fix(ingestion): build /mcp's bearer and the metrics wrap inside buildMux (#154) (#183)

Co-authored-by: Mathias <mthbqv@gmail.com>

ingestion/cmd/server/jwt_validator_test.go

@@ -12,6 +12,8 @@ import (
 
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
+
+	"github.com/mathiasbq/hyperguild/ingestion/internal/metrics"
 )
 
 // fakeIssuer answers OIDC discovery with 502 for the first fail requests, as
@@ -73,8 +75,37 @@ func TestBuildMuxAnswers503ToAJWTWhileTheIssuerIsUnresolved(t *testing.T) {
 	srv, _ := fakeIssuer(t, 1000)
 	v, err := buildJWTValidator(context.Background(), slog.New(slog.DiscardHandler), srv.URL, "brain")
 	require.NoError(t, err)
-	mux, err := buildMux(restAuth{token: routesTok, validator: v, realm: "brain"}, testHandler(t), nil)
+	mux, err := buildMux(restAuth{token: routesTok, validator: v, realm: "brain"}, testHandler(t),
+		okHandler("mcp"), metrics.New(), nil)
 	require.NoError(t, err)
 	code, _ := do(t, mux, "GET", "/pending", "Bearer aaa.bbb.ccc")
 	assert.Equal(t, http.StatusServiceUnavailable, code)
 }
+
+// PR #183 review: /mcp shares REST's bearer, but only the static token was
+// asserted on it. With the issuer unresolved, a JWT on /mcp answers 503 too,
+// so /mcp's middleware was built with the JWT validator.
+func TestBuildMuxAnswers503ToAJWTOnMCPWhileTheIssuerIsUnresolved(t *testing.T) {
+	srv, _ := fakeIssuer(t, 1000)
+	v, err := buildJWTValidator(context.Background(), slog.New(slog.DiscardHandler), srv.URL, "brain")
+	require.NoError(t, err)
+	mux, err := buildMux(restAuth{token: routesTok, validator: v, realm: "brain"}, testHandler(t),
+		okHandler("mcp"), metrics.New(), nil)
+	require.NoError(t, err)
+	code, _ := do(t, mux, "POST", "/mcp", "Bearer aaa.bbb.ccc")
+	assert.Equal(t, http.StatusServiceUnavailable, code)
+}
+
+// PR #183 review: a 401 on /mcp names the protected-resource metadata in
+// WWW-Authenticate (what the claude.ai OAuth connector discovers from), so
+// /mcp's middleware was built with resourceMetadataURL.
+func TestBuildMuxMCP401CarriesResourceMetadataInWWWAuthenticate(t *testing.T) {
+	const rm = "https://brain.example/.well-known/oauth-protected-resource"
+	mux, err := buildMux(restAuth{token: routesTok, realm: "brain", resourceMetadataURL: rm}, testHandler(t),
+		okHandler("mcp"), metrics.New(), nil)
+	require.NoError(t, err)
+	rec := httptest.NewRecorder()
+	mux.ServeHTTP(rec, httptest.NewRequest(http.MethodPost, "/mcp", nil))
+	require.Equal(t, http.StatusUnauthorized, rec.Code)
+	assert.Contains(t, rec.Header().Get("WWW-Authenticate"), `resource_metadata="`+rm+`"`)
+}

ingestion/cmd/server/main.go

@@ -516,8 +516,6 @@ func main() {
 		}
 	}
 
-	public["/mcp"] = chassisauth.BearerMiddleware(mcpToken, jwtValidator, "brain", resourceMetadataURL, mcpSrv)
-
 	// POST /capture (#53/#54): the uniform capture REST door. Needs a ticket
 	// tracker to file action items, so it only mounts when Gitea is
 	// configured. It reuses the MCP server's graph-wired brain store (one
@@ -541,7 +539,7 @@ func main() {
 		public["POST /capture"] = captureH
 		// Same use-case behind the MCP `capture` tool (#55 relay) so MCP-native
 		// harnesses (claude.ai, Crush, Pi, LLM Council) reach capture through
-		// the existing /mcp OAuth connector. mcpSrv is already wrapped above;
+		// the existing /mcp OAuth connector. buildMux wraps this same mcpSrv;
 		// WithCapture mutates the same instance, so the tool appears live.
 		mcpSrv.WithCapture(captureSvc, jwtValidator, mcpToken, "local-cli", resolver)
 		// hyperguild#97: a governance-file rewrite gets the same trail a
@@ -583,19 +581,18 @@ func main() {
 		os.Exit(1)
 	}
 
