fleetd #550: pin hash256's algorithm against a literal SHA-256 test vector

test_jar_id_defaults_to_live_and_reports_explicit_path's reference hash is computed by calling
hash256 itself (needed so it doesn't call the Linux-crashing bare shasum directly). That made
subject and reference the same instrument: they agree no matter which algorithm hash256 actually
runs, so a mutation swapping both of hash256's arms for the wrong algorithm was invisible to the
suite.

Adds test_hash256_computes_a_real_sha256, pinned against the published SHA-256 test vector for the
3-byte input "abc" (ba7816bf8f01...), written as a literal constant rather than computed by any
hasher at test time. Verified the constant myself both ways (sha256sum and shasum -a 256) before
writing it in.

.gitea/workflows/ci.yml

@@ -39,6 +39,10 @@ jobs:
         # that runs does so against the fake UDS herdr and fake ccs/claude stubs.
         run: mvn -B clean install
 
+      - name: javadoc reference lint
+        working-directory: fleetd
+        run: mvn -B -DskipTests javadoc:javadoc -Ddoclint=reference
+
       # Deliberately NOT actions/upload-artifact: this Gitea instance presents as GHES, and
       # @actions/artifact v2+ (i.e. upload-artifact@v4) refuses to run there —
       # "GHESNotSupportedError ... not currently supported on GHES", which red-Xes an otherwise
@@ -54,6 +58,23 @@ jobs:
           done
           exit 0
 
+  # fleetd #550 — nothing ran scripts/test-redeploy-fleetd.sh in CI before this, on any platform,
+  # so it had run only on macOS by hand and two Linux-only bugs (this issue's items 1 and 2)
+  # survived undetected: shasum is a macOS-only tool (it ships with Perl; GNU coreutils, i.e. every
+  # mainstream Linux distro including this runner's ubuntu-latest, does not have it and ships
+  # sha256sum instead). The gate here is the step's own exit code, nothing else: a `run:` step in
+  # Gitea/GitHub Actions already fails the job on a non-zero exit with no extra scripting needed,
+  # so this deliberately does NOT grep the output for a `FAIL:` count. That is the #550 item-2
+  # lesson one level up — a suite that dies before it runs a single test prints zero FAIL lines,
+  # which is exactly what a clean pass also prints, so counting FAIL lines can never be the gate.
+  shell-tests:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: redeploy-fleetd.sh shell suite
+        run: bash scripts/test-redeploy-fleetd.sh
+
   # CB-521 — actually run the AMQP contract test in CI, against a REAL broker. The broker is a
   # RabbitMQ SERVICE CONTAINER, not Testcontainers-with-Docker: the runner image has no Docker, so
   # AmqpReplyInboxContractTest reads AMQP_URI (set below to the service's network alias) and binds

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

@@ -696,11 +696,7 @@ public final class Fleetd {
         FleetMcp mcp = new FleetMcp(messages, workers, sessions, identity, presence,
                 primaryRegistry, callers, FleetMcp.AuthorizationMode.ENFORCED, metrics,
                 capacitySource(config, cfg, profile -> liveCountRef.get().apply(profile)),
-                new FleetMcp.HealthCoverageSource(() -> {
-                    var health = config.get().health();
-                    return FleetHealthMonitor.coverage(health != null && health.isEnabled(),
-                            health != null && health.notifications() != null && health.notifications().configured());
-                }),
+                healthCoverageSource(config),
                 quarantineSource,
                 leadMailbox,
                 outageSource,
@@ -1038,6 +1034,38 @@ public final class Fleetd {
                 cfg.profiles()::keySet, System::nanoTime);
     }
 
+    /**
+     * fleetd #426: package-private factory for {@code fleet_list}'s {@code healthCoverage} source,
+     * extracted out of {@code main} for the same reason {@link #capacitySource} and {@link
+     * #quarantineSource} were — and the same reason {@link #exhaustedPatternCoverageLine}/{@link
+     * #errorPatternCoverageLine} exist: {@link FleetHealthMonitor#coverage}'s three-branch method
+     * is easy to pin directly (a plain {@code (boolean, boolean) -> String} call), but that proves
+     * nothing about whether <em>this call site</em> pairs the right boolean with the right meaning.
+     * fleetd #415's measured lesson is the reason this matters here — swapping the two arguments at
+     * a call site like this one compiled clean and left the full suite green, because every existing
+     * test exercised the method in both directions without ever exercising the pairing.
+     *
+     * <p>{@code #407}'s "keep the config invalid, assert on the log line before the throw" option
+     * does not apply to this call site: the five reporters #407 covers all run in {@code main}
+     * <em>before</em> {@code cfg.validateAll()} (line ~171), so an invalid config still exercises
+     * them. This call site is built during {@code FleetMcp} construction, which runs only after
+     * {@code UnixSocketHerdrClient.connect} has already opened a real herdr socket (line ~188) —
+     * reaching it at all means main already performed real I/O, which the no-socket constraint on
+     * this ticket rules out. So the pairing is pinned by extracting it to this directly-callable
+     * factory instead, the same shape {@link #capacitySource}/{@link #quarantineSource} already use.
+     *
+     * <p>Reads {@code config.get().health()} live (health.notifications is a {@code SPLIT_KEYS}
+     * entry — see {@link ConfigRef#SPLIT_KEYS}), so a hot-reloaded notifications block changes what
+     * {@code fleet_list} reports without a restart, exactly like {@link #capacitySource}'s maxLoad.
+     */
+    static FleetMcp.HealthCoverageSource healthCoverageSource(ConfigRef config) {
+        return new FleetMcp.HealthCoverageSource(() -> {
+            var health = config.get().health();
+            return FleetHealthMonitor.coverage(health != null && health.isEnabled(),
+                    health != null && health.notifications() != null && health.notifications().configured());
+        });
+    }
+
     /**
      * fleetd #248: package-private factory for the member worktree/branch lookup {@link
      * CompletionResolver} uses to name a fallback report's worktree and branch (fleetd#241).

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

@@ -208,7 +208,7 @@ import java.util.function.Supplier;
  * that is correctly hot and for a key nobody triaged. Three times now — {@code worktreeGroup} (#323),
  * {@code primary}/{@code configReload} (#326), and {@code fleet.leaders} sitting in the escape hatch
  * (#333) — the second kind hid among the first. A top-level coverage checker in the
- * {@link ConfigRefProfileCoverageTest} shape (one level up, over {@code FleetConfig} itself rather
+ * {@code ConfigRefProfileCoverageTest} shape (one level up, over {@code FleetConfig} itself rather
  * than {@code FleetConfig.Profile}) proves this file's four classes exhaust the record's components
  * — see {@code ConfigRefTopLevelCoverageTest}. That test proves the record's <em>shape</em> is fully
  * triaged; it does NOT prove a {@code SPLIT_KEYS}/{@code COLD_KEYS}/{@code DEFERRED_KEYS} member has

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

@@ -1670,7 +1670,7 @@ public record FleetConfig(
      * <p>A herdr pane runs a login shell that re-sources the operator's own secret store, so a
      * member inherits every credential the operator's shell holds — measured at 31 names on this
      * host, of which only one ({@code GITEA_ACCESS_TOKEN}) used to be blocked, and that block was a
-     * single name hardcoded in {@link HerdrPeerLauncher} rather than driven by config (gitea issue
+     * single name hardcoded in {@link dev.ltms.fleet.member.HerdrPeerLauncher} rather than driven by config (gitea issue
      * #82). This record replaces that hardcoded shadow with a config-driven one.
      *
      * <p><b>deny-by-default, not a deny-list.</b> A deny-list (block these specific names, let

fleetd/src/main/java/dev/ltms/fleet/inject/Injector.java

@@ -306,7 +306,7 @@ public final class Injector {
         if (t == null) return;
 
         Pending sent = null;
-        RuntimeException sendError = null;
+        Throwable sendError = null;
         boolean turnCompleted = false;
         boolean turnFailed = false;
         boolean resubmit = false;
@@ -388,9 +388,14 @@ public final class Injector {
                                 t.turnObserved = false;
                                 t.injectableSincePickup = 0;
                                 sent = p;
-                            } catch (RuntimeException e) {
+                            } catch (Throwable e) {
                                 // Delivery failed at herdr; drop the poisoned message and surface it
-                                // rather than blocking the queue behind it.
+                                // rather than blocking the queue behind it. Catches Throwable, not just
+                                // RuntimeException: fleetd #546 — an Error escaping this send (e.g. a
+                                // NoClassDefFoundError, see #413) would otherwise leave the entry QUEUED
+                                // at the head of t.queue. Line :378 peeks rather than polls, so the next
+                                // onStatus round would re-enter this try and send the same text again,
+                                // typing the same brief into the member's pane a second time.
                                 t.queue.poll();
                                 p.state = Pending.State.NOT_DELIVERED;
                                 sent = p;

