fleetd #504 item 1: stop swallowing real launchctl/systemctl failures on the 'loaded but not running' path

The two 'loaded but not currently running' branches in the stop step (launchd/systemd, reached
when $OLD_PID is empty) ran 'launchctl unload'/'systemctl --user stop' with '2>/dev/null || true'
and printed 'ok' unconditionally. That swallowed a real supervisor failure (e.g. launchd or the
systemd user bus unreachable) exactly like a harmless already-stopped answer, and let the script
proceed to start a new daemon believing nothing was loaded -- the two-daemons failure fleetd #492
exists to prevent.

Adds unload_launchd_if_loaded/stop_systemd_if_loaded, applying systemd_loaded's own pattern
(capture stderr separately; a non-zero exit WITH stderr is a real failure, a non-zero exit with
empty stderr is a clean already-stopped answer) to the write side. The two call sites now use
these functions instead of the bare '|| true'.

Adds 5 tests: dies-on-real-failure and tolerates-clean-negative for each function, plus a
source-text check that the main flow calls the new functions instead of the original bare
'2>/dev/null || true'. All 5 verified by mutation (reintroducing the swallow, and separately
over-correcting to die unconditionally) -- each goes red with its own message, restores
byte-identical (full sha256), and passes a green control.

CLAUDE.md

@@ -96,8 +96,10 @@ below are the procedure — run them in order, every task, not only the big ones
    that answers it. **A worker's ask waits ~55 seconds, and no nudge makes that longer** — so never
    brief a worker to "ask me". Decide before you delegate, or give it an explicit default.
 6. **Verify yourself.** Re-run the build and the checks. A worker cannot run your IDE tooling, any
-   forge tools it appears to have hold a blocked credential and fail, and a piped command
-   (`… | tail`) hides failures behind a zero exit — never promote a worker's "clean" to a fact.
+   forge MCP server it appears to have holds a blocked credential and fails every call, and a piped
+   command (`… | tail`) hides failures behind a zero exit — never promote a worker's "clean" to a
+   fact. Its injected repo-scoped `GITEA_TOKEN` is a different credential and does work, so a worker
+   reporting that it opened its own PR is reporting something it really can do.
 7. **Review — fan out.** Spawn reviewers against the diff, one per dimension or per file, with
    `wait:false`. Never the implementer of the scope it reviews, and brief them from the diff — not
    from the implementer's rationale, which carries its own blind spot. Dispatch each PR's reviewers
@@ -200,8 +202,11 @@ simply complies has thrown away the reason there are two of you.
    assume them.** What you mount depends on your backend: an opencode member gets the bridge and
    nothing else, while a Claude Code member also inherits the operator's user-scope MCP servers,
    which the bridge never chose for you. Two rules follow. The primary's IDE tooling is still not
-   yours, whatever you see. And **a mounted tool is not a working tool** — the forge server you may
-   find there holds a deliberately blocked credential and fails every call, by design.
+   yours, whatever you see. And **a mounted tool is not a working tool** — the forge MCP server you
+   may find there holds a deliberately blocked credential and fails every call, by design. That is
+   not your only forge route, and the two must not be confused: the repo-scoped `GITEA_TOKEN` the
+   daemon injects into your environment does work, and using it to open your own PR is part of the
+   job. A blocked MCP tool is never a reason to skip that step.
 6. **Never merge.** Stage files explicitly — never `git add -A` — and leave alone anything the
    project marks as not-yours-to-commit.
 

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

@@ -1,15 +1,12 @@
 package dev.ltms.fleet;
 
 import ch.qos.logback.classic.Level;
-import ch.qos.logback.classic.Logger;
 import ch.qos.logback.classic.spi.ILoggingEvent;
-import ch.qos.logback.core.read.ListAppender;
 import dev.ltms.fleet.config.FleetConfig;
 import dev.ltms.fleet.msg.LeadMailbox;
+import dev.ltms.fleet.testing.CapturedLog;
 import org.junit.jupiter.api.Test;
-import org.slf4j.LoggerFactory;
 
-import java.util.List;
 import java.util.Map;
 
 import static org.junit.jupiter.api.Assertions.*;
@@ -52,29 +49,22 @@ class FleetdLeadMailboxSelectionTest {
         }
     }
 
-    private static ListAppender<ILoggingEvent> captureFleetdLogs() {
-        Logger logger = (Logger) LoggerFactory.getLogger(Fleetd.class);
-        ListAppender<ILoggingEvent> appender = new ListAppender<>();
-        appender.start();
-        logger.addAppender(appender);
-        return appender;
-    }
-
-    private static String joined(ListAppender<ILoggingEvent> appender, Level level) {
-        return appender.list.stream().filter(e -> e.getLevel() == level)
+    private static String joined(CapturedLog captured, Level level) {
+        return captured.events().stream().filter(e -> e.getLevel() == level)
                 .map(ILoggingEvent::getFormattedMessage).reduce("", (a, b) -> a + "\n" + b);
     }
 
     @Test
     void noCoordinatorBlockLeavesTheFeatureOffSilently() {
-        var appender = captureFleetdLogs();
-        var opener = new RecordingOpener();
+        try (var captured = CapturedLog.of(Fleetd.class)) {
+            var opener = new RecordingOpener();
 
-        assertNull(Fleetd.openLeadMailbox(null, Map.of(), opener));
+            assertNull(Fleetd.openLeadMailbox(null, Map.of(), opener));
 
-        assertNull(opener.offeredUri, "nothing configured means nothing is opened");
-        assertEquals("", joined(appender, Level.WARN),
-                "an opt-in feature nobody asked for must not warn on every boot");
+            assertNull(opener.offeredUri, "nothing configured means nothing is opened");
+            assertEquals("", joined(captured, Level.WARN),
+                    "an opt-in feature nobody asked for must not warn on every boot");
+        }
     }
 
