fix(ingestion): tripwire alerts survive outages; drop only permanent rejections

PR #160 round 2: dropping a message after 5 failures could not tell a
poison message from an ntfy outage. A 15-minute outage dropped the only
incident with nothing reported, and drops were not counted (74 of 300
removals were neither delivered nor counted).

- audit.NtfyNotifier returns *RejectedError for non-2xx; Permanent() is
  true for 4xx other than 429.
- The tripwire queue drops a message only on a permanent rejection and
  counts it; every other failure is an outage and is retried, however
  long. Each flush tries every queued message, so an undeliverable one
  cannot block the rest.
- Overflow and drops go into one counter, sent as a summary retried
  every tick and never dropped. Every removal is delivered or counted.

Tests (red first): an outage of ~100 ticks still delivers the only
incident; delivered + counted equals the number of removals; a
permanently rejected message is dropped and counted. The short-outage
test is renamed to say it is short. The Watch doc no longer claims
'loses none' unconditionally.

Refs #157

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

ingestion/internal/api/raw_race_test.go

@@ -8,6 +8,7 @@ import (
 	"os"
 	"path/filepath"
 	"strings"
+	"sync"
 	"sync/atomic"
 	"syscall"
 	"testing"
@@ -25,6 +26,18 @@ import (
 
 const raceSentinel = "SENTINEL-153-RACE"
 
+// startSwapper runs fn until the test ends, and waits for it to exit before
+// t.TempDir's cleanup removes the directory it writes into (hyperguild#162:
+// an unawaited swapper raced the cleanup, "directory not empty").
+func startSwapper(t *testing.T, fn func(stop *atomic.Bool)) {
+	t.Helper()
+	var stop atomic.Bool
+	var wg sync.WaitGroup
+	wg.Add(1)
+	go func() { defer wg.Done(); fn(&stop) }()
+	t.Cleanup(func() { stop.Store(true); wg.Wait() })
+}
+
 func swap(raw, outside string, fifo bool, stop *atomic.Bool) {
 	tgt := filepath.Join(raw, "r.md")
 	for i := 0; !stop.Load(); i++ {
@@ -76,9 +89,7 @@ func readers(brain, raw string) string {
 
 func TestRawReadersLeakNothingUnderASymlinkSwapRace(t *testing.T) {
 	brain, raw, outside := raceBrain(t)
-	var stop atomic.Bool
-	go swap(raw, outside, false, &stop)
-	defer stop.Store(true)
+	startSwapper(t, func(stop *atomic.Bool) { swap(raw, outside, false, stop) })
 	leaks, iters := 0, 0
 	for deadline := time.Now().Add(1500 * time.Millisecond); time.Now().Before(deadline); iters++ {
 		leaks += strings.Count(readers(brain, raw), raceSentinel)
@@ -91,8 +102,7 @@ func TestRawReadersLeakNothingUnderASymlinkSwapRace(t *testing.T) {
 
 func TestRawReadersDoNotHangUnderAFIFOSwapRace(t *testing.T) {
 	brain, raw, outside := raceBrain(t)
-	var stop atomic.Bool
-	go swap(raw, outside, true, &stop)
+	startSwapper(t, func(stop *atomic.Bool) { swap(raw, outside, true, stop) })
 	var iters atomic.Int64
 	done := make(chan struct{})
 	go func() {
@@ -104,9 +114,7 @@ func TestRawReadersDoNotHangUnderAFIFOSwapRace(t *testing.T) {
 	}()
 	select {
 	case <-done:
-		stop.Store(true)
 	case <-time.After(6 * time.Second):
-		stop.Store(true)
 		// Unblock the stuck open so the test process can exit.
 		if f, err := os.OpenFile(filepath.Join(raw, "r.md"), os.O_WRONLY|syscall.O_NONBLOCK, 0); err == nil {
 			_ = f.Close()
@@ -147,8 +155,7 @@ func TestDocmarkExtractionLeaksNothingUnderASymlinkSwapRace(t *testing.T) {
 	if err := os.WriteFile(tgt, []byte("inside docx bytes"), 0o644); err != nil {
 		t.Fatal(err)
 	}
-	var stop atomic.Bool
-	go func() {
+	startSwapper(t, func(stop *atomic.Bool) {
 		for i := 0; !stop.Load(); i++ {
 			tmp := filepath.Join(raw, ".swapdocx")
 			_ = os.Remove(tmp)
@@ -159,8 +166,7 @@ func TestDocmarkExtractionLeaksNothingUnderASymlinkSwapRace(t *testing.T) {
 			}
 			_ = os.Rename(tmp, tgt)
 		}
-	}()
-	defer stop.Store(true)
+	})
 	iters := 0
 	for deadline := time.Now().Add(1500 * time.Millisecond); time.Now().Before(deadline); iters++ {
 		_, _ = api.ReadIngestFile(brain, tgt)

