final java.util.concurrent.atomic.AtomicLong snapshotSeq = new java.util.concurrent.atomic.AtomicLong();
+        /** fleetd #316: successive {@code hasUncommitted} answers, one per call, last one sticky
+         *  once exhausted — models a worktree whose state changes between reads. Empty (the
+         *  default) falls back to the plain {@link #dirty} flag, so every existing test using this
+         *  fake keeps returning one fixed answer. */
+        private final List<Boolean> dirtySequence = new java.util.concurrent.CopyOnWriteArrayList<>();
+        private int failHasUncommittedOnCall = -1;
+        private RuntimeException hasUncommittedCallFailure;
+        private final java.util.concurrent.atomic.AtomicInteger hasUncommittedCalls =
+                new java.util.concurrent.atomic.AtomicInteger();
 
         RecordingWorktrees dirty(boolean dirty) {
             this.dirty = dirty;
             return this;
         }
 
+        /** fleetd #316: return {@code answers[0]} on the first {@code hasUncommitted} call,
+         *  {@code answers[1]} on the second, and so on; the last element repeats after that. */
+        RecordingWorktrees dirtySequence(boolean... answers) {
+            for (boolean a : answers) {
+                dirtySequence.add(a);
+            }
+            return this;
+        }
+
+        int hasUncommittedCallCount() {
+            return hasUncommittedCalls.get();
+        }
+
         RecordingWorktrees failHasUncommittedWith(RuntimeException e) {
             this.hasUncommittedFailure = e;
             return this;
         }
 
+        /**
+         * Throw from {@code hasUncommitted} on one specific call only, counting from 0. The
+         * whole-double {@link #failHasUncommittedWith} cannot express fleetd #316's fail-safe
+         * case, which needs the pre-stop read to succeed and only the late read to fail.
+         */
+        RecordingWorktrees failHasUncommittedOnCall(int call, RuntimeException e) {
+            this.failHasUncommittedOnCall = call;
+            this.hasUncommittedCallFailure = e;
+            return this;
+        }
+
         RecordingWorktrees failRemoveFor(String worktreePath) {
             failRemoveFor.add(worktreePath);
             return this;
@@ -127,9 +160,16 @@ class SessionManagerTest {
 
         @Override
         public boolean hasUncommitted(String worktreePath) {
+            int call = hasUncommittedCalls.getAndIncrement();
             if (hasUncommittedFailure != null) {
                 throw hasUncommittedFailure;
             }
+            if (call == failHasUncommittedOnCall) {
+                throw hasUncommittedCallFailure;
+            }
+            if (!dirtySequence.isEmpty()) {
+                return dirtySequence.get(Math.min(call, dirtySequence.size() - 1));
+            }
             return dirty;
         }
 
@@ -1207,6 +1247,104 @@ class SessionManagerTest {
                         + "dirty check threw");
     }
 