-	// /metrics — unauthenticated Prometheus endpoint. kube-prometheus-stack
+	// /metrics — unauthenticated Prometheus endpoint, scraped via the
-	// scrapes it via the ServiceMonitor in k3s/apps/supervisor/. The metrics
+	// ServiceMonitor in k3s/apps/supervisor/. buildMux registers it and wraps
-	// middleware below wraps every other registered handler so it observes
+	// the whole router in the registry's middleware.
-	// real request latency. /metrics itself is excluded from its own
+	//
-	// observation by registering it on the outer mux (post-wrap).
+	// The whole router, /mcp's bearer included, is built in buildMux so its
-	reg := metrics.New()
+	// tests are tests of production wiring (PR #152 review, hyperguild#154).
-	public["GET /metrics"] = http.HandlerFunc(reg.Handler())
+	// The REST API's credential is the same as /mcp's. Untested here: the
-
+	// restAuth fields and the public map this call passes in (e.g. a nil
-	// The whole router, built in one place so the test of buildMux is a test
+	// validator would drop JWT auth on /mcp and REST and the suite stays
-	// of production wiring (PR #152 review). The REST API's credential is the
+	// green), and ListenAndServe serving the result unchanged (hyperguild#182).
-	// same as /mcp's.
+	handler, err := buildMux(restAuth{token: mcpToken, validator: jwtValidator, realm: "brain",
-	mux, err := buildMux(restAuth{token: mcpToken, validator: jwtValidator, realm: "brain",
+		resourceMetadataURL: resourceMetadataURL}, h, mcpSrv, metrics.New(), public)
-		resourceMetadataURL: resourceMetadataURL}, h, public)
 	if err != nil {
 		logger.Error("build router", "err", err)
 		os.Exit(1)
@@ -616,7 +613,7 @@ func main() {
 		"watch_interval", watchIntervalLog,
 		"mcp_enabled", true,
 	)
-	if err := http.ListenAndServe(addr, reg.Middleware(mux)); err != nil {
+	if err := http.ListenAndServe(addr, handler); err != nil {
 		logger.Error("server stopped", "err", err)
 		os.Exit(1)
 	}

ingestion/cmd/server/routes.go

@@ -8,22 +8,22 @@ import (
 	chassisauth "git.d-ma.be/mathias/mcp-chassis/auth"
 
 	"github.com/mathiasbq/hyperguild/ingestion/internal/api"
+	"github.com/mathiasbq/hyperguild/ingestion/internal/metrics"
 )
 
-// publicRoutes is the complete set of patterns allowed on the outer mux,
+// publicRoutes is the complete set of patterns a caller may register on the
-// outside the REST API's auth (hyperguild#149). Each authenticates itself or is
+// outer mux, outside the REST API's auth (hyperguild#149). Each authenticates
-// meant to be open: /mcp (bearer/JWT), /capture (bearer/JWT), the webhook
+// itself or is meant to be open: /capture (bearer/JWT), the webhook (HMAC),
-// (HMAC), /oauth/token (client secret), OAuth metadata and /metrics (open).
+// /oauth/token (client secret) and OAuth metadata (open). /mcp and /metrics
-// A REST pattern added here would beat the authenticated catch-all by
+// are not here: buildMux builds them itself (hyperguild#154). A REST pattern
-// specificity and ship open, which is why buildMux refuses anything else.
+// added here would beat the authenticated catch-all by specificity and ship
+// open, which is why buildMux refuses anything else.
 var publicRoutes = map[string]bool{
-	"/mcp":                      true,
 	"POST /capture":             true,
 	"POST /webhooks/brain-sync": true,
 	"GET /.well-known/oauth-protected-resource":   true,
 	"GET /.well-known/oauth-authorization-server": true,
 	"POST /oauth/token":                           true,
-	"GET /metrics":                                true,
 }
 
 // restAuth is the credential the REST API checks: the same as /mcp's.
@@ -34,16 +34,25 @@ type restAuth struct {
 	resourceMetadataURL string
 }
 