ingestion/internal/audit/ntfy.go

@@ -49,7 +49,19 @@ func (n *NtfyNotifier) Notify(ctx context.Context, msg string) error {
 	}
 	defer func() { _ = resp.Body.Close() }()
 	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
-		return fmt.Errorf("ntfy notify: status %d", resp.StatusCode)
+		return &RejectedError{Status: resp.StatusCode}
 	}
 	return nil
 }
+
+// RejectedError is a non-2xx ntfy response.
+type RejectedError struct{ Status int }
+
+func (e *RejectedError) Error() string { return fmt.Sprintf("ntfy notify: status %d", e.Status) }
+
+// Permanent reports whether resending the same message cannot succeed: a 4xx
+// other than 429 (rate limited). Callers may drop such a message; anything
+// else (5xx, 429, transport errors) is an outage and must be retried.
+func (e *RejectedError) Permanent() bool {
+	return e.Status >= 400 && e.Status < 500 && e.Status != http.StatusTooManyRequests
+}

ingestion/internal/audit/ntfy_rejected_test.go

@@ -0,0 +1,27 @@
+package audit
+
+import (
+	"context"
+	"errors"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// A non-2xx ntfy response is a *RejectedError, and only a client-side refusal
+// other than rate limiting is permanent: callers may drop a message on a
+// permanent rejection and must retry the rest (hyperguild#157, PR #160).
+func TestNotifyClassifiesRejectionsAsPermanentOrNot(t *testing.T) {
+	for status, permanent := range map[int]bool{413: true, 400: true, 403: true, 429: false, 500: false, 503: false} {
+		srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) { w.WriteHeader(status) }))
+		err := NewNtfyNotifier(srv.URL, "").Notify(context.Background(), "m")
+		srv.Close()
+		var rej *RejectedError
+		require.True(t, errors.As(err, &rej), "status %d: %v", status, err)
+		assert.Equal(t, status, rej.Status)
+		assert.Equal(t, permanent, rej.Permanent(), "status %d", status)
+	}
+}

ingestion/internal/symlinkguard/symlinkguard.go

@@ -9,6 +9,7 @@ package symlinkguard
 
 import (
 	"context"
+	"errors"
 	"fmt"
 	"io/fs"
 	"log/slog"
@@ -21,7 +22,14 @@ import (
 )
 
 // DefaultInterval is used when Watch is given a non-positive interval.
-const DefaultInterval = 5 * time.Minute
+const DefaultInterval = 30 * time.Second
+
+// Alert bounds (PR #160 review): the planter chooses the paths, so neither an
+// alert's size nor the queue may grow with what was planted.
+const (
+	maxPathsPerAlert = 10 // further paths are counted, not listed
+	maxPending       = 5  // queued alerts; the rest become one summary
+)
 