     @Test
@@ -114,31 +104,33 @@ class FleetdLeadMailboxSelectionTest {
 
     @Test
     void warnsAndStaysOffWhenSelfIdIsMissing() {
-        var appender = captureFleetdLogs();
-        var opener = new RecordingOpener();
-        var coordinator = new FleetConfig.Coordinator(RESOLVED_URI, null, null, null, null);
+        try (var captured = CapturedLog.of(Fleetd.class)) {
+            var opener = new RecordingOpener();
+            var coordinator = new FleetConfig.Coordinator(RESOLVED_URI, null, null, null, null);
 
-        assertNull(Fleetd.openLeadMailbox(coordinator, Map.of(), opener));
+            assertNull(Fleetd.openLeadMailbox(coordinator, Map.of(), opener));
 
-        assertNull(opener.offeredUri, "a mailbox with no owning coord-id has no queue to declare");
-        String warns = joined(appender, Level.WARN);
-        assertTrue(warns.contains("coordinator.selfId"), () -> "say which key is missing: " + warns);
-        assertFalse(warns.contains(SECRET), () -> "the URI's password must never be logged: " + warns);
+            assertNull(opener.offeredUri, "a mailbox with no owning coord-id has no queue to declare");
+            String warns = joined(captured, Level.WARN);
+            assertTrue(warns.contains("coordinator.selfId"), () -> "say which key is missing: " + warns);
+            assertFalse(warns.contains(SECRET), () -> "the URI's password must never be logged: " + warns);
+        }
     }
 
     @Test
     void warnsAndStaysOffWhenTheBrokerIsUnreachableAtBoot() {
-        var appender = captureFleetdLogs();
-        var opener = new RecordingOpener();
-        opener.unreachable = true;
-        var coordinator = new FleetConfig.Coordinator(RESOLVED_URI, null, "mac-opus", null, null);
+        try (var captured = CapturedLog.of(Fleetd.class)) {
+            var opener = new RecordingOpener();
+            opener.unreachable = true;
+            var coordinator = new FleetConfig.Coordinator(RESOLVED_URI, null, "mac-opus", null, null);
 
-        assertNull(Fleetd.openLeadMailbox(coordinator, Map.of(), opener),
-                "a down coordination broker turns the feature off; it must never take the daemon down");
+            assertNull(Fleetd.openLeadMailbox(coordinator, Map.of(), opener),
+                    "a down coordination broker turns the feature off; it must never take the daemon down");
 
-        String warns = joined(appender, Level.WARN);
-        assertTrue(warns.contains("coord.example"), () -> "name the host that failed: " + warns);
-        assertFalse(warns.contains(SECRET), () -> "with credentials stripped: " + warns);
-        assertTrue(warns.contains("Connection refused"), () -> "and the real reason: " + warns);
+            String warns = joined(captured, Level.WARN);
+            assertTrue(warns.contains("coord.example"), () -> "name the host that failed: " + warns);
+            assertFalse(warns.contains(SECRET), () -> "with credentials stripped: " + warns);
+            assertTrue(warns.contains("Connection refused"), () -> "and the real reason: " + warns);
+        }
     }
 }

fleetd/src/test/java/dev/ltms/fleet/testing/CapturedLog.java

@@ -21,9 +21,29 @@ import java.util.List;
  * fork. try-with-resources makes "restored the appender but not the level" impossible to write,
  * because there is only one thing to close.
  *
- * <p>This is the one way to pin or capture a logger's level and output in this test tree — do not
- * hand-roll the {@code ListAppender} + {@code setLevel} + {@code finally detachAppender} pattern;
- * use {@link #at} or {@link #of} instead.
+ * <p>New code must use this rather than hand-rolling the {@code ListAppender} + {@code setLevel} +
+ * {@code finally detachAppender} pattern: use {@link #at} or {@link #of}. It is <em>not</em> yet
+ * the only instance of the pattern in this test tree, and the earlier wording here said it was —
+ * which would leave a reader who greps unable to tell a leftover from a violation.
+ *
+ * <p>Measured on main at af95897 (2026-09-12): nine test files still hand-roll it, with 42
+ * {@code setLevel} calls on a raw logback {@code Logger} between them — {@code
+ * FleetdStartupReportTest}, {@code GitHostShapeReportTest}, {@code MemberCredentialsGapReportTest},
+ * {@code MemberTrustModelReportTest}, {@code FleetHealthMonitorTest}, {@code
+ * ClaudeCodeLauncherTest}, {@code HerdrPeerLauncherAllowListWiringTest}, {@code
+ * HerdrPeerLauncherCharterTest} and {@code OpenCodeLauncherTest}. Every one of them pairs its pin
+ * with a restore, so none is the fleetd #525 leak and none was in fleetd #529's scope, which was
+ * the 19 <em>unrestored</em> pins only. They are unmigrated, not broken.
+ *
+ * <p>Re-measure with the two commands below, from the repo root. A file that appears in the first
+ * list and not the second still hand-rolls the pattern. When the first list comes back empty, this
+ * paragraph is spent and the sentence above can go back to saying "the one way" — delete the
+ * paragraph then rather than updating the count.
+ *
+ * <pre>{@code
+ * grep -rlE '\.setLevel\(' fleetd/src/test/java --include='*.java' | grep -v CapturedLog.java
+ * grep -rl 'CapturedLog' fleetd/src/test/java --include='*.java'
+ * }</pre>
  */
 public final class CapturedLog implements AutoCloseable {
     private final Logger logger;