-// buildMux is the server's whole router. The REST API is mounted at "/"
+// buildMux builds the handler main passes to ListenAndServe unchanged. The
-// behind bearer auth that buildMux constructs itself, so a caller cannot pass
+// REST API is mounted at "/" and /mcp is registered, both behind bearer auth
-// a no-op; public handlers are registered as given, and only for patterns on
+// that buildMux constructs itself from auth (one credential), so a caller
-// the publicRoutes allowlist. It returns an http.Handler, not the
+// cannot pass a no-op. GET /metrics serves reg, and the whole router is
-// *ServeMux, so nothing can register a route on the result afterwards and
+// wrapped in reg's middleware. Other public handlers are registered as given,
-// bypass the allowlist.
+// and only for patterns on the publicRoutes allowlist. It returns an
-func buildMux(auth restAuth, h *api.Handler, public map[string]http.Handler) (http.Handler, error) {
+// http.Handler, not the *ServeMux, so nothing can register a route on the
+// result. What stays untested is main's call site: the restAuth fields and
+// the public map main passes in, and ListenAndServe serving this handler
+// unchanged (hyperguild#182). The tests build their own arguments.
+func buildMux(
+	auth restAuth, h *api.Handler, mcpHandler http.Handler, reg *metrics.Registry, public map[string]http.Handler,
+) (http.Handler, error) {
 	if auth.token == "" {
 		return nil, errors.New("buildMux: REST auth needs a token")
 	}
+	if mcpHandler == nil || reg == nil {
+		return nil, errors.New("buildMux: needs the MCP handler and the metrics registry")
+	}
 	mux := http.NewServeMux()
 	for pattern, handler := range public {
 		if !publicRoutes[pattern] {
@@ -51,10 +60,15 @@ func buildMux(auth restAuth, h *api.Handler, public map[string]http.Handler) (ht
 		}
 		mux.Handle(pattern, handler)
 	}
-	mountREST(mux, h, func(next http.Handler) http.Handler {
+	bearer := func(next http.Handler) http.Handler {
 		return chassisauth.BearerMiddleware(auth.token, auth.validator, auth.realm, auth.resourceMetadataURL, next)
-	})
+	}
-	return mux, nil
+	mux.Handle("/mcp", bearer(mcpHandler))
+	// /metrics is registered on the mux and so observed by its own
+	// middleware, like any route.
+	mux.Handle("GET /metrics", reg.Handler())
+	mountREST(mux, h, bearer)
+	return reg.Middleware(mux), nil
 }
 
 // mountREST registers the REST API on its own mux and mounts it at "/" behind

ingestion/cmd/server/routes_observability_test.go

@@ -9,8 +9,6 @@ import (
 
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
-
-	"github.com/mathiasbq/hyperguild/ingestion/internal/metrics"
 )
 
 // wrongTokSentinel is a presented credential distinctive enough that any
@@ -53,8 +51,8 @@ func TestBuildMuxAcceptedLogLineOmitsTheBareStaticToken(t *testing.T) {
 // An authenticated REST request is labelled with the REST mux's pattern:
 // the inner mux sets it on the same *http.Request the outer middleware reads.
 func TestMetricsLabelAnAuthenticatedRESTRequestWithItsRoutePattern(t *testing.T) {
-	reg := metrics.New()
+	mux, reg := testMuxAndRegistry(t)
-	code, _ := do(t, reg.Middleware(testMux(t)), "GET", "/pending", "Bearer "+routesTok)
+	code, _ := do(t, mux, "GET", "/pending", "Bearer "+routesTok)
 	require.Equal(t, http.StatusOK, code)
 
 	rec := httptest.NewRecorder()
@@ -67,8 +65,8 @@ func TestMetricsLabelAnAuthenticatedRESTRequestWithItsRoutePattern(t *testing.T)
 // so it was labelled path="/" and a 4xx on /query could not be told from one
 // on /write. It is now labelled with the REST pattern it was addressed to.
 func TestMetricsLabelARejectedRESTRequestWithItsRoutePattern(t *testing.T) {
-	reg := metrics.New()
+	mux, reg := testMuxAndRegistry(t)
-	code, _ := do(t, reg.Middleware(testMux(t)), "POST", "/query", "")
+	code, _ := do(t, mux, "POST", "/query", "")
 	require.Equal(t, http.StatusUnauthorized, code)
 
 	rec := httptest.NewRecorder()
@@ -79,8 +77,8 @@ func TestMetricsLabelARejectedRESTRequestWithItsRoutePattern(t *testing.T) {
 // A path no REST route matches keeps the catch-all label, so an attacker
 // probing random URLs cannot grow the series set.
 func TestMetricsLabelARejectedUnknownPathWithTheCatchAll(t *testing.T) {
-	reg := metrics.New()
+	mux, reg := testMuxAndRegistry(t)
-	code, _ := do(t, reg.Middleware(testMux(t)), "POST", "/no-such-route-zq", "")
+	code, _ := do(t, mux, "POST", "/no-such-route-zq", "")
 	require.Equal(t, http.StatusUnauthorized, code)
 
 	rec := httptest.NewRecorder()

ingestion/cmd/server/routes_test.go

@@ -12,6 +12,7 @@ import (
 	"github.com/stretchr/testify/require"
 
 	"github.com/mathiasbq/hyperguild/ingestion/internal/api"
+	"github.com/mathiasbq/hyperguild/ingestion/internal/metrics"
 	"github.com/mathiasbq/hyperguild/ingestion/internal/pipeline"
 )
 