 // Notifier is the out-of-band alert sink (audit.Notifier satisfies it).
 type Notifier interface {
@@ -41,13 +49,7 @@ func (r Result) Clean() bool { return len(r.Symlinks) == 0 && len(r.Unreadable)
 // String describes the findings for a log line or alert. Paths are quoted, so
 // a name containing a newline cannot blur two findings together.
 func (r Result) String() string {
-	q := func(xs []string) string {
-		out := make([]string, len(xs))
-		for i, x := range xs {
-			out[i] = strconv.Quote(x)
-		}
-		return strings.Join(out, ", ")
-	}
+	q := r.quoted
 	var parts []string
 	if len(r.Symlinks) > 0 {
 		parts = append(parts, fmt.Sprintf("%d symlink(s): %s", len(r.Symlinks), q(r.Symlinks)))
@@ -58,6 +60,24 @@ func (r Result) String() string {
 	return strings.Join(parts, "; ")
 }
 
+// quoted joins at most maxPathsPerAlert paths, each strconv-quoted, and
+// counts the rest.
+func (Result) quoted(xs []string) string {
+	shown := xs
+	if len(shown) > maxPathsPerAlert {
+		shown = shown[:maxPathsPerAlert]
+	}
+	out := make([]string, len(shown))
+	for i, x := range shown {
+		out[i] = strconv.Quote(x)
+	}
+	s := strings.Join(out, ", ")
+	if extra := len(xs) - len(shown); extra > 0 {
+		s += fmt.Sprintf(" and %d more", extra)
+	}
+	return s
+}
+
 // Scan walks the brain dir and reports every symlink in it and every
 // directory it could not read, continuing past both. Only the ROOT .git is
 // skipped (it is brain-sync's, and no reader walks it); a nested .git is
@@ -106,15 +126,20 @@ func Scan(brainDir string) (Result, error) {
 }
 