fleetd/src/main/java/dev/ltms/fleet/inject/StatusPoller.java

@@ -64,34 +64,39 @@ public final class StatusPoller {
     }
 
     private void loop() {
-        while (running) {
-            Set<String> active = injector.activeTargets();
-            for (String target : active) {
-                if (!running) return;
-                try {
-                    // herdr's agent_status can misreport a settled worker as `unknown`; refine it
-                    // against the pane content before it drives delivery/completion (CB-115).
-                    // CB-185: refine THROUGH the same control the raw status came from — a router
-                    // splits lead/member targets across two herdr daemons, and reading a lead's pane
-                    // through the (fixed) member refiner never finds it, wedging that lead at UNKNOWN.
-                    AgentControl control = router != null ? router.agentsFor(target) : agents;
-                    AgentStatus status = refiner.refine(target, control.status(target), control);
-                    injector.onStatus(target, status);
-                } catch (HerdrException e) {
-                    // The worker's agent is gone — stop trying and unblock its waiters.
-                    if (e.code() != null && e.code().endsWith("_not_found")) {
-                        log.debug("target {} gone; dropping its queue", target);
-                        injector.drop(target, e);
-                    } else {
-                        log.debug("status poll for {} failed (will retry): {}", target, e.getMessage());
+        try {
+            while (running) {
+                Set<String> active = injector.activeTargets();
+                for (String target : active) {
+                    if (!running) return;
+                    try {
+                        // herdr's agent_status can misreport a settled worker as `unknown`; refine it
+                        // against the pane content before it drives delivery/completion (CB-115).
+                        // CB-185: refine THROUGH the same control the raw status came from — a router
+                        // splits lead/member targets across two herdr daemons, and reading a lead's pane
+                        // through the (fixed) member refiner never finds it, wedging that lead at UNKNOWN.
+                        AgentControl control = router != null ? router.agentsFor(target) : agents;
+                        AgentStatus status = refiner.refine(target, control.status(target), control);
+                        injector.onStatus(target, status);
+                    } catch (HerdrException e) {
+                        // The worker's agent is gone — stop trying and unblock its waiters.
+                        if (e.code() != null && e.code().endsWith("_not_found")) {
+                            log.debug("target {} gone; dropping its queue", target);
+                            injector.drop(target, e);
+                        } else {
+                            log.debug("status poll for {} failed (will retry): {}", target, e.getMessage());
+                        }
+                    } catch (Throwable e) {
+                        log.error("unexpected failure polling {}; skipping this round", target, e);
                     }
-                } catch (RuntimeException e) {
-                    // Never let one target's unexpected error (e.g. an odd agent.get shape) kill
-                    // the single poller thread and stall injection for every worker.
-                    log.warn("unexpected error polling {}; skipping this round", target, e);
                 }
+                sleep();
             }
-            sleep();
+        } finally {
+            if (running) {
+                log.error("status poller loop exited unexpectedly; it can be restarted");
+            }
+            running = false;
         }
     }
 

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

@@ -246,7 +246,7 @@ public final class AmqpReplyInbox implements ReplyInbox, AutoCloseable {
      * neither. So before this method existed with a requeue step, it dropped {@link #held}'s entries
      * for {@code target} while the broker still considered them outstanding: never acked, never
      * nacked, never requeued, and no longer reachable by {@link #peek} — permanently invisible. This
-     * is unlike {@link #handleRecovery} and {@link #close()}, whose bare {@code held.clear()} is
+     * is unlike {@link RecoveryListener#handleRecovery(Recoverable)} and {@link #close()}, whose bare {@code held.clear()} is
      * correct because each has already made the broker requeue (a real connection drop, or
      * {@code channel.close()} respectively) before clearing local state.
      *

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

@@ -713,7 +713,7 @@ public final class MessageService {
      * failure.
      *
      * <p><strong>Without {@code sweepAsking} on the release path, a target torn down while
-     * genuinely {@code ASKING} was unrecoverable.</strong> {@link #resolveQuestion} had already
+     * genuinely {@code ASKING} was unrecoverable.</strong> {@link Rendezvous#resolveQuestion(String, String, String)} had already
      * closed the forward waiter the instant the question surfaced (so the {@code waiter} branch
      * below finds nothing to fail), the {@code question == null} guard excluded the task from
      * {@code matching} (so the loop below skipped it too), and the worker's own {@code fleet_ask}

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

@@ -44,7 +44,7 @@ import java.util.stream.Collectors;
  * still holds an unacked message ({@link #pendingReplies}), tickets not yet collected
  * ({@link #pendingTickets}), and open questions not yet answered or lapsed
  * ({@link #pendingQuestions}) — and sends at most one combined nudge per tick
- * ({@link #injectNudge(String, int, int, int)}). Work that arrives while the lead is busy is
+ * ({@link #injectNudge(String, int, int, int, int, int)}). Work that arrives while the lead is busy is
  * never lost: it is re-read fresh on every tick until the lead is injectable or its own reminder
  * cap ({@link #maxReminders}) is reached — each source spends from its own budget, so one source
  * exhausting its cap does not stop nudges about the others (post-CB-590 regression fix; see

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

@@ -1074,6 +1074,29 @@ public final class SessionManager implements TurnListener {
      * members were live) and must still produce the line. Both {@link #drainSnapshot} passes (the
      * main snapshot and the straggler sweep) are folded into the one line: a caller reading two
      * lines could not tell a two-pass drain from two separate drains.
+     *
+     * <p><strong>Non-goal: this line must never move into a {@code finally} block, and this method
+     * must never grow one around it.</strong> "Every time" above means every time the drain
+     * <em>finishes</em>, not every time this method exits. The absence of the line is the signal
+     * that the drain died, so a {@code finally} would destroy the signal and print confident
+     * partial counts in the same edit — the line would appear after a drain that threw, carrying
+     * whatever {@code tally} it had reached. Both halves of the value are lost at once. The line
+     * has to be the last statement of the successful path and reachable only from it.
+     *
+     * <p>This is written down because it is the obvious review comment ("shouldn't we always log
+     * the drain result?"), it sounds like thoroughness, and the paragraph above reads as an
+     * invitation to it. Raised by the fleet01 lead on 2026-09-12, from their 2026-09-10 incident:
+     * we only know that drain died on that host because it <em>threw</em>, and a
+     * {@code NoClassDefFoundError} reached the JVM's uncaught handler. A drain that hung on one
+     * session, or returned early on a condition rather than an exception, would leave no stack
+     * trace, no {@code ERROR} token and no priority — only a missing line. That makes the loud
+     * variant the one we have seen and the quiet variants the ones this line exists to catch.
+     *
+     * <p>Related: {@code released} and {@code abandoned} are counted incrementally inside {@link
+     * #drainSnapshot}'s loop and folded with {@link DrainTally#plus}, rather than derived from a
+     * collection read at the end, for the same reason. If a partial report is ever wanted it must
+     * be a different line with a different verb. One line must not serve both, or a reader cannot
+     * tell a finished drain from an interrupted one by its wording.
      */
     void drainAll(long timeoutNanos) {
         long deadline = System.nanoTime() + timeoutNanos;

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

@@ -60,14 +60,21 @@ public final class SessionReaper {
     }
 
     private void loop() {
-        while (running) {
-            try {
-                sessions.reapIdle(idleTtlNanos);
-            } catch (RuntimeException e) {
-                log.warn("session reaper iteration failed; continuing", e);
+        try {
+            while (running) {
+                try {
+                    sessions.reapIdle(idleTtlNanos);
+                } catch (Throwable e) {
+                    log.error("session reaper iteration failed; continuing", e);
+                }
+                maybeSweepWipRefs();
+                sleep();
             }
-            maybeSweepWipRefs();
-            sleep();
+        } finally {
+            if (running) {
+                log.error("session reaper loop exited unexpectedly; it can be restarted");
+            }
+            running = false;
         }
     }
 
@@ -90,8 +97,8 @@ public final class SessionReaper {
                 log.info("refs/wip retention sweep deleted {} snapshot ref(s) older than 24h whose "
                         + "content was already reachable from main", deleted);
             }
-        } catch (RuntimeException e) {
-            log.warn("refs/wip retention sweep failed; continuing", e);
+        } catch (Throwable e) {
+            log.error("refs/wip retention sweep failed; continuing", e);
         }
         // Set even when the sweep threw, so a broken repo is retried on the slow cadence rather
         // than hammering git on every 5-second iteration.