@@ -21,20 +22,29 @@ func okHandler(body string) http.Handler {
 	return http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) { _, _ = w.Write([]byte(body)) })
 }
 
-// testMux builds the router exactly as main does: buildMux with the full
+// testMux builds the router through buildMux, the function main calls, with
-// public set main registers (stub handlers) and real bearer config.
+// every allowlisted public pattern (stub handlers), a stub MCP server and a
-// hyperguild#149: the REST API is authenticated as a whole; the PR #152
+// static token only: no JWT validator and no resourceMetadataURL (the tests
-// review showed a test of a COPY of the router let main.go regress unseen.
+// in jwt_validator_test.go pass those). It does not use main's own restAuth
+// or public map (hyperguild#182). hyperguild#149: the REST API is
+// authenticated as a whole; the PR #152 review showed a test of a COPY of
+// the router let main.go regress unseen.
 func testMux(t *testing.T) http.Handler {
 	t.Helper()
-	h := testHandler(t)
+	mux, _ := testMuxAndRegistry(t)
+	return mux
+}
+
+func testMuxAndRegistry(t *testing.T) (http.Handler, *metrics.Registry) {
+	t.Helper()
 	public := map[string]http.Handler{}
 	for p := range publicRoutes {
 		public[p] = okHandler("public " + p)
 	}
-	mux, err := buildMux(restAuth{token: routesTok, realm: "brain"}, h, public)
+	reg := metrics.New()
+	mux, err := buildMux(restAuth{token: routesTok, realm: "brain"}, testHandler(t), okHandler("mcp"), reg, public)
 	require.NoError(t, err)
-	return mux
+	return mux, reg
 }
 
 func testHandler(t *testing.T) *api.Handler {
@@ -79,8 +89,9 @@ func TestBuildMuxAuthenticatesAnUnregisteredPath(t *testing.T) {
 }
 
 // buildMux registers public handlers as given, without adding auth: each
-// public route owns its own (bearer on /mcp, HMAC on the webhook, client
+// public route owns its own (bearer/JWT on /capture, HMAC on the webhook,
-// secret on /oauth/token) or is meant to be open (/metrics, OAuth metadata).
+// client secret on /oauth/token) or is meant to be open (OAuth metadata).
+// /mcp and /metrics are not public routes: buildMux builds them itself.
 func TestBuildMuxDoesNotWrapPublicRoutes(t *testing.T) {
 	r := testMux(t)
 	for p := range publicRoutes {
@@ -98,14 +109,53 @@ func TestBuildMuxDoesNotWrapPublicRoutes(t *testing.T) {
 // over the authenticated catch-all (more specific pattern), so it is refused.
 func TestBuildMuxRefusesAPublicPatternOffTheAllowlist(t *testing.T) {
 	h := api.NewHandler(t.TempDir(), slog.New(slog.NewTextHandler(os.Stderr, nil)), pipeline.Config{})
-	for _, p := range []string{"POST /query", "GET /pending", "/", "POST /metrics"} {
+	// "/mcp" and "GET /metrics" are built by buildMux itself (hyperguild#154):
-		_, err := buildMux(restAuth{token: routesTok, realm: "brain"}, h, map[string]http.Handler{p: okHandler("x")})
+	// passing either as public would register them unauthenticated or twice.
+	for _, p := range []string{"POST /query", "GET /pending", "/", "POST /metrics", "/mcp", "GET /metrics"} {
+		_, err := buildMux(restAuth{token: routesTok, realm: "brain"}, h, okHandler("mcp"), metrics.New(),
+			map[string]http.Handler{p: okHandler("x")})
 		assert.Error(t, err, "public pattern %q must be refused", p)
 	}
 }
 