 // Watch rescans every interval (DefaultInterval if non-positive) until ctx
-// ends, and notifies once per distinct non-clean result. A result counts as
-// reported only after Notify succeeds, so an alert lost to a notifier outage
-// is retried on the next tick.
+// ends. Every symlink found is removed (the link, never its target:
+// hyperguild#157 decision), so one is readable for at most one interval.
+// Each removal is an alert; a persistent finding (a link it could not remove,
+// an unreadable dir, a scan error) alerts once per distinct state. Alerts
+// queue until Notify succeeds (see alertQueue): an outage delays them, and
+// any that do not fit the bounded queue, or are undeliverable, are counted in
+// a summary that is never dropped, so every incident is delivered or counted.
 func Watch(ctx context.Context, brainDir string, interval time.Duration, n Notifier) {
 	if interval <= 0 {
 		interval = DefaultInterval
 	}
 	t := time.NewTicker(interval)
 	defer t.Stop()
+	q := &alertQueue{}
 	reported := ""
 	for {
 		select {
@@ -123,31 +148,117 @@ func Watch(ctx context.Context, brainDir string, interval time.Duration, n Notif
 		case <-t.C:
 		}
 		r, err := Scan(brainDir)
-		var key, msg string
-		switch {
-		case err != nil:
-			key = "error: " + err.Error()
-			msg = "brain symlink tripwire (hyperguild#157): cannot scan the brain dir: " + err.Error()
-		case r.Clean():
-			reported = ""
-			continue
-		default:
-			key = r.String()
-			msg = "brain symlink tripwire (hyperguild#157): " + key +
-				". Readers may follow symlinks out of the brain; remove them."
+		key := ""
+		if err != nil {
+			key = "cannot scan the brain dir: " + err.Error()
+		} else {
+			removed, failed := removeSymlinks(brainDir, r.Symlinks)
+			if len(removed) > 0 {
+				slog.Error("brain symlink tripwire removed symlinks", "count", len(removed))
+				q.push(fmt.Sprintf("brain symlink tripwire (hyperguild#157): removed %d symlink(s) from the brain dir: %s",
+					len(removed), Result{}.quoted(removed)))
 			}
-		if key == reported {
-			continue
+			key = Result{Symlinks: failed, Unreadable: r.Unreadable}.String()
 		}
+		if key != reported {
+			reported = key
+			if key != "" {
 				slog.Error("brain symlink tripwire", "finding", key)
+				q.push("brain symlink tripwire (hyperguild#157): " + key +
+					". Readers may follow symlinks out of the brain; fix them.")
+			}
+		}
+		q.flush(ctx, n)
+	}
+}
+
+// alertQueue holds alerts until Notify accepts them. It is bounded (at most
+// maxPending queued) and loses nothing from the record: every overflow and
+// every dropped message increments suppressed, sent as a summary that is
+// retried every tick and never dropped.
+//
+// A message is dropped only when the notifier rejects it permanently (an
+// error with Permanent() == true, e.g. ntfy 413). Every other failure is an
+// outage, and nothing is dropped however long it lasts (PR #160 round 2:
+// dropping after N raw failures lost the only incident of a long outage).
+// Each flush tries every queued message, so one undeliverable message cannot
+// block the others.
+type alertQueue struct {
+	items      []string
+	suppressed int
+}
+
+func (q *alertQueue) push(msg string) {
+	if len(q.items) >= maxPending {
+		q.suppressed++
+		return
+	}
+	q.items = append(q.items, msg)
+}
+
+func (q *alertQueue) flush(ctx context.Context, n Notifier) {
 	if n == nil {
-			reported = key
-			continue
+		q.items, q.suppressed = nil, 0
+		return
 	}
-		if err := n.Notify(ctx, msg); err != nil {
+	kept := q.items[:0]
+	for _, msg := range q.items {
+		err := n.Notify(ctx, msg)
+		switch {
+		case err == nil:
+		case permanent(err):
+			slog.Error("symlink tripwire alert rejected permanently; counted in the summary", "err", err)
+			q.suppressed++
+		default:
 			slog.Warn("symlink tripwire notify failed; will retry", "err", err)
+			kept = append(kept, msg)
+		}
+	}
+	q.items = kept
+	if q.suppressed > 0 {
+		msg := fmt.Sprintf("brain symlink tripwire (hyperguild#157): %d further alert(s) were suppressed or undeliverable; check the ingestion logs", q.suppressed)
+		if err := n.Notify(ctx, msg); err == nil {
+			q.suppressed = 0
+		}
+	}
+}
+
+// permanent reports whether err says the message itself can never be sent.
+func permanent(err error) bool {
+	var p interface{ Permanent() bool }
+	return errors.As(err, &p) && p.Permanent()
+}
+
+// removeSymlinks unlinks each brain-relative symlink through an os.Root on
+// the brain dir, after re-checking it is still a symlink. Removing a symlink
+// never touches its target. It returns what was removed and what was not.
+func removeSymlinks(brainDir string, rels []string) (removed, failed []string) {
+	if len(rels) == 0 {
+		return nil, nil
+	}
+	resolved, err := filepath.EvalSymlinks(brainDir)
+	if err != nil {
+		return nil, rels
+	}
+	root, err := os.OpenRoot(resolved)
+	if err != nil {
+		return nil, rels
+	}
+	defer func() { _ = root.Close() }()
+	for _, rel := range rels {
+		fi, err := root.Lstat(rel)
+		if err != nil || fi.Mode()&os.ModeSymlink == 0 {
+			failed = append(failed, rel)
 			continue
 		}
-		reported = key
+		// root.Remove, not os.RemoveAll: if the link is swapped for a
+		// non-empty directory after the Lstat, Remove refuses rather than
+		// deleting it recursively, and it cannot act outside the root.
+		if err := root.Remove(rel); err != nil {
+			failed = append(failed, rel)
+			continue
 		}
+		removed = append(removed, rel)
+	}
+	return removed, failed
 }

ingestion/internal/symlinkguard/symlinkguard_test.go

@@ -3,9 +3,13 @@ package symlinkguard
 import (
 	"context"
 	"errors"
+	"fmt"
 	"os"
 	"path/filepath"
+	"strconv"
+	"strings"
 	"sync"
+	"sync/atomic"
 	"testing"
 	"time"
 