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

@@ -0,0 +1,160 @@
+package dev.ltms.fleet;
+
+import dev.ltms.fleet.config.ConfigRef;
+import dev.ltms.fleet.config.FleetConfig;
+import dev.ltms.fleet.mcp.FleetMcp;
+import org.junit.jupiter.api.DisplayName;
+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.assertEquals;
+import static org.junit.jupiter.api.Assertions.assertTrue;
+
+/**
+ * fleetd #426: {@code FleetHealthMonitor.coverage} had zero references anywhere in the test
+ * tree — not the method, not either output string, not the field it populates. {@code
+ * FleetHealthMonitorCoverageTest} (package {@code dev.ltms.fleet.health}) pins the three-branch
+ * method itself; that is the easy half.
+ *
+ * <p>The half that actually matters is this one: {@link Fleetd#healthCoverageSource} is the exact
+ * call site {@code Fleetd.main} wires into {@code FleetMcp}'s constructor, and it is what feeds
+ * {@code fleet_list}'s {@code healthCoverage} field (see {@code FleetMcp#listFleet}'s {@code
+ * result.put("healthCoverage", healthCoverage.value().get())}). Measured precedent on fleetd #423
+ * (for #415): swapping two arguments at a call site like this one — recreating #415's defect with
+ * the keys exchanged — compiled with 0 errors and ran the ENTIRE suite (1506 tests) green. A
+ * thoroughly-tested method proves nothing about whether the call site pairs its arguments correctly;
+ * only a test that drives the call site itself can catch that.
+ *
+ * <p><strong>Why this is not driven through a real {@code Fleetd.main} the way #407 drives its five
+ * reporters</strong> (keep the config invalid, assert on the log line emitted before {@code
+ * cfg.validateAll()} throws): all five of #407's reporters run in {@code main} before {@code
+ * validateAll()} (line ~171). {@link Fleetd#healthCoverageSource} is built during {@code FleetMcp}
+ * construction, which happens only after {@code UnixSocketHerdrClient.connect} has already opened a
+ * real herdr socket (line ~188) and after {@code sessions}/{@code workers} are constructed. Reaching
+ * this call site by actually running {@code main} would require a real socket connect — banned by
+ * this ticket's hard constraints — so #407's option 1 does not apply here. Instead {@link
+ * Fleetd#healthCoverageSource} is extracted to a directly-callable package-private factory, the same
+ * shape {@link Fleetd#capacitySource} and {@link Fleetd#quarantineSource} already use for the same
+ * reason (see {@code FleetdCapacitySourceWiringTest}, the direct precedent this test follows).
+ *
+ * <p>Uses a real {@link FleetConfig#load} + {@link ConfigRef} (no socket, no port bind, no spawn,
+ * nothing written outside {@code @TempDir}) so the fixture goes through the actual YAML parser and
+ * {@code FleetConfig.Health}/{@code Notifications} records, not a hand-built stand-in that could
+ * silently drift from what the parser actually produces.
+ */
+class FleetdHealthCoverageSourceWiringTest {
+
+    private static final String BASE = """
+            bind:
+              host: 127.0.0.1
+              port: 8765
+            herdrSocket: ~/.config/herdr/herdr.sock
+            """;
+
+    private static final String HEALTH_DISABLED_WITH_WEBHOOK = BASE + """
+            health:
+              enabled: false
+              notifications:
+                mode: webhook
+            """;
+
+    private static final String HEALTH_DETECTION_ONLY = BASE + """
+            health:
+              enabled: true
+            """;
+
+    private static final String HEALTH_FULL = BASE + """
+            health:
+              enabled: true
+              notifications:
+                mode: webhook
+            """;
+
+    private static final String HEALTH_ABSENT = BASE;
+
+    @Test
+    @DisplayName("enabled: false reports off, even with a webhook configured")
+    void disabledHealthReportsOff(@TempDir Path dir) throws Exception {
+        FleetMcp.HealthCoverageSource source = sourceFor(dir, HEALTH_DISABLED_WITH_WEBHOOK);
+
+        assertEquals("off", source.value().get(),
+                "health.enabled: false must report 'off' regardless of notifications — flipping "
+                        + "the 'enabled' argument at the HealthCoverageSource call site would report "
+                        + "'full' here instead");
+    }
+
+    @Test
+    @DisplayName("enabled with no notifications reports detection-only")
+    void enabledWithoutNotificationsReportsDetectionOnly(@TempDir Path dir) throws Exception {
+        FleetMcp.HealthCoverageSource source = sourceFor(dir, HEALTH_DETECTION_ONLY);
+
+        assertEquals("detection-only", source.value().get());
+    }
+
+    @Test
+    @DisplayName("enabled with a webhook configured reports full")
+    void enabledWithNotificationsReportsFull(@TempDir Path dir) throws Exception {
+        FleetMcp.HealthCoverageSource source = sourceFor(dir, HEALTH_FULL);
+
+        assertEquals("full", source.value().get(),
+                "health.enabled: true with notifications.mode: webhook must report 'full' — "
+                        + "swapping 'full' and 'detection-only' at the call site, or breaking the "
+                        + "enabled/notificationConfigured argument pairing, would report "
+                        + "'detection-only' here instead");
+    }
+
+    @Test
+    @DisplayName("an absent health: block reports off")
+    void absentHealthBlockReportsOff(@TempDir Path dir) throws Exception {
+        FleetMcp.HealthCoverageSource source = sourceFor(dir, HEALTH_ABSENT);
+
+        assertEquals("off", source.value().get());
+    }
+
+    /**
+     * fleetd #426, the live-wiring half: {@code health.notifications} is a {@code
+     * ConfigRef.SPLIT_KEYS} entry, and {@link Fleetd#healthCoverageSource} reads {@code
+     * config.get().health()} live (not the frozen startup {@code cfg}) — exactly like {@link
+     * Fleetd#capacitySource}'s {@code maxLoad} ({@code FleetdCapacitySourceWiringTest}'s {@code
+     * reloadedMaxLoadStillChangesWhatFleetListReports}). A hot-reloaded notifications block must
+     * change what {@code fleet_list} reports without a restart; a fix that froze the whole source
+     * against the startup snapshot would silently break that and every other test above would stay
+     * green, since none of them reload.
+     */
+    @Test
+    @DisplayName("a hot notifications reload still changes what fleet_list reports")
+    void reloadedNotificationsStillChangeWhatFleetListReports(@TempDir Path dir) throws Exception {
+        Path file = dir.resolve("fleetd.yaml");
+        Files.writeString(file, HEALTH_DETECTION_ONLY);
+        FleetConfig cfg = FleetConfig.load(file);
+        ConfigRef config = new ConfigRef(file, cfg);
+
+        FleetMcp.HealthCoverageSource source = Fleetd.healthCoverageSource(config);
+        assertEquals("detection-only", source.value().get(),
+                "sanity: detection-only before any reload");
+
+        Files.writeString(file, HEALTH_FULL);
+        assertTrue(config.reload().applied());
+        // The live snapshot now carries the webhook — proves the reload really happened and this
+        // test is not accidentally passing because nothing changed.
+        assertTrue(config.get().health().notifications() != null
+                        && config.get().health().notifications().configured(),
+                "sanity: the reloaded config really carries a configured webhook");
+
+        assertEquals("full", source.value().get(),
+                "the SAME HealthCoverageSource instance must reflect a reloaded notifications "
+                        + "block without a restart — health.notifications is read live off "
+                        + "config.get(), exactly like capacitySource's maxLoad");
+    }
+
+    private static FleetMcp.HealthCoverageSource sourceFor(Path dir, String yaml) throws Exception {
+        Path file = dir.resolve("fleetd.yaml");
+        Files.writeString(file, yaml);
+        FleetConfig cfg = FleetConfig.load(file);
+        ConfigRef config = new ConfigRef(file, cfg);
+        return Fleetd.healthCoverageSource(config);
+    }
+}

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

@@ -0,0 +1,40 @@
+package dev.ltms.fleet.health;
+
+import org.junit.jupiter.api.Test;
+
+import static org.junit.jupiter.api.Assertions.assertEquals;
+
+/**
+ * fleetd #426: {@link FleetHealthMonitor#coverage} had zero references anywhere in the test tree —
+ * not the method, not either output string, not the field it populates. This pins the method's own
+ * three branches directly.
+ *
+ * <p>This is the easy half. It proves the method words each combination correctly, but it proves
+ * nothing about whether {@code Fleetd.java}'s two call sites pass the right argument in the right
+ * position — see {@code FleetdHealthCoverageSourceWiringTest} (package {@code dev.ltms.fleet}) for
+ * the half that actually guards the call site, following the measured fleetd #415 lesson that a
+ * thoroughly-tested method and an untested argument pairing at its call site are different risks.
+ *
+ * <p><strong>The three output strings are load-bearing and must not change here.</strong> {@code
+ * "detection-only"} is read live off a running daemon's {@code fleet_list} today (measured
+ * 2026-09-12) — this test intentionally asserts the exact literal strings so a future edit to the
+ * wording trips it here first.
+ */
+class FleetHealthMonitorCoverageTest {
+
+    @Test
+    void disabledIsOffRegardlessOfNotificationConfig() {
+        assertEquals("off", FleetHealthMonitor.coverage(false, false));
+        assertEquals("off", FleetHealthMonitor.coverage(false, true));
+    }
+
+    @Test
+    void enabledWithoutNotificationsIsDetectionOnly() {
+        assertEquals("detection-only", FleetHealthMonitor.coverage(true, false));
+    }
+
+    @Test
+    void enabledWithNotificationsIsFull() {
+        assertEquals("full", FleetHealthMonitor.coverage(true, true));
+    }
+}

fleetd/src/test/java/dev/ltms/fleet/inject/InjectorTest.java

@@ -2,9 +2,11 @@ package dev.ltms.fleet.inject;
 