 func TestBuildMuxRefusesAnEmptyToken(t *testing.T) {
 	h := api.NewHandler(t.TempDir(), slog.New(slog.NewTextHandler(os.Stderr, nil)), pipeline.Config{})
-	_, err := buildMux(restAuth{token: "", realm: "brain"}, h, nil)
+	_, err := buildMux(restAuth{token: "", realm: "brain"}, h, okHandler("mcp"), metrics.New(), nil)
 	assert.Error(t, err)
 }
+
+// hyperguild#154: /mcp's bearer was built in main.go, untested; the buildMux
+// tests used stubs. buildMux now wraps the MCP handler itself, with the same
+// credential as the REST API.
+func TestBuildMuxRefusesMCPWithoutAValidToken(t *testing.T) {
+	r := testMux(t)
+	for _, authz := range []string{"", "Bearer wrong"} {
+		code, body := do(t, r, "POST", "/mcp", authz)
+		assert.Equal(t, http.StatusUnauthorized, code, "authz %q", authz)
+		assert.NotEqual(t, "mcp", body)
+	}
+	code, body := do(t, r, "POST", "/mcp", "Bearer "+routesTok)
+	assert.Equal(t, http.StatusOK, code)
+	assert.Equal(t, "mcp", body)
+}
+
+func TestBuildMuxRefusesANilMCPHandlerOrRegistry(t *testing.T) {
+	auth := restAuth{token: routesTok, realm: "brain"}
+	_, err := buildMux(auth, testHandler(t), nil, metrics.New(), nil)
+	assert.Error(t, err)
+	_, err = buildMux(auth, testHandler(t), okHandler("mcp"), nil, nil)
+	assert.Error(t, err)
+}
+
+// hyperguild#154: main wrapped buildMux's result in the metrics middleware
+// before ListenAndServe, untested. buildMux now returns the handler main
+// serves as-is, metrics included: a request through it is observed and
+// exposed on its own GET /metrics.
+func TestBuildMuxObservesRequestsAndServesMetrics(t *testing.T) {
+	r := testMux(t)
+	code, _ := do(t, r, "GET", "/pending", "Bearer "+routesTok)
+	require.Equal(t, http.StatusOK, code)
+	code, body := do(t, r, "GET", "/metrics", "")
+	require.Equal(t, http.StatusOK, code)
+	assert.Contains(t, body, `path="GET /pending"`)
+}

ingestion/internal/symlinkguard/symlinkguard.go

@@ -18,9 +18,14 @@ import (
 	"sort"
 	"strconv"
 	"strings"
+	"sync"
+	"sync/atomic"
 	"time"
 )
 
+// shutdownFlushTimeout bounds the final flush when Watch stops.
+var shutdownFlushTimeout = 5 * time.Second
+
 // DefaultInterval is used when Watch is given a non-positive interval.
 const DefaultInterval = 30 * time.Second
 
@@ -133,13 +138,32 @@ func Scan(brainDir string) (Result, error) {
 // 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.
+//
+// Delivery runs in its own goroutine (hyperguild#164): a hung notifier, with
+// a full queue, blocked one flush for several notify timeouts, and while the
+// scan waited on it a planted link stayed readable for several intervals.
+// The scan hands alerts over without blocking; any that do not fit are
+// counted in the summary. When ctx ends, Watch returns after a final flush
+// of what was handed over (bounded by shutdownFlushTimeout); alerts that
+// flush cannot deliver are counted in one error log line instead. That
+// shutdown path only runs for a caller that cancels ctx: cmd/server passes
+// context.Background(), so in production the flush on handoff is what keeps
+// alerts from waiting on a restart.
 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{}
+	alerts := make(chan string, maxPending)
+	var overflow atomic.Int64
+	var wg sync.WaitGroup
+	wg.Add(1)
+	go func() { defer wg.Done(); deliver(ctx, interval, n, alerts, &overflow) }()
+	// The scan is the only sender: closing alerts tells deliver to drain and
+	// flush once more, and Watch returns only after it has.
+	defer func() { close(alerts); wg.Wait() }()
+	q := alertSender{alerts: alerts, overflow: &overflow}
 	reported := ""
 	for {
 		select {
@@ -168,7 +192,76 @@ func Watch(ctx context.Context, brainDir string, interval time.Duration, n Notif
 					". Readers may follow symlinks out of the brain; fix them.")
 			}
 		}
-		q.flush(ctx, n)
+	}
+}
+
+// alertSender hands alerts to deliver without ever blocking the scan; an
+// alert that does not fit is counted in overflow, which deliver adds to the
+// summary.
+type alertSender struct {
+	alerts   chan<- string
+	overflow *atomic.Int64
+}
+
+func (s alertSender) push(msg string) {
+	select {
+	case s.alerts <- msg:
+	default:
+		s.overflow.Add(1)
+	}
+}
+
+// deliver owns the alertQueue. It flushes as soon as an alert is handed
+// over, and again every interval while any are pending (outage retries).
+// When alerts is closed (Watch is stopping) it drains it and flushes once
+// more, bounded by shutdownFlushTimeout, so an alert handed over just before
+// shutdown is not lost (PR #180 review). Whatever that flush cannot deliver
+// is logged as one error with its counts, never dropped silently.
+func deliver(ctx context.Context, interval time.Duration, n Notifier, alerts <-chan string, overflow *atomic.Int64) {
+	t := time.NewTicker(interval)
+	defer t.Stop()
+	q := &alertQueue{}
+	// drain moves everything handed over into the queue; it reports false
+	// once alerts is closed.
+	drain := func() bool {
+		defer func() { q.suppressed += int(overflow.Swap(0)) }()
+		for {
+			select {
+			case msg, ok := <-alerts:
+				if !ok {
+					return false
+				}
+				q.push(msg)
+			default:
+				return true
+			}
+		}
+	}
+	for {
+		select {
+		case msg, ok := <-alerts:
+			if ok {
+				q.push(msg)
+				ok = drain()
+			}
+			if !ok {
+				q.suppressed += int(overflow.Swap(0))
+				fctx, cancel := context.WithTimeout(context.WithoutCancel(ctx), shutdownFlushTimeout)
+				q.flush(fctx, n)
+				cancel()
+				if len(q.items) > 0 || q.suppressed > 0 {
+					slog.Error("symlink tripwire alerts not delivered at shutdown",
+						"undelivered", len(q.items), "suppressed", q.suppressed)
+				}
+				return
+			}
+			q.flush(ctx, n)
+		case <-t.C:
+			if !drain() {
+				continue // the closed channel is handled by the case above
+			}
+			q.flush(ctx, n)
+		}
 	}
 }
 