scripts/redeploy-fleetd.sh

@@ -42,6 +42,14 @@
 #   8. fleetd #492 — a post-restart check counts running fleetd processes and fails the whole run if
 #      more than one is alive. That is the one thing none of the checks above (healthz 200, jar id,
 #      the fresh "listening" line) can see: every one of them is satisfied by EITHER daemon.
+#   9. fleetd #512 — the ERROR-line count above is blind by construction to the exact failure #493
+#      is about: an uncaught exception in a shutdown thread never passes through the logger, so it
+#      never carries an ERROR (or SEVERE) token that any count could see. This script now also
+#      greps the previous daemon's shutdown window for that exception's real shape, and separately
+#      asserts that SessionManager's drain-complete line (fleetd #522) is present there — its
+#      absence is the real signal, because a drain that dies on its first session prints nothing
+#      else either. Warns loudly; never fails the redeploy, because by the time this is detectable
+#      the new daemon is already up and healthy.
 #
 # Usage:
 #   scripts/redeploy-fleetd.sh              # build, confirm, restart, verify
@@ -279,6 +287,49 @@ 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)"; 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)"; 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 "$err_file"
+}
+
 # fleetd #492: three real answers, not two — launchd, systemd, or genuinely unsupervised — plus a
 # fourth, "ambiguous", for the one case this script cannot tell apart: both signals firing at once.
 # That is exactly "I cannot tell who supervises this process", and guessing wrong here is how two
@@ -306,10 +357,12 @@ systemd_loaded() {
 #      never assigned to a global: a global set inside a `$( )` subshell dies with that subshell.
 #      This function packs BOTH values (kind and detail) onto that one stdout line, joined by
 #      $SUPERVISOR_DETAIL_SEP, and the caller unpacks them on its own side of the subshell boundary.
-#   2. This script runs under `set -euo pipefail` (line 50), so an unset variable is a loud
-#      failure. Do not add a `${VAR:-default}` anywhere downstream to paper over a value that
-#      should always be there — that hides a lost value instead of surfacing it (fleetd #497's
-#      defect class).
+#   2. This script runs under `set -u` (part of the `set -euo pipefail` at the top of the file), so
+#      an unset variable is a loud failure. Do not add a `${VAR:-default}` anywhere downstream to
+#      paper over a value that should always be there — that hides a lost value instead of
+#      surfacing it (fleetd #497's defect class). The mechanism is named rather than cited by line
+#      number on purpose: a line number in a comment goes stale on the next insert above it, and
+#      this one already had — it said line 50 while the `set` line was at 54.
 #   3. Every `case` on this function's return value needs an explicit final `*)` arm, chosen by
 #      whether that caller ACTS on the value (`die` — an unrecognised value must never be silently
 #      driven) or only DISPLAYS it (`echo`/`warn` and continue — a diagnostic must not go silent on
@@ -490,6 +543,114 @@ classify_amqp_connection_errors() {
   REDEPLOY_UNEXPLAINED_ERRORS=$((REDEPLOY_UNEXPLAINED_ERRORS + pending_inbox + pending_lead_mailbox))
 }
 