@@ -182,8 +186,313 @@ func TestWatchReportsASetAgainAfterItWasCleared(t *testing.T) {
 	go Watch(ctx, brain, 10*time.Millisecond, n)
 	require.NoError(t, os.Symlink(t.TempDir(), link))
 	require.Eventually(t, func() bool { return n.count() == 1 }, 2*time.Second, 10*time.Millisecond)
-	require.NoError(t, os.Remove(link))
-	time.Sleep(60 * time.Millisecond)
+	// The tripwire removed it; planting it again is a new event.
+	require.Eventually(t, func() bool { _, err := os.Lstat(link); return os.IsNotExist(err) }, 2*time.Second, 10*time.Millisecond)
 	require.NoError(t, os.Symlink(t.TempDir(), link))
 	require.Eventually(t, func() bool { return n.count() == 2 }, 2*time.Second, 10*time.Millisecond)
 }
+
+// hyperguild#157 decision: the runtime tripwire removes a symlink it finds
+// (the link, never its target), so it is readable for at most one interval
+// instead of until a human acts, and the alert says what was removed.
+func TestWatchRemovesAnOutsideFileLinkButNotItsTargetOrRegularFiles(t *testing.T) {
+	brain := t.TempDir()
+	outside := filepath.Join(t.TempDir(), "target.md")
+	require.NoError(t, os.WriteFile(outside, []byte("keep me"), 0o644))
+	require.NoError(t, os.MkdirAll(filepath.Join(brain, "wiki"), 0o755))
+	link := filepath.Join(brain, "wiki", "leak.md")
+	require.NoError(t, os.Symlink(outside, link))
+	regular := filepath.Join(brain, "wiki", "note.md")
+	require.NoError(t, os.WriteFile(regular, []byte("note"), 0o644))
+
+	n := &recNotifier{}
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+	go Watch(ctx, brain, 10*time.Millisecond, n)
+
+	require.Eventually(t, func() bool {
+		_, err := os.Lstat(link)
+		return os.IsNotExist(err)
+	}, 2*time.Second, 10*time.Millisecond, "the symlink is removed")
+	require.Eventually(t, func() bool { return n.count() >= 1 }, 2*time.Second, 10*time.Millisecond)
+	b, err := os.ReadFile(outside)
+	require.NoError(t, err)
+	assert.Equal(t, "keep me", string(b), "the target is untouched")
+	b, err = os.ReadFile(regular)
+	require.NoError(t, err)
+	assert.Equal(t, "note", string(b), "regular files are untouched")
+	n.mu.Lock()
+	defer n.mu.Unlock()
+	assert.Contains(t, n.msgs[0], "removed")
+	assert.Contains(t, n.msgs[0], "wiki/leak.md")
+}
+
+func TestDefaultIntervalIsThirtySeconds(t *testing.T) {
+	assert.Equal(t, 30*time.Second, DefaultInterval)
+}
+
+// PR #160 review: the Lstat re-check and the os.Root confinement in
+// removeSymlinks had no test. Anything that is not a symlink inside the brain
+// is reported failed and nothing is deleted.
+func TestRemoveSymlinksRefusesNonSymlinksEscapesAndPathsThroughLinks(t *testing.T) {
+	brain := t.TempDir()
+	require.NoError(t, os.MkdirAll(filepath.Join(brain, "wiki", "real"), 0o755))
+	regular := filepath.Join(brain, "wiki", "note.md")
+	require.NoError(t, os.WriteFile(regular, []byte("note"), 0o644))
+	inLinked := filepath.Join(brain, "wiki", "real", "x.md")
+	require.NoError(t, os.WriteFile(inLinked, []byte("x"), 0o644))
+	require.NoError(t, os.Symlink(filepath.Join(brain, "wiki", "real"), filepath.Join(brain, "wiki", "linkdir")))
+	outside := filepath.Join(filepath.Dir(brain), "outside-157.md")
+	require.NoError(t, os.WriteFile(outside, []byte("o"), 0o644))
+	t.Cleanup(func() { _ = os.Remove(outside) })
+
+	removed, failed := removeSymlinks(brain, []string{
+		"wiki/note.md", "wiki/linkdir/x.md", "../outside-157.md", "wiki/missing.md",