ingestion/internal/symlinkguard/symlinkguard_test.go

@@ -1,10 +1,12 @@
 package symlinkguard
 
 import (
+	"bytes"
 	"context"
 	"errors"
 	"fmt"
 	"os"
+	"log/slog"
 	"path/filepath"
 	"strconv"
 	"strings"
@@ -345,13 +347,19 @@ func TestWatchDropsAMessageThatNeverSendsSoLaterAlertsFlow(t *testing.T) {
 	}, 3*time.Second, 10*time.Millisecond, "a later alert is delivered despite a poison message")
 }
 
-// toggleNotifier fails while down is set.
+// toggleNotifier fails while down is set. Like the real audit.NtfyNotifier
+// (http.NewRequestWithContext), it also fails, without recording, when its
+// ctx is already done (PR #180 round 2: notifiers that ignored ctx let a
+// final flush on the cancelled ctx pass).
 type toggleNotifier struct {
 	recNotifier
 	down atomic.Bool
 }
 
 func (r *toggleNotifier) Notify(ctx context.Context, m string) error {
+	if err := ctx.Err(); err != nil {
+		return err
+	}
 	if r.down.Load() {
 		return errors.New("ntfy down")
 	}
@@ -496,3 +504,271 @@ func TestWatchDropsAPoisonMessageAndCountsIt(t *testing.T) {
 		return false
 	}, 3*time.Second, 5*time.Millisecond, "the dropped poison alert is counted in a summary")
 }
+
+// hungNotifier answers every Notify only after latency (or when ctx ends),
+// with an outage error: an ntfy that accepts connections but never responds
+// until its timeout.
+type hungNotifier struct{ latency time.Duration }
+
+func (h hungNotifier) Notify(ctx context.Context, _ string) error {
+	select {
+	case <-time.After(h.latency):
+	case <-ctx.Done():
+	}
+	return errors.New("ntfy timeout")
+}
+
+func waitGone(t *testing.T, link string) {
+	t.Helper()
+	require.Eventually(t, func() bool { _, err := os.Lstat(link); return os.IsNotExist(err) }, 10*time.Second, time.Millisecond)
+}
+
+// hyperguild#164: alerts were flushed in the scan goroutine, so with a hung
+// ntfy and a full queue one flush (6 x notify timeout) delayed the next scan.
+// The notifier here is slower than the interval, as in production (10 s
+// timeout vs 30 s interval with 6 messages), so the alert handoff is full:
+// a scan that waited on it would keep a link for seconds.
+func TestWatchLinkLifetimeStaysNearOneIntervalWithANotifierSlowerThanTheInterval(t *testing.T) {
+	const interval = 100 * time.Millisecond
+	brain := t.TempDir()
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+	go Watch(ctx, brain, interval, hungNotifier{latency: 400 * time.Millisecond})
+
+	for i := 0; i < 2*maxPending+3; i++ { // overfill queue and handoff
+		link := filepath.Join(brain, "fill"+strconv.Itoa(i))
+		require.NoError(t, os.Symlink(t.TempDir(), link))
+		waitGone(t, link)
+	}
+	var worst time.Duration
+	for i := 0; i < 6; i++ {
+		link := filepath.Join(brain, "probe"+strconv.Itoa(i))
+		require.NoError(t, os.Symlink(t.TempDir(), link))
+		planted := time.Now()
+		waitGone(t, link)
+		worst = max(worst, time.Since(planted))
+	}
+	assert.LessOrEqual(t, worst, interval+interval/2, "worst link lifetime %v with a hung notifier", worst)
+}
+
+// Review of PR #180: delivery waited for its own tick, about one interval
+// after the removal. An alert is now flushed as soon as it is handed over.
+func TestWatchDeliversARemovalAlertWellWithinOneInterval(t *testing.T) {
+	const interval = 400 * time.Millisecond
+	brain := t.TempDir()
+	n := &recNotifier{}
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+	go Watch(ctx, brain, interval, n)
+	link := filepath.Join(brain, "prompt")
+	require.NoError(t, os.Symlink(t.TempDir(), link))
+	waitGone(t, link)
+	removed := time.Now()
+	require.Eventually(t, func() bool { return n.count() == 1 }, 2*time.Second, time.Millisecond)