+# fleetd #512 part 2 — the negative check. #493's failure (an uncaught exception in a shutdown
+# thread) never passes through the logger: the JVM's default uncaught-exception handler prints
+# straight to stderr, so the line never carries a level, so classify_amqp_connection_errors's
+# ERROR/SEVERE token scan is structurally blind to it — measured on two hosts, including one where
+# even a syslog PRIORITY filter is blind to it too (fd 1 and fd 2 collapse to one socket there, so
+# every uncaught-exception line lands at priority 6/info). The fix is to grep the shape instead of
+# the level: `Exception in thread` at the start of a line (the handler's own banner) or
+# `NoClassDefFoundError` anywhere in it (the one real instance seen so far, but not the only shape
+# this could take). Kept as its own function, never folded into classify_amqp_connection_errors —
+# this is not an AMQP concern, and the two must stay independently readable and independently
+# testable.
+#
+# Sets REDEPLOY_UNCAUGHT_EXCEPTION_COUNT (lines matched) and REDEPLOY_UNCAUGHT_EXCEPTION_SAMPLE
+# (the first matching line, "" if none) so a caller can report both a count and a concrete quote
+# without re-reading the file. Pure: reads $1, sets globals, no side effects.
+scan_uncaught_exceptions() {
+  local log_file="$1" line
+  REDEPLOY_UNCAUGHT_EXCEPTION_COUNT=0
+  REDEPLOY_UNCAUGHT_EXCEPTION_SAMPLE=""
+  while IFS= read -r line || [ -n "$line" ]; do
+    case "$line" in
+      'Exception in thread'*|*NoClassDefFoundError*)
+        REDEPLOY_UNCAUGHT_EXCEPTION_COUNT=$((REDEPLOY_UNCAUGHT_EXCEPTION_COUNT + 1))
+        [ -n "$REDEPLOY_UNCAUGHT_EXCEPTION_SAMPLE" ] || REDEPLOY_UNCAUGHT_EXCEPTION_SAMPLE="$line"
+        ;;
+    esac
+  done < "$log_file"
+}
+
+# fleetd #512 part 2 — the positive check. fleetd #522 added a `log.info` at the very end of
+# SessionManager.drainAll's normal path (never in a `finally` — see the ticket discussion for why
+# that distinction matters): "drain complete: released=N abandoned=M (still BUSY at the shutdown
+# deadline)", printed once, on every successful drain, including the all-zero case. A drain that
+# dies partway through never reaches that statement, so the line's ABSENCE is a real signal — unlike
+# the ERROR-count check above, this one does not depend on the failure happening to throw.
+#
+# Sets REDEPLOY_DRAIN_COMPLETE_LINE to the matching line (last one, though drainAll runs at most
+# once per shutdown so there should never be more than one) or "" if absent. Pure, same shape as
+# scan_uncaught_exceptions above.
+find_drain_complete_line() {
+  local log_file="$1"
+  REDEPLOY_DRAIN_COMPLETE_LINE="$(grep -F 'drain complete: released=' "$log_file" | tail -1 || true)"
+}
+
+# fleetd #512 part 2 — THE TRAP, and the reason this is one function instead of two independent
+# checks the caller ORs together. Absence of the drain-complete line has TWO causes that need
+# OPPOSITE handling, and a naive "line absent -> the drain died" reading collapses them exactly the
+# way this whole ticket exists to stop: the line is emitted by the daemon being STOPPED, which is
+# running the OLD jar. Until a redeploy has landed fleetd #522 once, every previous daemon predates
+# the line and cannot emit it no matter how cleanly it drained — so on the very first redeploy after
+# #522 merged, "absent" means "too old to know how", not "died". Only once BOTH signals — this
+# line's absence AND scan_uncaught_exceptions' result — have been read together can the three real
+# outcomes be told apart:
+#
+#   complete -> the line is present: the drain finished. Name the counts it reported.
+#   died     -> the line is absent AND an uncaught-exception shape was found: the drain died. Name
+#               what was found.
+#   unknown  -> the line is absent AND no exception shape either: cannot tell. Say so, and say why
+#               (predates the line, or failed without throwing) — never worded as a pass or a
+#               failure, and never reassuring: "ok, no ERROR lines" one level up is the exact mistake
+#               this ticket exists to fix, and this outcome must not reproduce it.
+#
+# A fourth case, n/a, covers a cold start or a "loaded but wasn't running" restart: no previous
+# daemon was actually stopped THIS run, so there is no shutdown window in $log_file to have an
+# opinion about at all — scanning it anyway would read the NEW daemon's own startup lines and could
+# misreport "cannot tell" on every clean cold start. had_previous_daemon carries that fact in from
+# the caller (it already knows $OLD_PID) rather than this function re-deriving it from log content.
+#
+# Same shape as swap_if_built/refuse_drain_gate (fleetd #521/#528): the decision (which of the four
+# outcomes applies) and the action (which ok/warn line to print, and setting REDEPLOY_DRAIN_STATE
+# for the "result" section below to consult) live together in ONE function that the main flow calls
+# unconditionally — there is no guard left in the main flow to remove, invert, or bypass
+# independently of this function. Never calls die(): #512's own decision is to warn loudly and let
+# the redeploy stand, because by the time this is detectable the new daemon is already up and
+# healthy and failing here would give the operator nothing to do differently.
+report_shutdown_drain() {
+  local log_file="$1" had_previous_daemon="$2"
+  REDEPLOY_UNCAUGHT_EXCEPTION_COUNT=0
+  REDEPLOY_UNCAUGHT_EXCEPTION_SAMPLE=""
+  REDEPLOY_DRAIN_COMPLETE_LINE=""
+
+  if [ "$had_previous_daemon" != 1 ]; then
+    REDEPLOY_DRAIN_STATE="n/a"
+    ok "no previous daemon was running before this restart — nothing to check for a died shutdown drain"
+    return 0
+  fi
+
+  find_drain_complete_line "$log_file"
+  scan_uncaught_exceptions "$log_file"
+
+  if [ -n "$REDEPLOY_DRAIN_COMPLETE_LINE" ]; then
+    REDEPLOY_DRAIN_STATE="complete"
+    ok "previous daemon's shutdown drain finished: $REDEPLOY_DRAIN_COMPLETE_LINE"
+  elif [ "$REDEPLOY_UNCAUGHT_EXCEPTION_COUNT" -gt 0 ]; then
+    REDEPLOY_DRAIN_STATE="died"
+    warn "previous daemon's shutdown drain DIED — no drain-complete line, and an uncaught exception"
+    warn "was found in its shutdown window ($REDEPLOY_UNCAUGHT_EXCEPTION_COUNT line(s)):"
+    warn "    $REDEPLOY_UNCAUGHT_EXCEPTION_SAMPLE"
+    warn "Some sessions from the PREVIOUS daemon may not have been released."
+  else
+    REDEPLOY_DRAIN_STATE="unknown"
+    warn "cannot tell whether the previous daemon's shutdown drain finished — no drain-complete line"
+    warn "and no uncaught-exception shape either. This is NOT a pass and NOT a failure: it means"
+    warn "either that daemon predates fleetd #522's drain-complete log line, or its drain failed"
+    warn "without throwing (hung, or returned early)."
+  fi
+}
+
 # fleetd #517: extracted so the suite can call this decision directly, the same way #510 extracted
 # wait_for_daemon_exit so its ordering became checkable. Before this, the only test of the drain-gate
 # abort message was a grep of this script's own source for the wording — so mutating the `if` below
@@ -519,6 +680,32 @@ drain_gate_refusal() {
   fi
 }
 
+# fleetd #528 — drain_gate_refusal above is well tested (four cases, all direct), but nothing made
+# the MAIN FLOW's abort actually consult it. Before this, the main flow read
+# `die "$(drain_gate_refusal "$DO_BUILD" "$JAR_STAGED")"` directly, and mutating that one line to a
+# flat `die "aborted — nothing changed"` left the whole suite at exit 0 with zero FAIL lines and
+# byte-identical output to a clean run — every one of drain_gate_refusal's own tests still passed,
+# because they call the predicate directly and never touch this call site. That silently reinstated
+# the exact defect #517 was filed to fix. Same shape as #521/#526's should_swap/swap_if_built: a
+# predicate alone is not enough, because a test proving the predicate is right cannot also prove the
+# main flow consults it. So the decision (drain_gate_refusal) and the action (die) now live together
+# in ONE function, and the main flow calls it unconditionally instead of building the die() call
+# itself — there is no guard left in the main flow to remove, invert, or bypass independently of this
+# function. drain_gate_refusal stays separate and separately tested because the message-selection
+# logic is worth naming and testing on its own; refuse_drain_gate is the only thing that ever dies.
+#
+# What the behavioural tests above still cannot pin on their own: deleting the call to this function
+# from the main flow altogether — they call refuse_drain_gate directly, never through the main flow,
+# because sourcing stops before the main flow ever runs (see the SOURCED guard below). That gap is
+# closed the same way swap_if_built's is: test_refuse_drain_gate_call_site_present greps this script
+# for the real invocation, the same shape test_swap_ordered_after_wait_and_before_start already uses
+# for the swap call. Deliberately NOT written out here as a literal quoted string, so this comment
+# itself can never become a second match for that test's needle.
+refuse_drain_gate() {
+  local do_build="$1" staged_path="$2"
+  die "$(drain_gate_refusal "$do_build" "$staged_path")"
+}
+
 # CB-600: sourceable for testing. When this file is SOURCED (not executed) it stops here — nothing
 # below runs — so a test harness can `source` it to call check_log_path_matches_plist (or the
 # other pure helpers above) against a throwaway plist fixture without ever reaching the mutating
@@ -677,10 +864,11 @@ if [ -n "$OLD_PID" ] && [ "$ASSUME_YES" = 0 ]; then
   echo
   read -r -p "   Fleet drained? type yes to restart: " reply
   if [ "$reply" != "yes" ]; then
-    # fleetd #493 / #517: "nothing changed" would be a lie once a build has run and staged a jar —
-    # see drain_gate_refusal above for the full decision and why each of its four cases reads the
-    # way it does.
-    die "$(drain_gate_refusal "$DO_BUILD" "$JAR_STAGED")"
+    # fleetd #493 / #517 / #528: "nothing changed" would be a lie once a build has run and staged a
+    # jar — see drain_gate_refusal above for the full decision and why each of its four cases reads
+    # the way it does. refuse_drain_gate composes that message AND calls die itself, so this guard
+    # has nothing left of its own to get wrong beyond whether it calls refuse_drain_gate at all.
+    refuse_drain_gate "$DO_BUILD" "$JAR_STAGED"
   fi
 fi
 