+	})
+	assert.Empty(t, removed)
+	assert.ElementsMatch(t, []string{"wiki/note.md", "wiki/linkdir/x.md", "../outside-157.md", "wiki/missing.md"}, failed)
+	for _, p := range []string{regular, inLinked, outside} {
+		_, err := os.Stat(p)
+		assert.NoError(t, err, "%s must not be deleted", p)
+	}
+}
+
+// Dir targets and in-brain targets survive too, not only an outside file.
+func TestWatchRemovesDirAndInBrainLinksButNotTheirTargets(t *testing.T) {
+	brain := t.TempDir()
+	outsideDir := t.TempDir()
+	require.NoError(t, os.WriteFile(filepath.Join(outsideDir, "keep.md"), []byte("keep"), 0o644))
+	require.NoError(t, os.MkdirAll(filepath.Join(brain, "wiki"), 0o755))
+	inBrain := filepath.Join(brain, "wiki", "target.md")
+	require.NoError(t, os.WriteFile(inBrain, []byte("t"), 0o644))
+	require.NoError(t, os.Symlink(outsideDir, filepath.Join(brain, "wiki", "dirlink")))
+	require.NoError(t, os.Symlink(inBrain, filepath.Join(brain, "wiki", "inlink.md")))
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+	go Watch(ctx, brain, 10*time.Millisecond, nil)
+	require.Eventually(t, func() bool {
+		r, err := Scan(brain)
+		return err == nil && len(r.Symlinks) == 0
+	}, 2*time.Second, 10*time.Millisecond)
+	for _, p := range []string{filepath.Join(outsideDir, "keep.md"), inBrain} {
+		_, err := os.Stat(p)
+		assert.NoError(t, err, "%s must survive", p)
+	}
+}
+
+// PR #160 review: the planter chooses the paths. An alert names at most
+// maxPathsPerAlert of them and counts the rest, so its size stays bounded.
+func TestWatchAlertNamesBoundedPathsAndCountsTheRest(t *testing.T) {
+	brain := t.TempDir()
+	for i := 0; i < 50; i++ {
+		require.NoError(t, os.Symlink(t.TempDir(), filepath.Join(brain, strings.Repeat("n", 100)+strconv.Itoa(i))))
+	}
+	n := &recNotifier{}
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+	go Watch(ctx, brain, 10*time.Millisecond, n)
+	require.Eventually(t, func() bool { return n.count() >= 1 }, 2*time.Second, 10*time.Millisecond)
+	n.mu.Lock()
+	msg := n.msgs[0]
+	n.mu.Unlock()
+	assert.Contains(t, msg, "removed 50 symlink(s)")
+	assert.Contains(t, msg, "and 40 more")
+	assert.Less(t, len(msg), 2500, "alert size is bounded regardless of how many links were planted")
+}
+
+// rejectNotifier refuses every message containing a marker, forever.
+type rejectNotifier struct {
+	recNotifier
+	reject string
+}
+
+// permanentErr is a rejection the notifier says can never succeed (like ntfy 413).
+type permanentErr struct{}
+
+func (permanentErr) Error() string   { return "413 payload too large" }
+func (permanentErr) Permanent() bool { return true }
+
+func (r *rejectNotifier) Notify(ctx context.Context, m string) error {
+	if strings.Contains(m, r.reject) {
+		r.mu.Lock()
+		r.msgs = append(r.msgs, "REJECTED")
+		r.mu.Unlock()
+		return permanentErr{}
+	}
+	return r.recNotifier.Notify(ctx, m)
+}
+
+// PR #160 review: one message the server will never accept blocked every
+// later alert forever (head-of-line). It is dropped after maxAttempts.
+func TestWatchDropsAMessageThatNeverSendsSoLaterAlertsFlow(t *testing.T) {
+	brain := t.TempDir()
+	n := &rejectNotifier{reject: "POISON"}
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+	go Watch(ctx, brain, 10*time.Millisecond, n)
+	require.NoError(t, os.Symlink(t.TempDir(), filepath.Join(brain, "POISON")))
