feat: comprehensive audit + Tier 3 wiring (security/correctness/features)

Multi-agent audit + fix pass covering bugs, security, dead config wiring,
and missing features. All quality gates green: build/vet/test -race/govulncheck/
golangci-lint. Frontend tests 47/47.

SECURITY & CORRECTNESS

- Scanner pipeline fail-closed: cache.go scan timeout now 503 (was goto servePkg);
  scanner.go all-scanners-fail saves ScanStatusError; CheckVulnerabilities blocks
  on missing/error result instead of allowing.
- Scanner argv corrections: govulncheck source-mode + go.mod discovery;
  npm-audit generates lockfile via npm install --package-lock-only --ignore-scripts;
  pip-audit per-extension dispatch (wheel direct / sdist extract+pyproject).
- GHSA version-range filtering implemented (was always-include); url.QueryEscape
  on package name.
- pypi SSRF closed via host allowlist on original_url; url.QueryEscape on
  rewriteURL output.
- Path traversal: cache temp file uses os.CreateTemp; smb keyToPath sanitized;
  filesystem keyToPath returns error + filepath.Abs prefix-verify.
- Goroutine leaks plugged: cache.cleanupWorker stop channel; auth.ValidationCache
  Stop() with sync.Once; ws.unregister buffered with non-blocking send.
- WebSocket double-close panic fixed via sync.Once Client.closeSend().
- Race conditions: gormstore.registryCache sync.RWMutex (was concurrent map
  panic); auth.LastUsedAt no longer mutated under RLock; analytics strict
  lock-ordering invariant (statsMu before downloadsMu).
- Auth validator fails CLOSED on transport/unknown-status errors (was returning
  true,err 'allow cache fallback').
- Credential cache hash full SHA256 (was 8-byte truncate; eliminates 2^32
  collision risk).
- filesystem.fs.used incremented after successful rename (no quota inflation
  race).
- gormstore aggregation events deleted in same tx as insert (no double-counting).
- gormstore.SavePackage uses Updates(map) so zero-value security flags
  (RequiresAuth=false) can be cleared.
- partition_manager validates partition name regex before DROP TABLE.
- vcs/git: version regex validation rejects --option-injection; checkout uses
  --detach -- separator; removed TrimPrefix(repo,'v') that corrupted vault/
  vitess/vim-go.
- WebSocket CheckOrigin allowlist (was return true; CSWSH closed); same-origin
  default + ServerConfig.AllowedOrigins.
- fiber CVE GO-2026-4543 patched (v2.52.10 -> v2.52.12).

FUNCTIONAL FEATURES

- API key DB persistence: APIKeyModel + migration 202604280002, full CRUD,
  async LastUsedAt updates with WaitGroup-tracked goroutines drained on Close.
- Admin bootstrap via GOHOARDER_BOOTSTRAP_ADMIN_KEY env var; idempotent
  (skipped when non-revoked admin already exists).
- Auth middleware factory in pkg/app: RequireAuth, RequireRole, RequirePermission,
  OptionalAuth. Mounted on proxy + read APIs when Auth.Enabled=true.
  DELETE /api/packages/* requires admin role. /health public for k8s probes.
- NFS storage backend: pkg/storage/nfs wraps filesystem with /proc/mounts
  detection (Linux), post-write fsync, read-after-write health probe.
- WebSocket real-time pipeline: pkg/events Broadcaster interface; cache + scanner
  managers emit EventPackageCached / EventPackageDownloaded / EventScanComplete;
  app.go runs 30s EventStatsUpdate ticker. Frontend has full WS client (reconnect
  with exponential backoff 1s->30s, 25s heartbeat, subscriber pattern), Pinia
  realtime store, Dashboard live indicator + activity feed + reactive stats
  override.
- TLS termination: fiber.ListenTLS when Server.TLS.Enabled.
- Pre-warming: Prewarming.Enabled flag wired (was hardcoded false); Interval,
  MaxConcurrent, TopPackages plumbed.
- NetworkConfig wired (timeouts, retry, rate-limit, circuit-breaker).
- AuthConfig.BcryptCost wired.
- Security.Scanners.Static.MaxPackageSize plumbed to cache.io.LimitReader.
- Handlers.{Go,NPM,PyPI}.Enabled gates route mounting.
- Graceful shutdown: authManager.Close drains async writes.