@@ -736,16 +924,20 @@ 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 < "$OUT" 2>/dev/null || echo 0)"
@@ -873,6 +1065,14 @@ trap 'rm -f "$FRESH_LOG"' EXIT
 tail -n "+$((RESTART_MARK + 1))" "$OUT" > "$FRESH_LOG" 2>/dev/null || true
 classify_amqp_connection_errors "$FRESH_LOG"
 
+# fleetd #512 part 2: the previous daemon's shutdown drain, checked in the same fresh-log region —
+# see report_shutdown_drain above for the full decision (four outcomes, one of them a deliberate
+# "cannot tell"). HAD_OLD_PID crosses in whether a previous daemon was actually stopped this run;
+# see the function's own comment for why that matters.
+say "previous daemon's shutdown drain"
+HAD_OLD_PID=0; [ -n "$OLD_PID" ] && HAD_OLD_PID=1
+report_shutdown_drain "$FRESH_LOG" "$HAD_OLD_PID"
+
 # fleetd #492: checked here, after healthz and the fresh-log check have both had time to run, so a
 # supervisor that revives the OLD jar a few seconds late is caught too. Every check above (healthz
 # 200, jar id, the fresh 'listening' line) is satisfied by EITHER daemon if two are alive — this is
@@ -882,7 +1082,14 @@ assert_single_daemon "$(running_pid)"
 say "result"
 ok "pid $NEW_PID, jar $(jar_id)"
 if [ "$REDEPLOY_ERROR_COUNT" -eq 0 ]; then
-  ok "no ERROR lines since restart"
+  # fleetd #512 item 4: this line must not print when the shutdown-drain check above found the
+  # previous daemon's drain died, or could not tell — either would make "no ERROR lines" read as a
+  # clean bill of health it is not (an uncaught exception never carries an ERROR token to begin
+  # with, so this count alone cannot see that failure). "complete" and "n/a" are the only two
+  # outcomes report_shutdown_drain sets that mean nothing is wrong there.
+  if [ "$REDEPLOY_DRAIN_STATE" = "complete" ] || [ "$REDEPLOY_DRAIN_STATE" = "n/a" ]; then
+    ok "no ERROR lines since restart"
+  fi
 elif [ "$REDEPLOY_UNEXPLAINED_ERRORS" -eq 0 ]; then
   ok "$REDEPLOY_RECOVERED_AMQP_ERRORS AMQP connection reset ERROR lines recovered since restart"
 else

scripts/test-redeploy-fleetd.sh

@@ -521,6 +521,210 @@ test_drain_gate_refusal_no_build_staged_absent() {
   assert_equals "aborted — nothing changed" "$result" "no-build+staged-absent refusal wording"
 }
 