 import ch.qos.logback.classic.Level;
 import ch.qos.logback.classic.spi.ILoggingEvent;
+import com.fasterxml.jackson.databind.JsonNode;
 import dev.ltms.fleet.herdr.AgentControl;
 import dev.ltms.fleet.herdr.AgentStatus;
 import dev.ltms.fleet.herdr.FakeHerdr;
+import dev.ltms.fleet.herdr.HerdrClient;
 import dev.ltms.fleet.herdr.HerdrException;
 import dev.ltms.fleet.msg.TestTurnTokens;
 import dev.ltms.fleet.testing.CapturedLog;
@@ -664,4 +666,111 @@ class InjectorTest {
         ExecutionException ex = assertThrows(ExecutionException.class, f::get);
         assertInstanceOf(HerdrException.class, ex.getCause());
     }
+
+    /**
+     * A {@link HerdrClient} that throws a non-{@link RuntimeException} {@link Error} from {@code
+     * agent.prompt} instead of delegating — the fleetd #546 case: a stray non-RuntimeException
+     * throwable (e.g. a {@code NoClassDefFoundError}, fleetd #413) escaping the send seam at
+     * {@code Injector.java:383}. Records every call it sees itself, including the ones it throws
+     * for, since the delegate's own recording is never reached for {@code agent.prompt} — so a test
+     * can assert on exactly what this fake actually received.
+     */
+    private static final class ErrorOnPrompt implements HerdrClient {
+        private final FakeHerdr delegate;
+        private final List<FakeHerdr.Call> calls = new java.util.concurrent.CopyOnWriteArrayList<>();
+
+        private ErrorOnPrompt(FakeHerdr delegate) {
+            this.delegate = delegate;
+        }
+
+        List<FakeHerdr.Call> calls() {
+            return calls;
+        }
+
+        @Override
+        public JsonNode call(String method, Object params) {
+            calls.add(new FakeHerdr.Call(method, params));
+            if (method.equals("agent.prompt")) {
+                throw new AssertionError("simulated non-RuntimeException send failure (fleetd #546)");
+            }
+            return delegate.call(method, params);
+        }
+
+        @Override
+        public void close() {
+            delegate.close();
+        }
+    }
+
+    /**
+     * Mirrors StatusPoller's own per-target {@code catch (Throwable)} (fleetd #538 / PR #543): the
+     * production polling loop already swallows whatever escapes one target's round and comes back
+     * for the next one. A unit test that calls {@code onStatus} directly (bypassing StatusPoller)
+     * needs the same survival so it can observe what a SECOND round does, regardless of whether
+     * fleetd #546's fix is present.
+     */
+    private static void pollOnceSurviving(Injector inj, String target, AgentStatus status) {
+        try {
+            inj.onStatus(target, status);
+        } catch (Throwable ignored) {
+            // matches StatusPoller.loop's own catch (Throwable) added by PR #543
+        }
+    }
+
+    @Test
+    void anErrorFromSendRemovesTheMessageAndMarksItNotDelivered() {
+        // fleetd #546, acceptance test 1: an Error (not a RuntimeException) escaping the send seam
+        // at Injector.java:383 must still be caught, the message dropped from the queue, and its
+        // state set to NOT_DELIVERED — never left QUEUED at the head.
+        ErrorOnPrompt throwing = new ErrorOnPrompt(new FakeHerdr());
+        Injector inj = new Injector(new AgentControl(throwing));
+        Injector.Delivery delivery = inj.enqueue(T, "brief", TestTurnTokens.inert(T));
+
+        assertDoesNotThrow(() -> inj.onStatus(T, AgentStatus.IDLE),
+                "fleetd #546: an Error from the send seam must be caught inside onStatus, not "
+                        + "escape it");
+
+        assertTrue(delivery.completion().isCompletedExceptionally(),
+                "the delivery's future must surface the send failure");
+        assertEquals(Injector.Cancellation.NOT_DELIVERED, inj.cancel(delivery),
+                "the message must be dropped and marked NOT_DELIVERED, not left QUEUED at the "
+                        + "head of the queue");
+    }
+
+    @Test
+    void anErrorFromSendDoesNotRedeliverOnASecondRound() {
+        // fleetd #546, acceptance test 2 — the test this ticket exists for. Before the fix, an
+        // Error at the send seam left the message QUEUED (Injector.java:378 peeks, not polls), so a
+        // second onStatus round re-entered the same try and sent the same text again: the member's
+        // pane got the same brief typed into it twice.
+        ErrorOnPrompt throwing = new ErrorOnPrompt(new FakeHerdr());
+        Injector inj = new Injector(new AgentControl(throwing));
+        inj.enqueue(T, "brief", TestTurnTokens.inert(T));
+
+        pollOnceSurviving(inj, T, AgentStatus.IDLE);
+        pollOnceSurviving(inj, T, AgentStatus.IDLE);
+
+        long promptCalls = throwing.calls().stream().filter(c -> c.method().equals("agent.prompt")).count();
+        assertEquals(1, promptCalls, "fleetd #546: the poisoned text must be sent exactly once — a "
+                + "second onStatus round must not re-enter send for the same message");
+    }
+
+    @Test
+    void aHerdrExceptionFromSendStillProducesNotDeliveredUnchanged() {
+        // fleetd #546, acceptance test 3 (control): HerdrException extends RuntimeException, so it
+        // was already caught before this ticket's widening. This pins that the ordinary path is
+        // unchanged — still dropped, still NOT_DELIVERED, still surfaced to the caller.
+        FakeHerdr failing = new FakeHerdr().agentSendFailsWith("send_failed");
+        Injector inj = new Injector(new AgentControl(failing));
+        Injector.Delivery delivery = inj.enqueue(T, "boom", TestTurnTokens.inert(T));
+
+        inj.onStatus(T, AgentStatus.IDLE);
+
+        assertTrue(delivery.completion().isCompletedExceptionally(),
+                "a HerdrException at the send seam must still surface to the caller, unchanged by "
+                        + "fleetd #546's widening");
+        assertEquals(Injector.Cancellation.NOT_DELIVERED, inj.cancel(delivery),
+                "a HerdrException must still be dropped and marked NOT_DELIVERED, unchanged by the "
+                        + "wider Throwable catch");
+    }
 }

fleetd/src/test/java/dev/ltms/fleet/inject/StatusPollerResilienceTest.java

@@ -0,0 +1,99 @@
+package dev.ltms.fleet.inject;
+
+import com.fasterxml.jackson.databind.JsonNode;
+import dev.ltms.fleet.herdr.AgentControl;
+import dev.ltms.fleet.herdr.FakeHerdr;
+import dev.ltms.fleet.herdr.HerdrClient;
+import dev.ltms.fleet.msg.TestTurnTokens;
+import org.junit.jupiter.api.Test;
+
+import java.lang.reflect.Field;
+import java.util.Map;
+import java.util.concurrent.CompletableFuture;
+import java.util.concurrent.CountDownLatch;
+import java.util.concurrent.TimeUnit;
+import java.util.concurrent.atomic.AtomicBoolean;
+
+import static org.junit.jupiter.api.Assertions.assertFalse;
+import static org.junit.jupiter.api.Assertions.assertNotSame;
+import static org.junit.jupiter.api.Assertions.assertTrue;
+
+class StatusPollerResilienceTest {
+
+    @Test
+    void anErrorForOneTargetDoesNotStopPollingTheNextTarget() throws Exception {
+        FakeHerdr fake = new FakeHerdr().withAgent("worker", "term_b", "w2:p8", "w2:t8");
+        CountDownLatch errorThrown = new CountDownLatch(1);
+        AgentControl agents = new AgentControl(new ErrorOnceForFirstTarget(fake, errorThrown));
+        Injector injector = new Injector(agents);
+        StatusPoller poller = new StatusPoller(agents, injector, 1);
+        poller.start();
+        try {
+            injector.enqueue("term_a", "first", TestTurnTokens.inert("term_a"));
+            assertTrue(errorThrown.await(2, TimeUnit.SECONDS),
+                    "the first target must throw its test Error");
+
+            CompletableFuture<Void> delivered =
+                    injector.enqueue("term_b", "second", TestTurnTokens.inert("term_b")).completion();
+            delivered.get(2, TimeUnit.SECONDS);
+        } finally {
+            poller.stop();
+        }
+    }
+
+    @Test
+    void anAbnormalExitClearsRunningSoStartCreatesANewLoop() throws Exception {
+        StatusPoller poller = new StatusPoller(new AgentControl(new FakeHerdr()), new Injector(new AgentControl(new FakeHerdr())), -1);
+        poller.start();
+        Thread first = threadOf(poller);
+        first.join(2000);
+        assertFalse(runningOf(poller), "an abnormal loop exit must clear running");
+
+        poller.start();
+        Thread restarted = threadOf(poller);
+        try {
+            assertNotSame(first, restarted, "start() must create a new loop after an abnormal exit");
+            restarted.join(2000);
+        } finally {
+            poller.stop();
+        }
+    }
+
+    private static Thread threadOf(StatusPoller poller) throws ReflectiveOperationException {
+        Field field = StatusPoller.class.getDeclaredField("thread");
+        field.setAccessible(true);
+        return (Thread) field.get(poller);
+    }
+
+    private static boolean runningOf(StatusPoller poller) throws ReflectiveOperationException {
+        Field field = StatusPoller.class.getDeclaredField("running");
+        field.setAccessible(true);
+        return field.getBoolean(poller);
+    }
+
+    private static final class ErrorOnceForFirstTarget implements HerdrClient {
+        private final FakeHerdr delegate;
+        private final CountDownLatch errorThrown;
+        private final AtomicBoolean first = new AtomicBoolean(true);
+
+        private ErrorOnceForFirstTarget(FakeHerdr delegate, CountDownLatch errorThrown) {
+            this.delegate = delegate;
+            this.errorThrown = errorThrown;
+        }
+
+        @Override
+        public JsonNode call(String method, Object params) {
+            if (method.equals("agent.get") && params instanceof Map<?, ?> map
+                    && "w2:p7".equals(map.get("target")) && first.compareAndSet(true, false)) {
+                errorThrown.countDown();
+                throw new AssertionError("test Error from the first poll target");
+            }
+            return delegate.call(method, params);
+        }
+
+        @Override
+        public void close() {
+            delegate.close();
+        }
+    }
+}

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

@@ -0,0 +1,91 @@
+package dev.ltms.fleet.session;
+
+import dev.ltms.fleet.config.FleetConfig;
+import dev.ltms.fleet.guard.SubscriptionGuard;
+import dev.ltms.fleet.herdr.AgentControl;
+import dev.ltms.fleet.herdr.FakeHerdr;
+import dev.ltms.fleet.herdr.WorkspaceControl;
+import dev.ltms.fleet.member.ClaudeCodeLauncher;
+import org.junit.jupiter.api.Test;
+
+import java.lang.reflect.Field;
+import java.util.List;
+import java.util.Map;
+import java.util.Set;
+import java.util.concurrent.CountDownLatch;
+import java.util.concurrent.TimeUnit;
+import java.util.concurrent.atomic.AtomicBoolean;
+import java.util.function.LongSupplier;
+
+import static org.junit.jupiter.api.Assertions.assertFalse;
+import static org.junit.jupiter.api.Assertions.assertNotSame;
+import static org.junit.jupiter.api.Assertions.assertTrue;
+
+class SessionReaperResilienceTest {
+
+    @Test
+    void anErrorInOneIterationDoesNotStopTheNextIteration() throws Exception {
+        CountDownLatch errorThrown = new CountDownLatch(1);
+        CountDownLatch nextIteration = new CountDownLatch(1);
+        AtomicBoolean first = new AtomicBoolean(true);
+        LongSupplier clock = () -> {
+            if (first.compareAndSet(true, false)) {
+                errorThrown.countDown();
+                throw new AssertionError("test Error from the first reap iteration");
+            }
+            nextIteration.countDown();
+            return System.nanoTime();
+        };
+        SessionReaper reaper = new SessionReaper(sessionManager(clock), 60, 1);
+        reaper.start();
+        try {
+            assertTrue(errorThrown.await(2, TimeUnit.SECONDS),
+                    "the first reap iteration must throw its test Error");
+            assertTrue(nextIteration.await(2, TimeUnit.SECONDS),
+                    "the reaper must continue to the next iteration after an Error");
+        } finally {
+            reaper.stop();
+        }
+    }
+
+    @Test
+    void anAbnormalExitClearsRunningSoStartCreatesANewLoop() throws Exception {
+        SessionReaper reaper = new SessionReaper(sessionManager(System::nanoTime), 60, -1);
+        reaper.start();
+        Thread first = threadOf(reaper);
+        first.join(2000);
+        assertFalse(runningOf(reaper), "an abnormal loop exit must clear running");
+
+        reaper.start();
+        Thread restarted = threadOf(reaper);
+        try {
+            assertNotSame(first, restarted, "start() must create a new loop after an abnormal exit");
+            restarted.join(2000);
+        } finally {
+            reaper.stop();
+        }
+    }
+
+    private static SessionManager sessionManager(LongSupplier clock) {
+        FakeHerdr herdr = new FakeHerdr();
+        FleetConfig.Profile cfg = new FleetConfig.Profile(
+                "ltms-local", "http://gx00.gw:8000", "coder", null, "FLEETD_WORKER_TOKEN",
+                List.of("ccs", "ltms-local"), "tab", "fleetd-workers",
+                "worker: {profile} #{n}", null, null, null);
+        ClaudeCodeLauncher launcher = new ClaudeCodeLauncher(new AgentControl(herdr), new WorkspaceControl(herdr),
+                new SubscriptionGuard(Set.of("gx00.gw")), Map.of(cfg.profile(), cfg), cfg.profile(), _ -> null);
+        return new SessionManager(launcher, new FakeWorktrees(), clock);
+    }
+
+    private static Thread threadOf(SessionReaper reaper) throws ReflectiveOperationException {
+        Field field = SessionReaper.class.getDeclaredField("thread");
+        field.setAccessible(true);
+        return (Thread) field.get(reaper);
+    }
+
+    private static boolean runningOf(SessionReaper reaper) throws ReflectiveOperationException {
+        Field field = SessionReaper.class.getDeclaredField("running");
+        field.setAccessible(true);
+        return field.getBoolean(reaper);
+    }
+}