LINT/HYGIENE

- 80 golangci-lint issues resolved: errcheck (Close/Write/RemoveAll wrapped),
  gofmt, gosec G104/G306, govet shadow renames + fieldalignment reorders,
  staticcheck ST1000 pkg comments + QF1003 tagged switches + QF1008 embedded
  field selectors + QF1001 De Morgan's law.

BEHAVIOR CHANGES

- Scanner now fail-closed: deployments without scanner binaries
  (trivy/govulncheck/npm-audit/pip-audit/grype) BLOCK packages instead of
  serving unscanned. Add binaries or plan a future allow-on-scan-error flag.
- WS origin defaults to same-origin; cross-origin dashboards must set
  server.allowed_origins.
- Validator no longer fails open; private packages won't serve from cache when
  validation can't be performed.
- download_events retention dropped from 24h to ~5min (deleted in agg tx).
  Migration 202604280001 purges pre-upgrade events.
- DELETE /api/packages/* requires admin role when auth enabled.

NEW PACKAGES

- pkg/events  - Broadcaster interface
- pkg/storage/nfs  - NFS-aware filesystem wrapper

DEFERRED (out of scope this pass)

- Static scanner implementation (config defines AllowedLicenses/MaxPackageSize/
  BlockSuspicious but pkg/scanner/static/ does not exist).
- AuthConfig.AuditLog wiring (no audit log infra yet).
- CacheConfig.TTLOverrides passthrough (would touch cache.Config beyond
  integrator scope).

pkg/websocket/server.go

@@ -1,3 +1,5 @@
+// Package websocket implements the realtime event broadcasting server used
+// by the dashboard frontend.
 package websocket
 
 import (
@@ -37,6 +39,16 @@ type Client struct {
 	server        *Server
 	subscriptions map[EventType]bool
 	mu            sync.RWMutex
+	closeOnce     sync.Once
+}
+
+// closeSend safely closes the client's send channel exactly once.
+// Safe to call concurrently from multiple goroutines (e.g. run loop
+// unregister and shutdown closeAllClients).
+func (c *Client) closeSend() {
+	c.closeOnce.Do(func() {
+		close(c.send)
+	})
 }
 
 // Server manages WebSocket connections and event broadcasting
@@ -68,7 +80,7 @@ func NewServer(cfg Config) *Server {
 		clients:    make(map[*Client]bool),
 		broadcast:  make(chan Event, 256),
 		register:   make(chan *Client),
-		unregister: make(chan *Client),
+		unregister: make(chan *Client, 256),
 		upgrader: websocket.Upgrader{
 			ReadBufferSize:  cfg.ReadBufferSize,
 			WriteBufferSize: cfg.WriteBufferSize,
@@ -109,7 +121,7 @@ func (s *Server) run(ctx context.Context) {
 			s.mu.Lock()
 			if _, ok := s.clients[client]; ok {
 				delete(s.clients, client)
-				close(client.send)
+				client.closeSend()
 			}
 			s.mu.Unlock()
 			log.Debug().
@@ -147,10 +159,15 @@ func (s *Server) broadcastEvent(event Event) {
 			select {
 			case client.send <- message:
 			default:
-				// Client send buffer full - close connection
-				go func(c *Client) {
-					s.unregister <- c
-				}(client)
+				// Client send buffer full - schedule unregister.
+				// Non-blocking send; unregister chan is buffered, so we
+				// avoid spawning goroutines that pile up under a stuck
+				// run() consumer. If the buffer is full the client will
+				// be cleaned up on its next failed write/ping.
+				select {
+				case s.unregister <- client:
+				default:
+				}
 			}
 		}
 	}
@@ -173,9 +190,11 @@ func (s *Server) pingClients() {
 			time.Now().Add(10*time.Second),
 		); err != nil {
 			log.Debug().Err(err).Msg("Failed to ping client")
-			go func(c *Client) {
-				s.unregister <- c
-			}(client)
+			// Non-blocking send; see broadcastEvent for rationale.
+			select {
+			case s.unregister <- client:
+			default:
+			}
 		}
 	}
 }
@@ -186,8 +205,8 @@ func (s *Server) closeAllClients() {
 	defer s.mu.Unlock()
 
 	for client := range s.clients {
-		client.conn.Close() // #nosec G104 -- Cleanup, error not critical
-		close(client.send)
+		_ = client.conn.Close() // #nosec G104 -- Cleanup, error not critical
+		client.closeSend()
 	}
 	s.clients = make(map[*Client]bool)
 }
@@ -207,6 +226,33 @@ func (s *Server) Broadcast(eventType EventType, data map[string]interface{}) {
 	}
 }
 