+# fleetd #528 — the four tests above pin drain_gate_refusal(), and that is ALL they pin: they call
+# the predicate directly and never touch the main flow's call site. That was measured to be not
+# enough, the same way test_should_swap_true_when_build_ran/test_should_swap_false_when_build_skipped
+# were not enough for #521: with the main flow reading `die "$(drain_gate_refusal "$DO_BUILD"
+# "$JAR_STAGED")"`, replacing that whole line with a flat `die "aborted — nothing changed"` left this
+# suite at exit 0 with zero FAIL lines and byte-identical output to a clean run. Nothing above could
+# tell the difference, because none of it calls anything at or above the call site itself.
+#
+# So these four call refuse_drain_gate() — the function the main flow actually calls, holding the
+# composed message and the die() together — with die() stubbed to RECORD whether it was called and
+# with what message, instead of exiting the process. That fails if refuse_drain_gate stops consulting
+# drain_gate_refusal, mangles what it passes it, or simply never calls die.
+#
+# What none of these four can catch: deleting the `refuse_drain_gate "$DO_BUILD" "$JAR_STAGED"` line
+# from the main flow altogether — see the comment above refuse_drain_gate in redeploy-fleetd.sh for
+# why no test in this file can do better than that (sourcing stops before the main flow runs).
+DIED_CALLED=0
+DIED_MESSAGE=""
+stub_die_recorder() {
+  DIED_CALLED=0
+  DIED_MESSAGE=""
+  die() { DIED_CALLED=1; DIED_MESSAGE="$*"; }
+}
+
+test_refuse_drain_gate_build_ran_staged_present() {
+  source "$ROOT/scripts/redeploy-fleetd.sh"
+  local dir staged
+  dir="$TMP/refuse-drain-build-staged"; mkdir -p "$dir"
+  staged="$dir/fleetd-new.jar"
+  printf 'staged jar bytes' > "$staged"
+  stub_die_recorder
+  refuse_drain_gate 1 "$staged"
+  [ "$DIED_CALLED" = 1 ] \
+    || fail "refuse_drain_gate build-ran+staged-present must call die, and did not"
+  printf '%s' "$DIED_MESSAGE" | grep -qF "$staged" \
+    || fail "refuse_drain_gate build-ran+staged-present die message does not name the staged jar"
+  printf '%s' "$DIED_MESSAGE" | grep -qF 'Rerun WITHOUT --no-build' \
+    || fail "refuse_drain_gate build-ran+staged-present die message is missing the rerun instruction"
+  if printf '%s' "$DIED_MESSAGE" | grep -qF 'nothing changed'; then
+    fail "refuse_drain_gate build-ran+staged-present must not claim nothing changed — the jar already moved"
+  fi
+  source "$ROOT/scripts/redeploy-fleetd.sh"
+}
+
+test_refuse_drain_gate_build_ran_staged_absent() {
+  source "$ROOT/scripts/redeploy-fleetd.sh"
+  local dir
+  dir="$TMP/refuse-drain-build-no-staged"; mkdir -p "$dir"
+  stub_die_recorder
+  refuse_drain_gate 1 "$dir/fleetd-new.jar"
+  [ "$DIED_CALLED" = 1 ] \
+    || fail "refuse_drain_gate build-ran+staged-absent must call die, and did not"
+  assert_equals "aborted — nothing changed" "$DIED_MESSAGE" "refuse_drain_gate build-ran+staged-absent die message"
+  source "$ROOT/scripts/redeploy-fleetd.sh"
+}
+
+test_refuse_drain_gate_no_build_staged_present() {
+  source "$ROOT/scripts/redeploy-fleetd.sh"
+  local dir staged
+  dir="$TMP/refuse-drain-no-build-staged"; mkdir -p "$dir"
+  staged="$dir/fleetd-new.jar"
+  printf 'leftover staged jar bytes' > "$staged"
+  stub_die_recorder
+  refuse_drain_gate 0 "$staged"
+  [ "$DIED_CALLED" = 1 ] \
+    || fail "refuse_drain_gate no-build+staged-present must call die, and did not"
+  assert_equals "aborted — nothing changed" "$DIED_MESSAGE" "refuse_drain_gate no-build+staged-present die message"
+  source "$ROOT/scripts/redeploy-fleetd.sh"
+}
+
+test_refuse_drain_gate_no_build_staged_absent() {
+  source "$ROOT/scripts/redeploy-fleetd.sh"
+  local dir
+  dir="$TMP/refuse-drain-no-build-no-staged"; mkdir -p "$dir"
+  stub_die_recorder
+  refuse_drain_gate 0 "$dir/fleetd-new.jar"
+  [ "$DIED_CALLED" = 1 ] \
+    || fail "refuse_drain_gate no-build+staged-absent must call die, and did not"
+  assert_equals "aborted — nothing changed" "$DIED_MESSAGE" "refuse_drain_gate no-build+staged-absent die message"
+  source "$ROOT/scripts/redeploy-fleetd.sh"
+}
+
+# fleetd #528 — closes the one gap the four behavioural tests above cannot: they call
+# refuse_drain_gate directly, and sourcing stops before the main flow ever runs (the SOURCED guard),
+# so none of them can prove the main flow still CALLS refuse_drain_gate at all. Same shape as
+# test_swap_ordered_after_wait_and_before_start: a source-text grep for the real call site. This is
+# what actually kills the item-1 mutation from the ticket — replacing the main flow's call with a
+# flat `die "aborted — nothing changed"` removes this exact needle, where none of the behavioural
+# tests above would even notice.
+#
+# The grep ends `|| true`: this file runs under `set -euo pipefail`, so an ABSENT needle would fail
+# the assignment and `set -e` would kill the whole suite before the `[ -n ... ] || fail` guard below
+# ever ran — the exact dead-check shape fleetd #528 also flags as a sweep finding (see the PR body).
+test_refuse_drain_gate_call_site_present() {
+  local src="$ROOT/scripts/redeploy-fleetd.sh" call_line
+  call_line="$(grep -Fn 'refuse_drain_gate "$DO_BUILD" "$JAR_STAGED"' "$src" | head -1 | cut -d: -f1 || true)"
+  [ -n "$call_line" ] \
+    || fail "could not find the main flow's refuse_drain_gate call site in redeploy-fleetd.sh"
+}
+
+# fleetd #504 — the "loaded but not currently running" branches for launchd/systemd used to run
+# `launchctl unload`/`systemctl --user stop` with `2>/dev/null || true` and print `ok`
+# unconditionally, so a real supervisor failure (e.g. it cannot reach launchd/the systemd user bus)
+# read exactly like a harmless already-stopped answer. unload_launchd_if_loaded/
+# stop_systemd_if_loaded (redeploy-fleetd.sh, right after systemd_loaded) apply systemd_loaded's own
+# "capture stderr separately — only a non-zero exit WITH stderr is a real failure" pattern to the
+# WRITE side. Both call the real `launchctl`/`systemctl` binaries directly (they are not overridable
+# wrapper functions the way launchd_loaded/systemd_loaded are), so these tests put a stub binary
+# first on PATH — the same technique test_detect_supervisor_systemd_probe_error_is_unclear above
+# already uses for `systemctl`.
+test_unload_launchd_if_loaded_dies_on_real_failure() {
+  local bin_dir output rc=0 saved_plist="$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
@@ -720,6 +924,164 @@ test_unattributable_quiet_mutation_is_caught() {
   printf 'Unattributable mutation: FAIL: cross-unattributable recovered: expected 0, got 2\n'
 }
 