scripts/redeploy-fleetd.sh

@@ -96,13 +96,23 @@ SYSTEMD_UNIT='fleetd'
 # any of the prose detail strings and needs no escaping in a `case`/glob pattern.
 SUPERVISOR_DETAIL_SEP=$'\x1f'
 
-# fleetd #492 follow-up: set by systemd_loaded/systemd_installed when the underlying `systemctl`
-# call could not answer cleanly — it exited non-zero AND wrote something to stderr, which is a real
-# tool failure (e.g. it cannot reach the user bus over a non-lingering ssh session), never the same
-# fact as a clean negative answer ("not active", no stderr). Initialized here, not just inside the
-# probes, so detect_supervisor can read them under `set -u` even before either probe has ever run,
-# and so a test that stubs a probe with a plain `return 0`/`return 1` body (leaving these untouched)
-# reads a deterministic 0 rather than whatever a previous probe call left behind.
+# fleetd #492 follow-up, refined by fleetd #545: three states, not two, set by
+# systemd_loaded/systemd_installed —
+#   0 = no error, the probe ran and gave a clean answer.
+#   1 = the probe RAN and answered badly: `systemctl` exited non-zero AND wrote something to
+#       stderr, a real tool failure (e.g. it cannot reach the user bus over a non-lingering ssh
+#       session), never the same fact as a clean negative answer ("not active", no stderr).
+#   2 = the probe could not even be SET UP: the `mktemp` call that makes a place to capture
+#       `systemctl`'s stderr failed before `systemctl` ever ran. This is a fleetd #545 fix: on GNU
+#       coreutils (every Linux distribution) a template with no `X`s made `mktemp` fail every
+#       single time, and the two states were folded into one flag and one message that named
+#       cause 1 ("systemctl exited non-zero and reported an error on stderr") for a failure that
+#       was actually cause 2 — systemctl was never executed at all. One flag with two meanings
+#       needing different messages was the defect; a third value is the fix, not a second flag.
+# Initialized here, not just inside the probes, so detect_supervisor can read them under `set -u`
+# even before either probe has ever run, and so a test that stubs a probe with a plain
+# `return 0`/`return 1` body (leaving these untouched) reads a deterministic 0 rather than whatever
+# a previous probe call left behind.
 SYSTEMD_LOADED_ERRORED=0
 SYSTEMD_INSTALLED_ERRORED=0
 # fleetd #492 follow-up: SUPERVISOR_UNCLEAR_DETAIL is the specific supervisor/reason that
@@ -130,11 +140,35 @@ ok()   { printf '   ok    %s\n' "$*"; }
 warn() { printf '   WARN  %s\n' "$*"; }
 die()  { printf '\n   FAIL  %s\n\n' "$*" >&2; exit 1; }
 
+# fleetd #550 — shasum is macOS-only (it ships with Perl, which Debian/Ubuntu/etc. do not install
+# by default); GNU coreutils (every mainstream Linux distro) ships sha256sum instead and has no
+# shasum at all. Prefer sha256sum, fall back to shasum -a 256 — same idiom as
+# probe-member-credentials.sh's `hasher` selection — and when NEITHER is on PATH, say so plainly.
+# That third answer matters: without it, a missing hasher makes `cut` succeed on empty input, and
+# under `set -o pipefail` the pipeline as a whole still fails, so a caller's own `|| echo "absent"`
+# then reports a file that is right there as though it were gone. `hash256` never does that — it
+# only ever hashes or says it could not.
+hash256() {
+  local f="$1"
+  if command -v sha256sum >/dev/null 2>&1; then
+    sha256sum "$f" | cut -c1-12
+  elif command -v shasum >/dev/null 2>&1; then
+    shasum -a 256 "$f" | cut -c1-12
+  else
+    echo "unhashable"
+  fi
+}
+
 # Reports the hash of $JAR by default, or of whatever path is passed — used to report the STAGED
 # jar right after a build (before it has been swapped in) without ever changing what a bare
 # `jar_id` (no args) means: the live path, $JAR. --check and the final "pid ..., jar ..." line
 # both call it with no args on purpose, so neither can ever be fooled by a leftover staged file.
-jar_id() { local f="${1:-$JAR}"; [ -f "$f" ] && shasum -a 256 "$f" | cut -c1-12 || echo "absent"; }
+# fleetd #550 — THREE distinct answers now, not two: `[ -f "$f" ]` already separates "the jar is
+# not there" (-> "absent") from "the jar is there"; for the second case, hash256 itself separates
+# "hashed it" (a 12-char hex string) from "could not hash it" (-> "unhashable", when no hasher is
+# on PATH). "absent" must never be the answer for a file that exists — that conflation, on Linux,
+# was the whole defect this ticket fixes.
+jar_id() { local f="${1:-$JAR}"; [ -f "$f" ] && hash256 "$f" || echo "absent"; }
 running_pid() { pgrep -f "$PATTERN" || true; }
 
 # fleetd #493 — three small, independently testable pieces of "never build into the path a
@@ -238,12 +272,17 @@ launchd_loaded() { launchctl list "$LAUNCHD_LABEL" >/dev/null 2>&1; }
 # fleetd #492 follow-up: both functions used to throw `systemctl`'s stderr straight into
 # /dev/null, which meant "systemctl answered no" and "systemctl could not answer at all" (e.g. it
 # cannot reach the user bus over a non-lingering ssh session) looked identical — both a plain
-# nonzero exit. They now capture stderr separately and set their own *_ERRORED flag ONLY when the
+# nonzero exit. They now capture stderr separately and set their own *_ERRORED flag to 1 when the
 # call exited non-zero AND wrote something to stderr — a real tool failure, never a clean "not
 # installed"/"not active" answer (which exits non-zero with empty stderr). detect_supervisor reads
 # the flag right after calling the probe, so a probe that could not answer routes to "unclear",
 # never silently becomes "none".
 #
+# fleetd #545: the flag has a third value, 2, set when the `mktemp` call that sets up the probe's
+# own stderr capture fails, before `systemctl` ever runs — see the SYSTEMD_LOADED_ERRORED /
+# SYSTEMD_INSTALLED_ERRORED comment above their initialization for why this is a third value on the
+# same flag, not a second flag.
+#
 # "installed": a unit FILE by this name exists, regardless of its current state — the systemd
 # analogue of the plist file existing on disk. `list-unit-files` reads unit definitions without
 # depending on runtime state, so this stays read-only and safe under --check.
@@ -251,8 +290,8 @@ systemd_installed() {
   SYSTEMD_INSTALLED_ERRORED=0
   command -v systemctl >/dev/null 2>&1 || return 1
   local err_file out rc=0
-  if ! err_file="$(mktemp -t systemd-installed-err)"; then
-    SYSTEMD_INSTALLED_ERRORED=1
+  if ! err_file="$(mktemp -t systemd-installed-err.XXXXXX)"; then
+    SYSTEMD_INSTALLED_ERRORED=2
     return 1
   fi
   out="$(systemctl --user list-unit-files "$SYSTEMD_UNIT.service" --no-legend 2>"$err_file")" || rc=$?
@@ -275,8 +314,8 @@ systemd_loaded() {
   SYSTEMD_LOADED_ERRORED=0
   command -v systemctl >/dev/null 2>&1 || return 1
   local err_file rc=0
-  if ! err_file="$(mktemp -t systemd-loaded-err)"; then
-    SYSTEMD_LOADED_ERRORED=1
+  if ! err_file="$(mktemp -t systemd-loaded-err.XXXXXX)"; then
+    SYSTEMD_LOADED_ERRORED=2
     return 1
   fi
   systemctl --user is-active "$SYSTEMD_UNIT" >/dev/null 2>"$err_file" || rc=$?
@@ -287,6 +326,49 @@ systemd_loaded() {
   return "$rc"
 }
 