+// BroadcastEvent satisfies the events.Broadcaster contract used by
+// cache/scanner sub-systems. It accepts a string event type and an
+// arbitrary payload (typically map[string]any) and forwards to the
+// typed Broadcast path.
+//
+// Non-blocking: enqueues onto the broadcast channel and drops the
+// event if the channel is full. Safe for concurrent use.
+func (s *Server) BroadcastEvent(eventType string, payload any) {
+	var data map[string]interface{}
+	switch p := payload.(type) {
+	case map[string]interface{}:
+		data = p
+	case nil:
+		data = map[string]interface{}{}
+	default:
+		// Wrap unknown payloads so the wire format stays consistent.
+		// Sub-systems are expected to pass map[string]any; this branch
+		// is defensive only.
+		log.Warn().
+			Str("event_type", eventType).
+			Msg("BroadcastEvent received non-map payload; wrapping under 'payload' key")
+		data = map[string]interface{}{"payload": payload}
+	}
+
+	s.Broadcast(EventType(eventType), data)
+}
+
 // HandleWebSocket upgrades HTTP connection to WebSocket
 func (s *Server) HandleWebSocket(w http.ResponseWriter, r *http.Request) {
 	conn, err := s.upgrader.Upgrade(w, r, nil)
@@ -237,7 +283,7 @@ func (s *Server) HandleWebSocket(w http.ResponseWriter, r *http.Request) {
 func (c *Client) readPump() {
 	defer func() {
 		c.server.unregister <- c
-		c.conn.Close() // #nosec G104 -- Cleanup, error not critical
+		_ = c.conn.Close() // #nosec G104 -- Cleanup, error not critical
 	}()
 
 	_ = c.conn.SetReadDeadline(time.Now().Add(60 * time.Second)) // #nosec G104 -- Websocket deadline
@@ -265,7 +311,7 @@ func (c *Client) writePump() {
 	ticker := time.NewTicker(54 * time.Second)
 	defer func() {
 		ticker.Stop()
-		c.conn.Close() // #nosec G104 -- Cleanup, error not critical
+		_ = c.conn.Close() // #nosec G104 -- Cleanup, error not critical
 	}()
 