+	assert.Less(t, time.Since(removed), interval/4)
+}
+
+// Review of PR #180: on ctx end, deliver returned without draining what the
+// scan had handed over, so a removal just before shutdown was neither
+// delivered nor counted. Watch now drains and flushes before returning.
+func TestWatchDeliversAnAlertHandedOverJustBeforeShutdown(t *testing.T) {
+	for i := 0; i < 20; i++ {
+		brain := t.TempDir()
+		n := &recNotifier{}
+		ctx, cancel := context.WithCancel(context.Background())
+		done := make(chan struct{})
+		go func() { Watch(ctx, brain, 5*time.Millisecond, n); close(done) }()
+		link := filepath.Join(brain, "late")
+		require.NoError(t, os.Symlink(t.TempDir(), link))
+		waitGone(t, link)
+		cancel()
+		<-done
+		require.Equal(t, 1, n.count(), "iteration %d: the removal alert was lost at shutdown", i)
+	}
+}
+
+// blockingNotifier holds every Notify until its ctx ends, tracking calls in
+// flight.
+type blockingNotifier struct{ inFlight atomic.Int32 }
+
+func (b *blockingNotifier) Notify(ctx context.Context, _ string) error {
+	b.inFlight.Add(1)
+	defer b.inFlight.Add(-1)
+	<-ctx.Done()
+	return ctx.Err()
+}
+
+// Watch does not return while its delivery goroutine is still notifying.
+func TestWatchReturnsOnlyAfterDeliveryHasStopped(t *testing.T) {
+	prev := shutdownFlushTimeout
+	shutdownFlushTimeout = 50 * time.Millisecond
+	t.Cleanup(func() { shutdownFlushTimeout = prev })
+	brain := t.TempDir()
+	n := &blockingNotifier{}
+	ctx, cancel := context.WithCancel(context.Background())
+	done := make(chan struct{})
+	go func() { Watch(ctx, brain, 5*time.Millisecond, n); close(done) }()
+	link := filepath.Join(brain, "x")
+	require.NoError(t, os.Symlink(t.TempDir(), link))
+	waitGone(t, link)
+	require.Eventually(t, func() bool { return n.inFlight.Load() == 1 }, 2*time.Second, time.Millisecond)
+	cancel()
+	select {
+	case <-done:
+	case <-time.After(5 * time.Second):
+		t.Fatal("Watch did not return after cancel")
+	}
+	assert.Zero(t, n.inFlight.Load(), "Notify still running after Watch returned")
+}
+
+// slowToggleNotifier fails slowly (a hung ntfy) while down, then records.
+type slowToggleNotifier struct {
+	recNotifier
+	down    atomic.Bool
+	latency time.Duration
+}
+
+func (r *slowToggleNotifier) Notify(ctx context.Context, m string) error {
+	if r.down.Load() {
+		select {
+		case <-time.After(r.latency):
+		case <-ctx.Done():
+		}
+		return errors.New("ntfy timeout")
+	}
+	return r.recNotifier.Notify(ctx, m)
+}
+
+// Review of PR #180: with a slow notifier the scan's handoff overflows, and
+// those alerts exist only in the overflow counter. Every removal is still
+// delivered or counted in the summary.
+func TestWatchCountsHandoffOverflowWithASlowNotifier(t *testing.T) {
+	brain := t.TempDir()
+	n := &slowToggleNotifier{latency: 30 * time.Millisecond}
+	n.down.Store(true)
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+	go Watch(ctx, brain, 2*time.Millisecond, n)
+	const plants = 40
+	for i := 0; i < plants; i++ {
+		link := filepath.Join(brain, "ov"+strconv.Itoa(i))
+		require.NoError(t, os.Symlink(t.TempDir(), link))
+		waitGone(t, link)
+	}
+	n.down.Store(false)
+	delivered, counted := 0, 0
+	require.Eventually(t, func() bool {
+		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") {
+				var c int
+				_, _ = fmt.Sscanf(m[i+3:], "%d", &c)
+				counted += c
+			}
+		}
+		return delivered+counted == plants
+	}, 5*time.Second, 10*time.Millisecond, "delivered %d + counted %d of %d", delivered, counted, plants)
+}
+
+// An alert still queued when Watch stops (its send failed, the notifier has