+# fleetd #504: the "loaded but not currently running" branches in the main stop step (case
+# launchd/systemd, reached when $OLD_PID is empty) used to run `launchctl unload`/`systemctl --user
+# stop` with `2>/dev/null || true` and then print `ok` unconditionally — the exact conflation
+# systemd_loaded/systemd_installed above already fixed on the READ side (fleetd #492 follow-up): a
+# genuine "already stopped" answer (nonzero exit, nothing on stderr) is harmless, but a real tool
+# failure (nonzero exit WITH a stderr message — e.g. launchd or the systemd user bus is
+# unreachable) is not, and reporting `ok` on THAT means the start step below can register a fresh
+# load on top of a supervisor that never actually let go: the exact two-daemons failure fleetd #492
+# exists to prevent, reached from the one state (already odd) where a false `ok` is least
+# affordable. These two functions apply the same "capture stderr separately, flag only a nonzero
+# exit WITH stderr as a real failure" pattern to the WRITE side. No ${VAR:-default} anywhere here —
+# see the #492 follow-up constraints comment above detect_supervisor for why a default would hide a
+# lost value instead of surfacing it (fleetd #497's defect class).
+unload_launchd_if_loaded() {
+  local err_file rc=0
+  if ! err_file="$(mktemp -t launchd-unload-err.XXXXXX)"; then
+    die "could not create a temp file to capture 'launchctl unload' stderr — cannot tell a real
+        failure from a clean already-unloaded answer, so refusing to guess. The daemon's
+        supervision state was NOT touched."
+  fi
+  launchctl unload -w "$LAUNCHD_PLIST" >/dev/null 2>"$err_file" || rc=$?
+  if [ "$rc" -ne 0 ] && [ -s "$err_file" ]; then
+    die "'launchctl unload -w $LAUNCHD_PLIST' failed: $(cat "$err_file")
+        The daemon may still be under supervision; investigate before retrying."
+  fi
+  rm -f "$err_file"
+}
+
+stop_systemd_if_loaded() {
+  local err_file rc=0
+  if ! err_file="$(mktemp -t systemd-stop-err.XXXXXX)"; then
+    die "could not create a temp file to capture 'systemctl --user stop' stderr — cannot tell a
+        real failure from a clean already-stopped answer, so refusing to guess. The daemon's
+        supervision state was NOT touched."
+  fi
+  systemctl --user stop "$SYSTEMD_UNIT" >/dev/null 2>"$err_file" || rc=$?
+  if [ "$rc" -ne 0 ] && [ -s "$err_file" ]; then
+    die "'systemctl --user stop $SYSTEMD_UNIT' failed: $(cat "$err_file")
+        The daemon may still be under supervision; investigate before retrying."
+  fi
+  rm -f "$err_file"
+}
+
 # fleetd #492: three real answers, not two — launchd, systemd, or genuinely unsupervised — plus a
 # fourth, "ambiguous", for the one case this script cannot tell apart: both signals firing at once.
 # That is exactly "I cannot tell who supervises this process", and guessing wrong here is how two
@@ -306,6 +388,14 @@ systemd_loaded() {
 # "none" now means only: neither supervisor is installed, neither is loaded, and neither probe
 # errored.
 #
+# fleetd #545: *_ERRORED carries a THIRD state (2 = the probe's own mktemp setup failed, before
+# `systemctl` ever ran — see the flag's own comment above its initialization), and it must never be
+# reported with the same detail text as state 1 (`systemctl` ran and answered badly on stderr). The
+# two are different facts about different failures, and conflating them makes the "unclear" message
+# assert a cause ("systemctl exited non-zero and reported an error on stderr") that was never
+# measured when the real cause was state 2. detect_supervisor below picks the detail text off the
+# flag's value, not off a single "errored at all" boolean.
+#
 # fleetd #492 follow-up — constraints every caller of this function depends on (learned the hard
 # way: an earlier version of this fix set a SUPERVISOR_UNCLEAR_DETAIL global from inside here and
 # it was silently lost, because every real call site invokes this as `$(detect_supervisor)`):
@@ -339,7 +429,10 @@ detect_supervisor() {
   launchd_installed && li=1
   systemd_installed && si=1
 
-  if [ "$SYSTEMD_LOADED_ERRORED" = 1 ] || [ "$SYSTEMD_INSTALLED_ERRORED" = 1 ]; then
+  if [ "$SYSTEMD_LOADED_ERRORED" = 2 ] || [ "$SYSTEMD_INSTALLED_ERRORED" = 2 ]; then
+    detail="the systemd --user probe for '$SYSTEMD_UNIT' could not even be set up (a temp file to capture systemctl's stderr could not be created) — systemctl was never run, so this says nothing about systemd, the user bus, or the unit itself"
+    kind="unclear"
+  elif [ "$SYSTEMD_LOADED_ERRORED" = 1 ] || [ "$SYSTEMD_INSTALLED_ERRORED" = 1 ]; then
     detail="the systemd --user probe for '$SYSTEMD_UNIT' could not answer cleanly (systemctl exited non-zero and reported an error on stderr, not a clean negative — e.g. it cannot reach the user bus)"
     kind="unclear"
   elif [ "$ld" = 1 ] && [ "$sd" = 1 ]; then
@@ -787,7 +880,7 @@ if [ "$DO_BUILD" = 1 ]; then
   # fleetd #493: wipe a leftover staged jar from a previous failed/interrupted run BEFORE doing
   # anything else, so that run's leftovers can never be mistaken for this run's output.
   rm -f "$JAR_STAGED"
-  BUILD_LOG="$(mktemp -t fleetd-build)"
+  BUILD_LOG="$(mktemp -t fleetd-build.XXXXXX)"
   echo "   log: $BUILD_LOG"
   if ! mvn -f "$MODULE/pom.xml" clean install > "$BUILD_LOG" 2>&1; then
     grep -E 'ERROR|BUILD FAILURE|Tests run:.*Failures: [1-9]|Tests run:.*Errors: [1-9]' "$BUILD_LOG" \
@@ -881,16 +974,20 @@ if [ -n "$OLD_PID" ]; then
 elif [ "$SUPERVISOR_KIND" = "launchd" ]; then
   # Loaded but not currently running (e.g. throttled after a crash loop). Unload it anyway so the
   # start step below does a clean load, never a load stacked on an already-loaded label.
+  # fleetd #504: unload_launchd_if_loaded (above) tolerates a genuine already-unloaded answer but
+  # dies on a real `launchctl` failure — never a bare `|| true` that would print `ok` either way.
   say "stop"
   RESTART_MARK="$(wc -l < "$OUT" 2>/dev/null || echo 0)"
-  launchctl unload -w "$LAUNCHD_PLIST" 2>/dev/null || true
+  unload_launchd_if_loaded
   ok "launchd agent unloaded (was already not running)"
 elif [ "$SUPERVISOR_KIND" = "systemd" ]; then
   # Same case for systemd: the unit is known/active-capable but not currently running. `stop` on an
   # already-stopped unit is a harmless no-op — kept for symmetry with the launchd branch above.
+  # fleetd #504: stop_systemd_if_loaded (above) tolerates that genuine no-op but dies on a real
+  # `systemctl` failure — never a bare `|| true` that would print `ok` either way.
   say "stop"
   RESTART_MARK="$(wc -l < "$OUT" 2>/dev/null || echo 0)"
-  systemctl --user stop "$SYSTEMD_UNIT" 2>/dev/null || true
+  stop_systemd_if_loaded
   ok "systemd --user unit stopped (was already not running)"
 else
   RESTART_MARK="$(wc -l < "$OUT" 2>/dev/null || echo 0)"
@@ -1013,7 +1110,7 @@ tail -n "+$((RESTART_MARK + 1))" "$OUT" 2>/dev/null \
 
 # Errors since the restart, anchored to the marker so old noise cannot leak in. Keep the fresh
 # region in a file because the classifier must preserve the order of errors and recoveries.
-FRESH_LOG="$(mktemp -t fleetd-fresh-log)"
+FRESH_LOG="$(mktemp -t fleetd-fresh-log.XXXXXX)"
 trap 'rm -f "$FRESH_LOG"' EXIT
 tail -n "+$((RESTART_MARK + 1))" "$OUT" > "$FRESH_LOG" 2>/dev/null || true
 classify_amqp_connection_errors "$FRESH_LOG"