+# fleetd #512 part 2 — the negative check (scan_uncaught_exceptions). The heart of this half of the
+# ticket: a fixture with the uncaught-exception shape and NO line carrying an ERROR token at all,
+# proving the scan finds it without one. A fixture that also carried an ERROR line would pass for
+# the wrong reason.
+test_scan_uncaught_exceptions_finds_shape_without_error_token() {
+  cat > "$TMP/scan-died.log" <<'LOG'
+2026-09-12 10:15:00 INFO fleetd listening on 127.0.0.1:8765
+Exception in thread "Thread-0" java.lang.NoClassDefFoundError: reactor/core/Exceptions
+	at dev.ltms.fleet.session.SessionManager.drainAll(SessionManager.java:1081)
+LOG
+  local error_count
+  error_count="$(grep -c ' ERROR ' "$TMP/scan-died.log" || true)"
+  [ "$error_count" = "0" ] \
+    || fail "test fixture error: scan-died.log unexpectedly carries an ERROR token"
+  scan_uncaught_exceptions "$TMP/scan-died.log"
+  assert_equals 1 "$REDEPLOY_UNCAUGHT_EXCEPTION_COUNT" "scan must find the exception without an ERROR token"
+  printf '%s' "$REDEPLOY_UNCAUGHT_EXCEPTION_SAMPLE" | grep -qF 'NoClassDefFoundError' \
+    || fail "scan did not capture the matching line as the sample"
+}
+
+test_scan_uncaught_exceptions_clean_control() {
+  cat > "$TMP/scan-clean.log" <<'LOG'
+2026-09-12 10:15:00 INFO fleetd listening on 127.0.0.1:8765
+2026-09-12 10:15:05 INFO dev.ltms.fleet.session.SessionManager - drain complete: released=0 abandoned=0 (still BUSY at the shutdown deadline)
+LOG
+  scan_uncaught_exceptions "$TMP/scan-clean.log"
+  assert_equals 0 "$REDEPLOY_UNCAUGHT_EXCEPTION_COUNT" "clean control must find no uncaught exception"
+  assert_equals "" "$REDEPLOY_UNCAUGHT_EXCEPTION_SAMPLE" "clean control sample must be empty"
+}
+
+# fleetd #512 part 2 — the positive check (find_drain_complete_line). Both halves of #522's line:
+# present, and absent.
+test_find_drain_complete_line_present() {
+  cat > "$TMP/drain-line-present.log" <<'LOG'
+2026-09-12 10:15:05 INFO dev.ltms.fleet.session.SessionManager - drain complete: released=2 abandoned=1 (still BUSY at the shutdown deadline)
+LOG
+  find_drain_complete_line "$TMP/drain-line-present.log"
+  printf '%s' "$REDEPLOY_DRAIN_COMPLETE_LINE" | grep -qF 'released=2 abandoned=1' \
+    || fail "find_drain_complete_line did not capture the present line"
+}
+
+test_find_drain_complete_line_absent() {
+  cat > "$TMP/drain-line-absent.log" <<'LOG'
+2026-09-12 10:15:00 INFO fleetd listening on 127.0.0.1:8765
+LOG
+  find_drain_complete_line "$TMP/drain-line-absent.log"
+  assert_equals "" "$REDEPLOY_DRAIN_COMPLETE_LINE" "find_drain_complete_line must report empty when absent"
+}
+
+# fleetd #512 part 2 — report_shutdown_drain, the composite decision+action function the main flow
+# calls unconditionally (same shape as swap_if_built/refuse_drain_gate, #521/#528). These four cover
+# the four outcomes named in the ticket's "trap": complete, died, unknown ("cannot tell" — neither a
+# pass nor a failure), and n/a (no previous daemon was actually stopped this run).
+#
+# Deliberately NOT run inside `$(...)`: report_shutdown_drain sets REDEPLOY_DRAIN_STATE as a global
+# side effect that these tests need to read back afterward, and a command substitution forks a
+# subshell that global assignment would not survive (the exact trap documented above
+# detect_supervisor in redeploy-fleetd.sh, for the same reason). Plain output redirection to a file
+# does not fork a subshell, so it is used to capture what was printed instead.
+test_report_shutdown_drain_died_without_error_token() {
+  cat > "$TMP/drain-died.log" <<'LOG'
+2026-09-12 10:15:00 INFO fleetd listening on 127.0.0.1:8765
+2026-09-12 10:15:05 INFO dev.ltms.fleet.Fleetd - shutting down
+Exception in thread "Thread-0" java.lang.NoClassDefFoundError: reactor/core/Exceptions
+	at dev.ltms.fleet.session.SessionManager.drainAll(SessionManager.java:1081)
+LOG
+  local error_count
+  error_count="$(grep -c ' ERROR ' "$TMP/drain-died.log" || true)"
+  [ "$error_count" = "0" ] \
+    || fail "test fixture error: drain-died.log unexpectedly carries an ERROR token"
+
+  report_shutdown_drain "$TMP/drain-died.log" 1 > "$TMP/drain-died-output" 2>&1
+  assert_equals "died" "$REDEPLOY_DRAIN_STATE" "died fixture must set REDEPLOY_DRAIN_STATE=died"
+  assert_equals 1 "$REDEPLOY_UNCAUGHT_EXCEPTION_COUNT" "died fixture uncaught-exception count"
+  grep -qF 'NoClassDefFoundError' "$TMP/drain-died-output" \
+    || fail "report_shutdown_drain did not report the uncaught-exception shape it found"
+  grep -qF 'DIED' "$TMP/drain-died-output" \
+    || fail "report_shutdown_drain did not report the drain as DIED"
+}
+
+test_report_shutdown_drain_complete_control() {
+  cat > "$TMP/drain-complete.log" <<'LOG'
+2026-09-12 10:15:00 INFO fleetd listening on 127.0.0.1:8765
+2026-09-12 10:15:05 INFO dev.ltms.fleet.Fleetd - shutting down
+2026-09-12 10:15:05 INFO dev.ltms.fleet.session.SessionManager - drain complete: released=3 abandoned=0 (still BUSY at the shutdown deadline)
+LOG
+  report_shutdown_drain "$TMP/drain-complete.log" 1 > "$TMP/drain-complete-output" 2>&1
+  assert_equals "complete" "$REDEPLOY_DRAIN_STATE" "complete-control fixture must set REDEPLOY_DRAIN_STATE=complete"
+  assert_equals 0 "$REDEPLOY_UNCAUGHT_EXCEPTION_COUNT" "complete-control fixture must find no uncaught exception"
+  grep -qF 'released=3 abandoned=0' "$TMP/drain-complete-output" \
+    || fail "report_shutdown_drain did not report the drain-complete counts"
+}
+