+// since recovered, the retry tick has not come) is flushed once more before
+// Watch returns, on a ctx that is not the cancelled one: toggleNotifier fails
+// on a done ctx, as the real ntfy notifier does.
+func TestWatchFinalFlushSendsAQueuedAlertOnACtxThatOutlivesTheCancel(t *testing.T) {
+	brain := t.TempDir()
+	n := &toggleNotifier{}
+	n.down.Store(true)
+	ctx, cancel := context.WithCancel(context.Background())
+	done := make(chan struct{})
+	go func() { Watch(ctx, brain, 300*time.Millisecond, n); close(done) }()
+	link := filepath.Join(brain, "queued")
+	require.NoError(t, os.Symlink(t.TempDir(), link))
+	waitGone(t, link)
+	time.Sleep(20 * time.Millisecond) // let the immediate send fail
+	n.down.Store(false)
+	cancel()
+	<-done
+	assert.Equal(t, 1, n.count(), "the queued alert was not flushed at shutdown")
+}
+
+// lockedBuffer is a goroutine-safe log sink: Watch goroutines of earlier
+// tests may still log while one test captures.
+type lockedBuffer struct {
+	mu  sync.Mutex
+	buf bytes.Buffer
+}
+
+func (b *lockedBuffer) Write(p []byte) (int, error) {
+	b.mu.Lock()
+	defer b.mu.Unlock()
+	return b.buf.Write(p)
+}
+
+func (b *lockedBuffer) String() string { b.mu.Lock(); defer b.mu.Unlock(); return b.buf.String() }
+
+func captureLog(t *testing.T) *lockedBuffer {
+	t.Helper()
+	b := &lockedBuffer{}
+	prev := slog.Default()
+	slog.SetDefault(slog.New(slog.NewTextHandler(b, nil)))
+	t.Cleanup(func() { slog.SetDefault(prev) })
+	return b
+}
+
+// PR #180 round 2: when the final flush fails (ntfy hung at shutdown) the
+// undelivered alerts vanished with only "will retry" warnings. Watch now logs
+// one error with both counts.
+func TestWatchLogsUndeliveredAndSuppressedCountsWhenTheFinalFlushFails(t *testing.T) {
+	prev := shutdownFlushTimeout
+	shutdownFlushTimeout = 50 * time.Millisecond
+	t.Cleanup(func() { shutdownFlushTimeout = prev })
+	logs := captureLog(t)
+	brain := t.TempDir()
+	ctx, cancel := context.WithCancel(context.Background())
+	done := make(chan struct{})
+	go func() { Watch(ctx, brain, 5*time.Millisecond, hungNotifier{latency: time.Hour}); close(done) }()
+	link := filepath.Join(brain, "lost")
+	require.NoError(t, os.Symlink(t.TempDir(), link))
+	waitGone(t, link)
+	cancel()
+	<-done
+	out := logs.String()
+	assert.Contains(t, out, "level=ERROR")
+	assert.Contains(t, out, `msg="symlink tripwire alerts not delivered at shutdown"`)
+	assert.Contains(t, out, "undelivered=1")
+	assert.Contains(t, out, "suppressed=0")
+}
+
+// deliver, given a handoff that already holds alerts and is closed, moves them
+// into the queue and runs the final flush, even though the ctx is cancelled:
+// the drain-on-close path, without depending on when the close lands.
+func TestDeliverFlushesAlertsDrainedFromAClosedHandoffOnACancelledCtx(t *testing.T) {
+	alerts := make(chan string, 3)
+	alerts <- "a1"
+	alerts <- "a2"
+	close(alerts)
+	ctx, cancel := context.WithCancel(context.Background())
+	cancel()
+	n := &toggleNotifier{}
+	var overflow atomic.Int64
+	deliver(ctx, time.Hour, n, alerts, &overflow)
+	assert.Equal(t, []string{"a1", "a2"}, n.msgs)
+}
+
+// PR #180 round 2: when the first receive after close finds the handoff empty,
+// overflow counted by the scan was never moved into the summary.
+func TestDeliverCountsOverflowWhenTheHandoffIsClosedAndEmpty(t *testing.T) {
+	alerts := make(chan string)
+	close(alerts)
+	n := &recNotifier{}
+	var overflow atomic.Int64
+	overflow.Store(3)
+	deliver(context.Background(), time.Hour, n, alerts, &overflow)
+	require.Len(t, n.msgs, 1)
+	assert.Contains(t, n.msgs[0], "3 further alert(s)")
+}