scripts/test-redeploy-fleetd.sh

@@ -128,8 +128,112 @@ STUB
   launchd_loaded()    { return 1; }
   result="$(PATH="$bin_dir:$PATH" detect_supervisor)"
   assert_equals "unclear" "$(supervisor_kind_of "$result")" "a systemd probe error must read as unclear, not none"
-  printf '%s' "$(supervisor_detail_of "$result")" | grep -qF "$SYSTEMD_UNIT" \
+  local detail
+  detail="$(supervisor_detail_of "$result")"
+  printf '%s' "$detail" | grep -qF "$SYSTEMD_UNIT" \
     || fail "detail does not name the systemd unit whose probe errored"
+  # fleetd #545: this is the PROBE-RAN-AND-ANSWERED-BADLY case (systemctl actually executed and
+  # wrote to stderr) — it must carry that story and never the SET-UP-FAILED story (mktemp never
+  # even ran here), or the two "unclear" causes have collapsed back into one message that asserts a
+  # cause it did not measure, which is the exact defect this ticket exists to fix.
+  printf '%s' "$detail" | grep -qF "systemctl exited non-zero and reported an error on stderr" \
+    || fail "detail does not say systemctl ran and answered with stderr: $detail"
+  printf '%s' "$detail" | grep -qF "could not even be set up" \
+    && fail "detail wrongly claims the probe could not be set up, but systemctl actually ran and answered on stderr: $detail"
+  return 0
+}
+
+# fleetd #545 — the companion case to the probe-error test above: here `mktemp` itself fails
+# (whatever the reason — the historical bug was a GNU-mktemp-rejects-a-template-with-no-Xs case,
+# but this stub simulates ANY reason the probe's own stderr-capture temp file cannot be created,
+# e.g. a full or unwritable temp dir) and `systemctl` is never invoked at all. Before this ticket,
+# this collapsed into the SAME "systemctl exited non-zero and reported an error on stderr" detail
+# as the sibling test above, which asserts a cause (systemctl ran and answered badly) that was
+# never measured, because systemctl never ran. This proves the SET-UP-FAILED detail is distinct and
+# does not claim systemctl said anything.
+test_detect_supervisor_systemd_probe_setup_failure_is_unclear() {
+  # Re-source first for the same reason test_detect_supervisor_systemd_probe_error_is_unclear does:
+  # restore the REAL probe bodies before driving them through a stub PATH.
+  source "$ROOT/scripts/redeploy-fleetd.sh"
+  local bin_dir result rc=0
+  bin_dir="$TMP/stub-bin-mktemp-fails"
+  mkdir -p "$bin_dir"
+  # A systemctl stub that would fail loudly if it were ever actually invoked — proves the mktemp
+  # failure short-circuits the probe before systemctl runs, not merely that this test forgot to
+  # supply a working systemctl.
+  cat > "$bin_dir/systemctl" <<'STUB'
+#!/usr/bin/env bash
+echo "systemctl must never run when mktemp already failed" >&2
+exit 1
+STUB
+  chmod +x "$bin_dir/systemctl"
+  cat > "$bin_dir/mktemp" <<'STUB'
+#!/usr/bin/env bash
+echo "mktemp: cannot create temp file" >&2
+exit 1
+STUB
+  chmod +x "$bin_dir/mktemp"
+
+  PATH="$bin_dir:$PATH" systemd_loaded && rc=0 || rc=$?
+  [ "$rc" -ne 0 ] \
+    || fail "systemd_loaded must not report loaded=true when its own mktemp setup failed"
+  assert_equals "2" "$SYSTEMD_LOADED_ERRORED" \
+    "systemd_loaded must flag a SETUP failure (2), distinct from a probe-answered-with-stderr failure (1)"
+
+  launchd_installed() { return 1; }
+  launchd_loaded()    { return 1; }
+  result="$(PATH="$bin_dir:$PATH" detect_supervisor)"
+  assert_equals "unclear" "$(supervisor_kind_of "$result")" "a systemd probe setup failure must read as unclear, not none"
+  local detail
+  detail="$(supervisor_detail_of "$result")"
+  printf '%s' "$detail" | grep -qF "could not even be set up" \
+    || fail "detail does not say the probe could not be SET UP: $detail"
+  printf '%s' "$detail" | grep -qF "systemctl exited non-zero and reported an error on stderr" \
+    && fail "detail wrongly asserts systemctl exited non-zero and reported an error on stderr, but systemctl was never run: $detail"
+  return 0
+}
+
+# fleetd #545 — source-text check: every `mktemp -t` template in redeploy-fleetd.sh must contain an
+# `X` placeholder. BSD mktemp (macOS) tolerates a bare template with no `X`s and just appends its
+# own random suffix, which is exactly why six such sites survived undetected here — GNU mktemp
+# (every Linux distribution) refuses a template with fewer than three `X`s and exits non-zero. There
+# is no BSD-vs-GNU seam to stub on this Mac, so this is a source-text check rather than a
+# behavioural one, the same shape as test_refuse_drain_gate_call_site_present above. Anchored on
+# `mktemp -t ` (with the trailing space) so it inspects only the `-t`-style templates this ticket is
+# about, never the `mktemp -d` calls this file and test-probe-member-credentials.sh already use
+# (both already carry their own `XXXXXX` and are a different mktemp mode entirely).
+test_mktemp_dash_t_templates_have_x_placeholders() {
+  local src="$ROOT/scripts/redeploy-fleetd.sh" bad
+  bad="$(grep -n 'mktemp -t ' "$src" | grep -v 'XXX' || true)"
+  [ -z "$bad" ] \
+    || fail "mktemp -t template(s) with no X placeholder (fails under GNU coreutils): $bad"
+}
+
+# fleetd #550 — the shape, not the named lines: #545 showed the exact same failure mode (a
+# macOS-only idiom used with no portable fallback) spread from two sites to six across 91 commits
+# before anyone tested the SHAPE rather than specific lines. This is the shasum sibling: any script
+# under scripts/ that actually INVOKES the macOS-only hasher to compute a hash (as opposed to
+# merely probing whether it exists with `command -v`, or mentioning it in prose) must also check
+# for the portable one first in that same file — the prefer-portable-fall-back-to-macOS-only idiom
+# probe-member-credentials.sh:273-279 and this ticket's own hash256 helper both follow.
+#
+# The needle is built from two concatenated pieces, deliberately never written as one literal
+# string in this file: written whole, it would match THIS CHECK'S OWN source line once the loop
+# below reaches this very file, and the check would then "pass" by matching itself rather than any
+# real invocation elsewhere — a zero-findings result indistinguishable from a clean file.
+test_no_unguarded_macos_only_hasher_calls() {
+  local needle f bad="" usage
+  needle='shasum'; needle="$needle -a"
+  for f in "$ROOT"/scripts/*.sh; do
+    [ -f "$f" ] || continue
+    usage="$(grep -Fn "$needle" "$f" || true)"
+    if [ -n "$usage" ]; then
+      grep -q 'command -v sha256sum' "$f" \
+        || bad="$bad$(basename "$f") "
+    fi
+  done
+  [ -z "$bad" ] \
+    || fail "script(s) invoke the macOS-only hasher with no portable-hasher-first fallback guard in the same file: $bad"
 }
 
 # fleetd #492 follow-up (Item 1): this must go through the REAL call-site shape at :437-440, not a
@@ -218,8 +322,11 @@ test_jar_id_defaults_to_live_and_reports_explicit_path() {
   JAR="$dir/fleetd.jar"; JAR_STAGED="$dir/fleetd-new.jar"
   printf 'live jar bytes' > "$JAR"
   printf 'staged jar bytes, not the same content' > "$JAR_STAGED"
-  live_hash="$(shasum -a 256 "$JAR" | cut -c1-12)"
-  staged_hash="$(shasum -a 256 "$JAR_STAGED" | cut -c1-12)"
+  # fleetd #550: this reference hash must be computed the same portable way jar_id() itself now
+  # computes one — a bare, unguarded call to the macOS-only hasher here was exactly the item-2
+  # defect, dying with "command not found" on any Linux runner that has no such hasher at all.
+  live_hash="$(hash256 "$JAR")"
+  staged_hash="$(hash256 "$JAR_STAGED")"
   default_result="$(jar_id)"
   explicit_result="$(jar_id "$JAR_STAGED")"
   JAR="$saved_jar"; JAR_STAGED="$saved_staged"
@@ -228,6 +335,26 @@ test_jar_id_defaults_to_live_and_reports_explicit_path() {
   assert_equals "$staged_hash" "$explicit_result" "jar_id \"\$JAR_STAGED\" must report the hash of the staged jar, not fall back to \$JAR"
 }
 
+# fleetd #550 — closes a gap the test above leaves open. That test's own reference hash is now ALSO
+# computed by calling hash256 (needed for item 2: the old bare macOS-only-hasher call there was the
+# Linux crash), so its subject (jar_id, via hash256) and its reference (also hash256) share one
+# instrument — they agree no matter which algorithm hash256 actually runs, so a mutation that swaps
+# BOTH of hash256's arms for the wrong algorithm is invisible to it. This test's expected value
+# comes from neither hasher: it is the published SHA-256 test vector for the 3-byte input "abc"
+# (no trailing newline), written here as a literal constant, so it can still tell "hashed
+# correctly" from "hashed, just with the wrong algorithm" — which is what this whole ticket is
+# about.
+test_hash256_computes_a_real_sha256() {
+  local dir f result
+  dir="$TMP/hash256-known-vector"; mkdir -p "$dir"
+  f="$dir/abc.txt"
+  printf 'abc' > "$f"
+  result="$(hash256 "$f")"
+  # SHA-256("abc") = ba7816bf8f01cfea414140de5dae2223b00361a396177a9cb410ff61f20015ad, the standard
+  # FIPS 180 test vector — first 12 hex chars, matching hash256's own `cut -c1-12`.
+  assert_equals "ba7816bf8f01" "$result" "hash256 of the literal 3-byte input 'abc' must be the known SHA-256 prefix, not some other algorithm's"
+}
+
 # fleetd #517 — jar_id()'s "absent" branch was unpinned by any test: the existing test above (#511)
 # proves both halves of the present-file contract but never exercises the missing-file path. This
 # word matters more than a string usually would: "absent" is the #413 signal that a `mvn clean`
@@ -246,6 +373,37 @@ test_jar_id_reports_absent_for_missing_file() {
   assert_equals "absent" "$explicit_result" "jar_id with an explicit missing path must report absent"
 }
 
+# fleetd #550 — the whole point of this ticket: a jar that IS there but could not be hashed must
+# never read the same as a jar that is not there at all. Drives the REAL hash256/jar_id bodies
+# (never stubbed) through a stub PATH that contains neither of the two hashers this script knows —
+# same technique test_detect_supervisor_systemd_probe_setup_failure_is_unclear uses for `mktemp`,
+# except here the stub directory is used to REPLACE PATH rather than prepend to it, because the
+# point is to make BOTH hashers unreachable, not to intercept one specific command while leaving
+# everything else on the real PATH reachable. `[ -f ... ]` and the shell's own `command`/`echo`
+# builtins need no PATH at all, so this is safe even with PATH reduced to an empty directory.
+test_jar_id_reports_unhashable_when_no_hasher_on_path() {
+  local bin_dir dir saved_jar="$JAR" default_result explicit_result
+  bin_dir="$TMP/stub-bin-no-hasher"; mkdir -p "$bin_dir"
+  dir="$TMP/jar-id-no-hasher"; mkdir -p "$dir"
+  JAR="$dir/fleetd.jar"
+  printf 'a real jar that exists but nothing here can hash' > "$JAR"
+  [ -f "$JAR" ] || fail "test fixture error: \$JAR does not exist at $JAR"
+  default_result="$(PATH="$bin_dir" jar_id)"
+  explicit_result="$(PATH="$bin_dir" jar_id "$JAR")"
+  JAR="$saved_jar"
+  [ "$default_result" != "absent" ] \
+    || fail "jar_id reported absent for a file that exists, only because no hasher was on PATH"
+  [ "$explicit_result" != "absent" ] \
+    || fail "jar_id (explicit path) reported absent for a file that exists, only because no hasher was on PATH"
+  # A 12-char hex hash is the OTHER wrong answer here: with no hasher at all, nothing could have
+  # produced one, so a value that merely happens to look like one would mean the stub failed to
+  # hide the real hashers rather than that jar_id degraded correctly.
+  printf '%s' "$default_result" | grep -Eq '^[0-9a-f]{12}$' \
+    && fail "test fixture error: PATH stub did not actually hide the real hasher(s) — got what looks like a real hash"
+  assert_equals "unhashable" "$default_result" "jar_id with no hasher on PATH must report a third, distinct state — never absent, never a hash"
+  assert_equals "unhashable" "$explicit_result" "jar_id (explicit path) with no hasher on PATH must report the same third state"
+}
+
 # fleetd #493 — never build into the path a running process holds. stage_built_jar/swap_staged_jar
 # are exercised directly against real files on disk (not stubs), because the whole point is file
 # behavior (does the content move, does the source disappear, does a failure leave both sides
@@ -621,6 +779,110 @@ test_refuse_drain_gate_call_site_present() {
     || fail "could not find the main flow's refuse_drain_gate call site in redeploy-fleetd.sh"
 }
 