+	time.Sleep(30 * time.Millisecond)
+	require.NoError(t, os.Symlink(t.TempDir(), filepath.Join(brain, "later")))
+	require.Eventually(t, func() bool {
+		n.mu.Lock()
+		defer n.mu.Unlock()
+		for _, m := range n.msgs {
+			if strings.Contains(m, "later") {
+				return true
+			}
+		}
+		return false
+	}, 3*time.Second, 10*time.Millisecond, "a later alert is delivered despite a poison message")
+}
+
+// toggleNotifier fails while down is set.
+type toggleNotifier struct {
+	recNotifier
+	down atomic.Bool
+}
+
+func (r *toggleNotifier) Notify(ctx context.Context, m string) error {
+	if r.down.Load() {
+		return errors.New("ntfy down")
+	}
+	return r.recNotifier.Notify(ctx, m)
+}
+
+// PR #160 review: an outage queued one message per removal, unbounded. The
+// queue holds at most maxPending messages; the rest become one summary.
+func TestWatchQueueIsBoundedDuringAnOutage(t *testing.T) {
+	brain := t.TempDir()
+	n := &toggleNotifier{}
+	n.down.Store(true)
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+	go Watch(ctx, brain, 5*time.Millisecond, n)
+	for i := 0; i < 30; i++ {
+		link := filepath.Join(brain, "p"+strconv.Itoa(i))
+		require.NoError(t, os.Symlink(t.TempDir(), link))
+		require.Eventually(t, func() bool { _, err := os.Lstat(link); return os.IsNotExist(err) }, 2*time.Second, 2*time.Millisecond)
+	}
+	n.down.Store(false)
+	require.Eventually(t, func() bool { return n.count() >= 1 }, 2*time.Second, 5*time.Millisecond)
+	time.Sleep(100 * time.Millisecond)
+	assert.LessOrEqual(t, n.count(), maxPending+1, "queued alerts plus at most one suppression summary")
+	n.mu.Lock()
+	defer n.mu.Unlock()
+	found := false
+	for _, m := range n.msgs {
+		if strings.Contains(m, "suppressed") {
+			found = true
+		}
+	}
+	assert.True(t, found, "suppressed alerts are counted in a summary")
+}
+
+// Within the bound, every alert queued during an outage is delivered after
+// it, each naming its own removal (PR #160 review: truncating the queue after
+// the first send stayed green).
+func TestWatchDeliversEveryQueuedAlertAfterAShortOutage(t *testing.T) {
+	brain := t.TempDir()
+	n := &toggleNotifier{}
+	n.down.Store(true)
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+	go Watch(ctx, brain, 5*time.Millisecond, n)
+	for _, name := range []string{"qa", "qb", "qc"} {
+		link := filepath.Join(brain, name)
+		require.NoError(t, os.Symlink(t.TempDir(), link))
+		require.Eventually(t, func() bool { _, err := os.Lstat(link); return os.IsNotExist(err) }, 2*time.Second, 2*time.Millisecond)
+	}
+	n.down.Store(false)
+	require.Eventually(t, func() bool { return n.count() == 3 }, 2*time.Second, 5*time.Millisecond)
+	n.mu.Lock()
+	defer n.mu.Unlock()
+	for i, name := range []string{"qa", "qb", "qc"} {
+		assert.Contains(t, n.msgs[i], strconv.Quote(name))
+	}
+}
+
+// PR #160 round 2: dropping after N failures could not tell a poison message
+// from an outage, so a 15-minute outage silently lost the only incident. A
+// message is only dropped when it keeps failing while others send; through
+// an outage nothing is dropped, however long.
+func TestWatchLosesNoAlertThroughALongOutage(t *testing.T) {
+	brain := t.TempDir()
+	n := &toggleNotifier{}
+	n.down.Store(true)
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+	go Watch(ctx, brain, 2*time.Millisecond, n)
+	link := filepath.Join(brain, "only-incident")