 	for {

pkg/websocket/server_test.go

@@ -0,0 +1,184 @@
+package websocket
+
+import (
+	"context"
+	"net/http"
+	"net/http/httptest"
+	"net/url"
+	"runtime"
+	"sync"
+	"testing"
+	"time"
+
+	gws "github.com/gorilla/websocket"
+)
+
+// TestCloseSendIdempotent ensures double-close on client.send is safe.
+func TestCloseSendIdempotent(t *testing.T) {
+	tests := []struct {
+		name  string
+		calls int
+	}{
+		{name: "single", calls: 1},
+		{name: "double", calls: 2},
+		{name: "many", calls: 16},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			c := &Client{send: make(chan []byte, 1)}
+			var wg sync.WaitGroup
+			wg.Add(tc.calls)
+			for i := 0; i < tc.calls; i++ {
+				go func() {
+					defer wg.Done()
+					defer func() {
+						if r := recover(); r != nil {
+							t.Errorf("closeSend panicked: %v", r)
+						}
+					}()
+					c.closeSend()
+				}()
+			}
+			wg.Wait()
+		})
+	}
+}
+
+// TestUnregisterChannelBuffered verifies unregister channel is buffered so
+// non-blocking sends from broadcastEvent/pingClients don't block or leak
+// goroutines under a stuck consumer.
+func TestUnregisterChannelBuffered(t *testing.T) {
+	s := NewServer(Config{})
+	if cap(s.unregister) < 256 {
+		t.Fatalf("unregister channel cap=%d, want >=256", cap(s.unregister))
+	}
+}
+
+// TestConcurrentConnectionsCloseNoPanic spawns N concurrent websocket clients,
+// closes them all, and asserts no panic. Also performs a coarse goroutine-leak
+// check (pre/post counts within a small delta).
+func TestConcurrentConnectionsCloseNoPanic(t *testing.T) {
+	startGoros := runtime.NumGoroutine()
+
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+
+	srv := NewServer(Config{})
+	srv.Start(ctx)
+
+	httpSrv := httptest.NewServer(http.HandlerFunc(srv.HandleWebSocket))
+	defer httpSrv.Close()
+
+	wsURL, err := url.Parse(httpSrv.URL)
+	if err != nil {
+		t.Fatalf("parse url: %v", err)
+	}
+	wsURL.Scheme = "ws"
+
+	const N = 32
+	conns := make([]*gws.Conn, 0, N)
+	for i := 0; i < N; i++ {
+		c, _, err := gws.DefaultDialer.Dial(wsURL.String(), nil)
+		if err != nil {
+			t.Fatalf("dial %d: %v", i, err)
+		}
+		conns = append(conns, c)
+	}
+
+	// Allow registrations to settle.
+	deadline := time.Now().Add(2 * time.Second)
+	for time.Now().Before(deadline) {
+		srv.mu.RLock()
+		n := len(srv.clients)
+		srv.mu.RUnlock()
+		if n == N {
+			break
+		}
+		time.Sleep(10 * time.Millisecond)
+	}
+
+	// Concurrent close from many goroutines simulates burst disconnect.
+	var wg sync.WaitGroup
+	for _, c := range conns {
+		wg.Add(1)
+		go func(c *gws.Conn) {
+			defer wg.Done()
+			defer func() {
+				if r := recover(); r != nil {
+					t.Errorf("close panicked: %v", r)
+				}
+			}()
+			_ = c.Close()
+		}(c)
+	}
+	wg.Wait()
+
+	// Trigger shutdown — closeAllClients runs concurrently with any pending
+	// in-loop unregisters. This is the double-close race scenario.
+	cancel()
+
+	// Give run loop and pumps time to exit.
+	time.Sleep(500 * time.Millisecond)
+	runtime.GC()
+
+	endGoros := runtime.NumGoroutine()
+	// Allow some slack: test runtime, pollers, etc. We just want to catch
+	// catastrophic per-connection leaks (would scale with N=32).
+	if endGoros > startGoros+8 {
+		t.Errorf("possible goroutine leak: start=%d end=%d", startGoros, endGoros)
+	}
+}
+
+// TestBroadcastWithFullClientBufferDoesNotLeakGoroutines verifies the
+// non-blocking unregister send path: a client with a full send buffer should
+// be unregistered without spawning per-event goroutines.
+func TestBroadcastWithFullClientBufferDoesNotLeakGoroutines(t *testing.T) {
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+
+	s := NewServer(Config{})
+	// Don't Start — drive run-loop interactions manually so we can test the
+	// non-blocking unregister send without races on the real loop.
+	go s.run(ctx)
+
+	// Inject a fake client with a full send buffer.
+	c := &Client{
+		send:          make(chan []byte, 1),
+		subscriptions: make(map[EventType]bool),
+	}
+	c.send <- []byte("filler") // buffer now full
+	s.register <- c
+
+	// Wait for registration.
+	deadline := time.Now().Add(time.Second)
+	for time.Now().Before(deadline) {
+		s.mu.RLock()
+		_, ok := s.clients[c]
+		s.mu.RUnlock()
+		if ok {
+			break
+		}
+		time.Sleep(5 * time.Millisecond)
+	}
+
+	// Broadcast many events. Buffer is full, so each will hit the default
+	// branch and try to non-blocking-send to s.unregister.
+	for i := 0; i < 1000; i++ {
+		s.Broadcast(EventStatsUpdate, map[string]interface{}{"n": i})
+	}
+
+	// Wait for client to be unregistered.
+	deadline = time.Now().Add(2 * time.Second)
+	for time.Now().Before(deadline) {
+		s.mu.RLock()
+		_, ok := s.clients[c]
+		n := len(s.clients)
+		s.mu.RUnlock()
+		if !ok && n == 0 {
+			return
+		}
+		time.Sleep(10 * time.Millisecond)
+	}
+	t.Fatal("client was not unregistered after broadcast spam with full buffer")
+}