+# fleetd #504 — the "loaded but not currently running" branches for launchd/systemd used to run
+# `launchctl unload`/`systemctl --user stop` with `2>/dev/null || true` and print `ok`
+# unconditionally, so a real supervisor failure (e.g. it cannot reach launchd/the systemd user bus)
+# read exactly like a harmless already-stopped answer. unload_launchd_if_loaded/
+# stop_systemd_if_loaded (redeploy-fleetd.sh, right after systemd_loaded) apply systemd_loaded's own
+# "capture stderr separately — only a non-zero exit WITH stderr is a real failure" pattern to the
+# WRITE side. Both call the real `launchctl`/`systemctl` binaries directly (they are not overridable
+# wrapper functions the way launchd_loaded/systemd_loaded are), so these tests put a stub binary
+# first on PATH — the same technique test_detect_supervisor_systemd_probe_error_is_unclear above
+# already uses for `systemctl`.
+test_unload_launchd_if_loaded_dies_on_real_failure() {
+  local bin_dir output rc=0 saved_plist="$LAUNCHD_PLIST"
+  bin_dir="$TMP/stub-bin-launchctl-error"; mkdir -p "$bin_dir"
+  cat > "$bin_dir/launchctl" <<'STUB'
+#!/usr/bin/env bash
+echo "Could not find specified service" >&2
+exit 1
+STUB
+  chmod +x "$bin_dir/launchctl"
+  LAUNCHD_PLIST="$TMP/fake-fail.plist"
+  output="$(PATH="$bin_dir:$PATH" unload_launchd_if_loaded 2>&1)" || rc=$?
+  LAUNCHD_PLIST="$saved_plist"
+  [ "$rc" -ne 0 ] \
+    || fail "unload_launchd_if_loaded must die when launchctl exits non-zero AND writes to stderr"
+  printf '%s' "$output" | grep -qF 'launchctl unload' \
+    || fail "die message does not name the failing launchctl unload command"
+}
+
+# Captured via $(...) rather than called bare: unload_launchd_if_loaded's own die() does a hard
+# `exit`, and calling it directly at this level would let a regression that makes it die on this
+# clean-negative case kill the WHOLE suite before the `|| fail` below ever ran — printing die's own
+# message instead of this test's. Inside a command substitution, that `exit` only ends the subshell
+# (a-guard-is-defeated-by-its-calling-context: the same reason the *_dies_on_real_failure tests
+# above capture this way), so this test's own message is what actually reaches the report.
+test_unload_launchd_if_loaded_tolerates_clean_negative() {
+  local bin_dir saved_plist="$LAUNCHD_PLIST" output rc=0
+  bin_dir="$TMP/stub-bin-launchctl-noop"; mkdir -p "$bin_dir"
+  cat > "$bin_dir/launchctl" <<'STUB'
+#!/usr/bin/env bash
+exit 1
+STUB
+  chmod +x "$bin_dir/launchctl"
+  LAUNCHD_PLIST="$TMP/fake-noop.plist"
+  output="$(PATH="$bin_dir:$PATH" unload_launchd_if_loaded 2>&1)" || rc=$?
+  LAUNCHD_PLIST="$saved_plist"
+  [ "$rc" -eq 0 ] \
+    || fail "unload_launchd_if_loaded must tolerate a clean already-unloaded answer (non-zero exit, empty stderr): $output"
+}
+
+test_stop_systemd_if_loaded_dies_on_real_failure() {
+  local bin_dir output rc=0
+  bin_dir="$TMP/stub-bin-systemctl-stop-error"; mkdir -p "$bin_dir"
+  cat > "$bin_dir/systemctl" <<'STUB'
+#!/usr/bin/env bash
+echo "Failed to connect to bus: No such file or directory" >&2
+exit 1
+STUB
+  chmod +x "$bin_dir/systemctl"
+  output="$(PATH="$bin_dir:$PATH" stop_systemd_if_loaded 2>&1)" || rc=$?
+  [ "$rc" -ne 0 ] \
+    || fail "stop_systemd_if_loaded must die when systemctl exits non-zero AND writes to stderr"
+  printf '%s' "$output" | grep -qF 'systemctl --user stop' \
+    || fail "die message does not name the failing systemctl --user stop command"
+}
+
+# Same subshell-capture reasoning as test_unload_launchd_if_loaded_tolerates_clean_negative above:
+# stop_systemd_if_loaded's own die() does a hard `exit`, so this must run inside $(...) or a
+# regression here would kill the whole suite with die's message instead of this test's.
+test_stop_systemd_if_loaded_tolerates_clean_negative() {
+  local bin_dir output rc=0
+  bin_dir="$TMP/stub-bin-systemctl-stop-noop"; mkdir -p "$bin_dir"
+  cat > "$bin_dir/systemctl" <<'STUB'
+#!/usr/bin/env bash
+exit 1
+STUB
+  chmod +x "$bin_dir/systemctl"
+  output="$(PATH="$bin_dir:$PATH" stop_systemd_if_loaded 2>&1)" || rc=$?
+  [ "$rc" -eq 0 ] \
+    || fail "stop_systemd_if_loaded must tolerate a clean already-stopped answer (non-zero exit, empty stderr): $output"
+}
+
+# Closes the same gap test_refuse_drain_gate_call_site_present closes for the drain gate: the four
+# tests above call unload_launchd_if_loaded/stop_systemd_if_loaded directly, and sourcing stops
+# before the main flow ever runs (the SOURCED guard), so none of them can prove the main flow still
+# CALLS these two functions instead of the original bare `2>/dev/null || true`. A source-text check,
+# like test_swap_ordered_after_wait_and_before_start. The call-site needle is anchored (`^  name$`)
+# so it cannot be satisfied by the comment lines above each call site that merely mention the
+# function by name.
+test_stop_branches_call_tolerant_helpers_not_bare_or_true() {
+  local src="$ROOT/scripts/redeploy-fleetd.sh" unload_call_line stop_call_line
+  unload_call_line="$(grep -n '^  unload_launchd_if_loaded$' "$src" | head -1 | cut -d: -f1 || true)"
+  stop_call_line="$(grep -n '^  stop_systemd_if_loaded$' "$src" | head -1 | cut -d: -f1 || true)"
+  [ -n "$unload_call_line" ] \
+    || fail "could not find the main flow's call to unload_launchd_if_loaded in redeploy-fleetd.sh"
+  [ -n "$stop_call_line" ] \
+    || fail "could not find the main flow's call to stop_systemd_if_loaded in redeploy-fleetd.sh"
+  if grep -qF 'launchctl unload -w "$LAUNCHD_PLIST" 2>/dev/null || true' "$src"; then
+    fail "the bare 'launchctl unload ... 2>/dev/null || true' defect (fleetd #504) is back in redeploy-fleetd.sh"
+  fi
+  if grep -qF 'systemctl --user stop "$SYSTEMD_UNIT" 2>/dev/null || true' "$src"; then
+    fail "the bare 'systemctl --user stop ... 2>/dev/null || true' defect (fleetd #504) is back in redeploy-fleetd.sh"
+  fi
+}
+
 test_no_errors() {
   cat > "$TMP/no-errors.log" <<'LOG'
 2026-09-05 12:00:00 INFO fleetd listening
@@ -984,6 +1246,9 @@ test_detect_supervisor_none
 test_detect_supervisor_systemd_installed_not_loaded_is_unclear
 test_detect_supervisor_launchd_installed_not_loaded_is_unclear
 test_detect_supervisor_systemd_probe_error_is_unclear
+test_detect_supervisor_systemd_probe_setup_failure_is_unclear
+test_mktemp_dash_t_templates_have_x_placeholders
+test_no_unguarded_macos_only_hasher_calls
 test_require_drivable_supervisor_refuses_ambiguous
 test_require_drivable_supervisor_refuses_unclear
 test_require_drivable_supervisor_accepts_known_kinds
@@ -991,7 +1256,9 @@ test_count_daemon_pids
 test_assert_single_daemon_accepts_one_pid
 test_assert_single_daemon_rejects_two_pids
 test_jar_id_defaults_to_live_and_reports_explicit_path
+test_hash256_computes_a_real_sha256
 test_jar_id_reports_absent_for_missing_file
+test_jar_id_reports_unhashable_when_no_hasher_on_path
 test_stage_built_jar_moves_off_live_path
 test_stage_built_jar_dies_when_build_produced_nothing
 test_swap_staged_jar_moves_staged_onto_live
@@ -1016,6 +1283,11 @@ test_refuse_drain_gate_build_ran_staged_absent
 test_refuse_drain_gate_no_build_staged_present
 test_refuse_drain_gate_no_build_staged_absent
 test_refuse_drain_gate_call_site_present
+test_unload_launchd_if_loaded_dies_on_real_failure
+test_unload_launchd_if_loaded_tolerates_clean_negative
+test_stop_systemd_if_loaded_dies_on_real_failure
+test_stop_systemd_if_loaded_tolerates_clean_negative
+test_stop_branches_call_tolerant_helpers_not_bare_or_true
 test_no_errors
 test_recovery_patterns_match_source
 test_attributed_recovered_connection_error