+test_report_shutdown_drain_unknown_cannot_tell() {
+  cat > "$TMP/drain-unknown.log" <<'LOG'
+2026-09-12 10:15:00 INFO fleetd listening on 127.0.0.1:8765
+2026-09-12 10:15:05 INFO dev.ltms.fleet.Fleetd - shutting down
+LOG
+  report_shutdown_drain "$TMP/drain-unknown.log" 1 > "$TMP/drain-unknown-output" 2>&1
+  assert_equals "unknown" "$REDEPLOY_DRAIN_STATE" "cannot-tell fixture must set REDEPLOY_DRAIN_STATE=unknown"
+  grep -qF 'cannot tell' "$TMP/drain-unknown-output" \
+    || fail "report_shutdown_drain did not say it could not tell"
+  if grep -qF '   ok' "$TMP/drain-unknown-output"; then
+    fail "cannot-tell outcome must not be printed via ok() — it is neither a pass nor a failure"
+  fi
+}
+
+# A cold start (or a restart where nothing was actually stopped) has no previous-daemon shutdown
+# window to have an opinion about at all. This fixture's log content looks exactly like a died drain
+# — proving the had_previous_daemon=0 gate is actually consulted, not merely documented: without it,
+# this would misreport "died" or "unknown" on every clean cold start.
+test_report_shutdown_drain_no_previous_daemon_is_na() {
+  cat > "$TMP/drain-na.log" <<'LOG'
+Exception in thread "Thread-0" java.lang.NoClassDefFoundError: reactor/core/Exceptions
+LOG
+  report_shutdown_drain "$TMP/drain-na.log" 0 > "$TMP/drain-na-output" 2>&1
+  assert_equals "n/a" "$REDEPLOY_DRAIN_STATE" "no-previous-daemon fixture must set REDEPLOY_DRAIN_STATE=n/a even though the log content looks like a died drain"
+  grep -qF 'nothing to check' "$TMP/drain-na-output" \
+    || fail "report_shutdown_drain did not report that there was nothing to check"
+}
+
+# fleetd #512 — closes the gap none of the seven tests above can: they call report_shutdown_drain
+# directly, and sourcing stops before the main flow ever runs (the SOURCED guard), so none of them
+# can prove the main flow still calls it at all. Same shape as test_refuse_drain_gate_call_site_present
+# and test_swap_ordered_after_wait_and_before_start: a source-text grep for the real call site, plus
+# an ordering check against its neighbours in the verify/result flow.
+test_report_shutdown_drain_call_site_present() {
+  local src="$ROOT/scripts/redeploy-fleetd.sh" call_line
+  call_line="$(grep -Fn 'report_shutdown_drain "$FRESH_LOG" "$HAD_OLD_PID"' "$src" | head -1 | cut -d: -f1 || true)"
+  [ -n "$call_line" ] \
+    || fail "could not find the main flow's report_shutdown_drain call site in redeploy-fleetd.sh"
+}
+
+test_report_shutdown_drain_ordered_after_classify_and_before_result() {
+  local src="$ROOT/scripts/redeploy-fleetd.sh" classify_line drain_line result_line
+  classify_line="$(grep -Fn 'classify_amqp_connection_errors "$FRESH_LOG"' "$src" | tail -1 | cut -d: -f1 || true)"
+  drain_line="$(grep -Fn 'report_shutdown_drain "$FRESH_LOG" "$HAD_OLD_PID"' "$src" | head -1 | cut -d: -f1 || true)"
+  result_line="$(grep -Fn 'say "result"' "$src" | head -1 | cut -d: -f1 || true)"
+  [ -n "$classify_line" ] || fail "could not find the classify_amqp_connection_errors call site"
+  [ -n "$drain_line" ] || fail "could not find the report_shutdown_drain call site"
+  [ -n "$result_line" ] || fail "could not find the result section"
+  [ "$drain_line" -gt "$classify_line" ] \
+    || fail "report_shutdown_drain (line $drain_line) is not after classify_amqp_connection_errors (line $classify_line)"
+  [ "$drain_line" -lt "$result_line" ] \
+    || fail "report_shutdown_drain (line $drain_line) is not before the result section (line $result_line)"
+}
+
+# fleetd #512 item 4 — the summary line must not read as reassurance when the shutdown-drain check
+# found something wrong (or could not tell). Sourcing stops before the main flow runs, so this is a
+# source-text check like test_drain_gate_abort_message_says_no_no_build above.
+test_no_error_lines_message_gated_by_drain_state() {
+  local src="$ROOT/scripts/redeploy-fleetd.sh" block
+  block="$(grep -B2 -F 'ok "no ERROR lines since restart"' "$src")"
+  [ -n "$block" ] || fail "could not find the 'no ERROR lines since restart' line in redeploy-fleetd.sh"
+  printf '%s' "$block" | grep -qF 'REDEPLOY_DRAIN_STATE' \
+    || fail "'no ERROR lines since restart' is not guarded by the shutdown-drain outcome (fleetd #512 item 4)"
+}
+
 test_detect_supervisor_launchd_only
 test_detect_supervisor_systemd_only
 test_detect_supervisor_none
@@ -753,6 +1115,16 @@ test_drain_gate_refusal_build_ran_staged_present
 test_drain_gate_refusal_build_ran_staged_absent
 test_drain_gate_refusal_no_build_staged_present
 test_drain_gate_refusal_no_build_staged_absent
+test_refuse_drain_gate_build_ran_staged_present
+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
@@ -764,4 +1136,15 @@ test_other_error_is_unexplained
 test_recovery_requirement_mutation_is_caught
 test_shared_counter_mutation_is_caught
 test_unattributable_quiet_mutation_is_caught
+test_scan_uncaught_exceptions_finds_shape_without_error_token
+test_scan_uncaught_exceptions_clean_control
+test_find_drain_complete_line_present
+test_find_drain_complete_line_absent
+test_report_shutdown_drain_died_without_error_token
+test_report_shutdown_drain_complete_control
+test_report_shutdown_drain_unknown_cannot_tell
+test_report_shutdown_drain_no_previous_daemon_is_na
+test_report_shutdown_drain_call_site_present
+test_report_shutdown_drain_ordered_after_classify_and_before_result
+test_no_error_lines_message_gated_by_drain_state
 printf 'PASS: redeploy log classifier\n'