+    // --- fleetd #316: the dirty check must be re-taken after the worker is stopped, not trusted
+    // stale from before it ------------------------------------------------------------------------
+
+    @Test
+    void releaseDoesNotRemoveAWorktreeThatBecameDirtyBetweenTheFirstCheckAndRemoval() {
+        // Models the exact race #316 reports: hasUncommitted answers clean while the worker is
+        // still running (call 1), the worker then writes new work, and by the time release is
+        // about to force-remove the worktree a second read (call 2) would see it as dirty. Without
+        // the fix this test fails: release() never re-reads and force-removes the worktree anyway.
+        FakeHerdr herdr = new FakeHerdr();
+        RecordingWorktrees worktrees = new RecordingWorktrees().dirtySequence(false, true);
+        SessionManager sessions = sessionManager(herdr, worktrees);
+        MemberSession s = sessions.acquire("ltms-local", null, "/caller/proj", null,
+                new WorktreeRequest("cb-316a", null));
+
+        sessions.release(s.paneId());
+
+        assertTrue(worktrees.removeCalls().isEmpty(),
+                "a worktree that turned dirty between the pre-stop read and removal must be preserved");
+        assertEquals(2, worktrees.hasUncommittedCallCount(),
+                "the fix re-reads hasUncommitted exactly once more, immediately before removal");
+    }
+
+    @Test
+    void releasePreservesAWorktreeWhoseLateRecheckCannotBeRead() {
+        // fleetd #316 invariant 1, which no test pinned when the fix landed: the late re-check
+        // fails toward PRESERVING. Found by mutation — flipping dirtyImmediatelyBeforeRemoval's
+        // catch from `return true` to `return false` turned the guard into a cause of the very
+        // data loss it was added to stop, and the whole suite stayed green. The pre-stop read
+        // succeeds and says clean (call 0); the read that authorises the removal throws (call 1).
+        FakeHerdr herdr = new FakeHerdr();
+        RecordingWorktrees worktrees = new RecordingWorktrees()
+                .dirtySequence(false)
+                .failHasUncommittedOnCall(1, new WorktreeException("git status exited 128"));
+        SessionManager sessions = sessionManager(herdr, worktrees);
+        MemberSession s = sessions.acquire("ltms-local", null, "/caller/proj", null,
+                new WorktreeRequest("cb-316c", null));
+
+        sessions.release(s.paneId());
+
+        assertTrue(worktrees.removeCalls().isEmpty(),
+                "a worktree whose state cannot be read immediately before removal must be kept: "
+                        + "preserving costs disk, deleting on a guess destroys work with no other copy");
+        assertEquals(2, worktrees.hasUncommittedCallCount(),
+                "the late re-check still runs — it is the throwing call, not a skipped one");
+    }
+
+    @Test
+    void releaseSnapshotsWorkFoundOnlyByTheLateRecheck() {
+        // #316's second half: the pre-stop dirty=false means trySnapshot never ran for this
+        // session, so the late-discovered work would otherwise have no refs/wip/* copy at all —
+        // only the on-disk preserve. The re-check path must snapshot it too.
+        FakeHerdr herdr = new FakeHerdr();
+        RecordingWorktrees worktrees = new RecordingWorktrees().dirtySequence(false, true);
+        SessionManager sessions = sessionManager(herdr, worktrees);
+        MemberSession s = sessions.acquire("ltms-local", null, "/caller/proj", null,
+                new WorktreeRequest("cb-316b", null));
+
+        sessions.release(s.paneId());
+
+        assertEquals(java.util.List.of(s.worktree()), worktrees.snapshotCalls(),
+                "the newly-dirty worktree is snapshotted even though the pre-stop check saw it clean");
+    }
+
+    @Test
+    void releaseStillRemovesAWorktreeThatStaysCleanOnTheLateRecheck() {
+        // The ordinary, non-racing case: nothing else changes behaviour when the second read
+        // agrees with the first.
+        FakeHerdr herdr = new FakeHerdr();
+        RecordingWorktrees worktrees = new RecordingWorktrees().dirty(false);
+        SessionManager sessions = sessionManager(herdr, worktrees);
+        MemberSession s = sessions.acquire("ltms-local", null, "/caller/proj", null,
+                new WorktreeRequest("cb-316c", null));
+
+        sessions.release(s.paneId());
+
+        assertEquals(java.util.List.of(s.worktree()), worktrees.removeCalls(),
+                "a worktree that is still clean on the late recheck is removed as before");
+    }
+
+    @Test
+    void releaseNeverReChecksAWorktreeAlreadyPreservedByTheFirstDirtyCheck() {
+        // Invariant 4 from #316: no second unconditional git status. A release that already
+        // decided to preserve (the ordinary CB-576 dirty path) must not pay for a second read.
+        FakeHerdr herdr = new FakeHerdr();
+        RecordingWorktrees worktrees = new RecordingWorktrees().dirty(true);
+        SessionManager sessions = sessionManager(herdr, worktrees);
+        MemberSession s = sessions.acquire("ltms-local", null, "/caller/proj", null,
+                new WorktreeRequest("cb-316d", null));
+
+        sessions.release(s.paneId());
+
+        assertEquals(1, worktrees.hasUncommittedCallCount(),
+                "a release that already preserves on the first read must not re-check before "
+                        + "skipping the removal it was never going to do");
+        assertTrue(worktrees.removeCalls().isEmpty());
+    }
+
     /**
      * fleetd #283 defect 1 changed this test's own premise, so its assertions are updated along
      * with the production fix. Before #283, the middle session's worktree-removal failure escaped