+	require.NoError(t, os.Symlink(t.TempDir(), link))
+	require.Eventually(t, func() bool { _, err := os.Lstat(link); return os.IsNotExist(err) }, 2*time.Second, 2*time.Millisecond)
+	time.Sleep(200 * time.Millisecond) // ~100 ticks of outage, far past any attempt limit
+	n.down.Store(false)
+	require.Eventually(t, func() bool { return n.count() >= 1 }, 2*time.Second, 5*time.Millisecond)
+	n.mu.Lock()
+	defer n.mu.Unlock()
+	assert.Contains(t, n.msgs[0], strconv.Quote("only-incident"))
+}
+
+// Every removal is either delivered or counted in the suppression summary:
+// none disappears from the record (round 2: 74 of 300 were in neither).
+func TestWatchAccountsForEveryRemovalDuringAnOutage(t *testing.T) {
+	brain := t.TempDir()
+	n := &toggleNotifier{}
+	n.down.Store(true)
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+	go Watch(ctx, brain, 2*time.Millisecond, n)
+	const plants = 25
+	for i := 0; i < plants; i++ {
+		link := filepath.Join(brain, "acct"+strconv.Itoa(i))
+		require.NoError(t, os.Symlink(t.TempDir(), link))
+		require.Eventually(t, func() bool { _, err := os.Lstat(link); return os.IsNotExist(err) }, 2*time.Second, 1*time.Millisecond)
+	}
+	n.down.Store(false)
+	require.Eventually(t, func() bool {
+		n.mu.Lock()
+		defer n.mu.Unlock()
+		for _, m := range n.msgs {
+			if strings.Contains(m, "suppressed") {
+				return true
+			}
+		}
+		return false
+	}, 2*time.Second, 5*time.Millisecond)
+	time.Sleep(50 * time.Millisecond)
+	n.mu.Lock()
+	defer n.mu.Unlock()
+	delivered, counted := 0, 0
+	for _, m := range n.msgs {
+		if strings.Contains(m, "removed 1 symlink(s)") {
+			delivered++
+		}
+		if i := strings.Index(m, "): "); i >= 0 && strings.Contains(m, "suppressed") {
+			_, _ = fmt.Sscanf(m[i+3:], "%d", &counted)
+		}
+	}
+	assert.Equal(t, plants, delivered+counted, "delivered %d + counted %d must equal %d removals", delivered, counted, plants)
+}
+
+// A poison message is still dropped (others send while it fails), and the
+// drop is counted in the summary.
+func TestWatchDropsAPoisonMessageAndCountsIt(t *testing.T) {
+	brain := t.TempDir()
+	n := &rejectNotifier{reject: "POISON"}
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+	go Watch(ctx, brain, 5*time.Millisecond, n)
+	require.NoError(t, os.Symlink(t.TempDir(), filepath.Join(brain, "POISON")))
+	time.Sleep(20 * time.Millisecond)
+	require.NoError(t, os.Symlink(t.TempDir(), filepath.Join(brain, "fine")))
+	require.Eventually(t, func() bool {
+		n.mu.Lock()
+		defer n.mu.Unlock()
+		for _, m := range n.msgs {
+			if strings.Contains(m, "1 further alert(s)") {
+				return true
+			}
+		}
+		return false
+	}, 3*time.Second, 5*time.Millisecond, "the dropped poison alert is counted in a summary")
+}

ingestion/internal/watcher/watcher_test.go

@@ -8,6 +8,7 @@ import (
 	"os"
 	"path/filepath"
 	"strings"
+	"sync"
 	"sync/atomic"
 	"syscall"
 	"testing"
@@ -343,7 +344,12 @@ func TestProcessDir_DoesNotHangUnderAFIFOSwapRace(t *testing.T) {
 	brainDir := setupBrainDir(t)
 	raw := filepath.Join(brainDir, "raw")
 	var stop atomic.Bool
+	var wg sync.WaitGroup
+	wg.Add(1)
+	// Awaited before t.TempDir's cleanup (hyperguild#162).
+	t.Cleanup(func() { stop.Store(true); wg.Wait() })
 	go func() {
+		defer wg.Done()
 		for i := 0; !stop.Load(); i++ {
 			tmp := filepath.Join(raw, ".swap")
 			_ = os.Remove(tmp)