diff --git a/cmd/archivmail/cmd_import.go b/cmd/archivmail/cmd_import.go
index 0c8307e..9a46eeb 100644
--- a/cmd/archivmail/cmd_import.go
+++ b/cmd/archivmail/cmd_import.go
@@ -285,6 +285,7 @@ Commands:
   import-piler   Aus mailpiler migrieren (pilerexport oder direkte Store-Methode)
   export         E-Mails exportieren (EML, MBOX)
   reindex        Index neu aufbauen (alle oder pro Mandant)
+  purge          Mails mit abgelaufener Aufbewahrungsfrist löschen (cron-fähig)
   recompress     Bestehende Mails nachträglich gzip-komprimieren
   rethread       Thread-IDs rückwirkend aus In-Reply-To/References befüllen
   ocr-reprocess  OCR für Anhänge nachholen (alle oder pro Mandant/Status)
diff --git a/cmd/archivmail/cmd_purge.go b/cmd/archivmail/cmd_purge.go
new file mode 100644
index 0000000..d497c17
--- /dev/null
+++ b/cmd/archivmail/cmd_purge.go
@@ -0,0 +1,127 @@
+package main
+
+import (
+	"context"
+	"flag"
+	"log/slog"
+	"os"
+
+	"archivmail/config"
+	"archivmail/internal/audit"
+	"archivmail/internal/index"
+	"archivmail/internal/storage"
+)
+
+// runPurge deletes all mails whose retain_until has passed, removes them
+// from the search index, and writes one audit log entry per deleted mail
+// (GoBD-Nachvollziehbarkeit). Intended to be cron-driven (PROJ-56c), mirrors
+// the manual /api/admin/purge endpoint but adds index cleanup + audit trail,
+// which the plain Store.Purge() helper intentionally does not do.
+//
+// Usage: archivmail purge [-config /path/to/config.yml] [-dry-run]
+func runPurge(args []string) {
+	fs := flag.NewFlagSet("purge", flag.ExitOnError)
+	configPath := fs.String("config", "/etc/archivmail/config.yml", "path to config file")
+	dryRun := fs.Bool("dry-run", false, "list expired mails without deleting them")
+	fs.Parse(args)
+
+	logger := slog.New(slog.NewTextHandler(os.Stdout, &slog.HandlerOptions{Level: slog.LevelInfo}))
+
+	cfg, err := config.Load(*configPath)
+	if err != nil {
+		logger.Error("failed to load config", "err", err)
+		os.Exit(1)
+	}
+
+	storeCfg := storage.Config{
+		Dir:             cfg.Storage.StorePath,
+		Keyfile:         cfg.Storage.Keyfile,
+		DSN:             cfg.Database.DSN(),
+		CompressEnabled: cfg.Storage.Compress,
+	}
+	mailStore, err := storage.New(storeCfg)
+	if err != nil {
+		logger.Error("storage init failed", "err", err)
+		os.Exit(1)
+	}
+	defer mailStore.Close()
+
+	ctx := context.Background()
+
+	ids, err := mailStore.ListExpiredMailIDs(ctx)
+	if err != nil {
+		logger.Error("purge: list expired failed", "err", err)
+		os.Exit(1)
+	}
+	if len(ids) == 0 {
+		logger.Info("purge: nothing to do, no expired mails")
+		return
+	}
+
+	if *dryRun {
+		logger.Info("purge: dry-run, would delete", "count", len(ids))
+		for _, id := range ids {
+			logger.Info("purge: dry-run candidate", "id", id)
+		}
+		return
+	}
+
+	// Index + audit log are best-effort extras (PROJ-56c); the OCR/index
+	// backends and audit DB can be unreachable without that blocking the
+	// actual deletion, which is what GoBD-Löschsperre/Retention requires.
+	var idxMgr index.TenantIndexer
+	indexBackend := cfg.Index.Backend
+	if indexBackend == "manticore" {
+		dsn := cfg.Index.ManticoreDSN
+		if dsn == "" {
+			dsn = "manticore@tcp(127.0.0.1:9306)/"
+		}
+		if m, err := index.NewManticoreTenantManager(dsn); err == nil {
+			idxMgr = m
+			defer m.Close()
+		} else {
+			logger.Warn("purge: index init failed, skipping index cleanup", "err", err)
+		}
+	}
+
+	var audlog *audit.Logger
+	if a, err := audit.New(cfg.Database.DSN(), cfg.Audit.ResolvedLogPath(), logger); err == nil {
+		audlog = a
+		defer audlog.Close()
+	} else {
+		logger.Warn("purge: audit log init failed, deletions will not be audited", "err", err)
+	}
+
+	deleted := 0
+	failed := 0
+	for _, id := range ids {
+		tenantID, _ := mailStore.GetTenantForMail(ctx, id)
+
+		if err := mailStore.Delete(id); err != nil {
+			logger.Warn("purge: delete failed", "id", id, "err", err)
+			failed++
+			continue
+		}
+
+		if idxMgr != nil {
+			if err := idxMgr.ForTenant(tenantID).Delete(id); err != nil {
+				logger.Warn("purge: index cleanup failed", "id", id, "err", err)
+			}
+		}
+
+		if audlog != nil {
+			audlog.Log(audit.Entry{
+				EventType: "mail_purged",
+				Username:  "cron:purge",
+				TenantID:  tenantID,
+				MailID:    id,
+				Success:   true,
+				Detail:    "automatischer Purge nach Ablauf der Aufbewahrungsfrist (retain_until)",
+			})
+		}
+
+		deleted++
+	}
+
+	logger.Info("purge: complete", "total", len(ids), "deleted", deleted, "failed", failed)
+}
diff --git a/cmd/archivmail/main.go b/cmd/archivmail/main.go
index 1dc6c65..203c7b6 100644
--- a/cmd/archivmail/main.go
+++ b/cmd/archivmail/main.go
@@ -58,6 +58,9 @@ func main() {
 		case "reindex":
 			runReindex(os.Args[2:])
 			return
+		case "purge":
+			runPurge(os.Args[2:])
+			return
 		case "recompress":
 			runRecompress(os.Args[2:])
 			return
diff --git a/deploy/cron.d/archivmail b/deploy/cron.d/archivmail
index ceed789..3c4909c 100644
--- a/deploy/cron.d/archivmail
+++ b/deploy/cron.d/archivmail
@@ -26,6 +26,11 @@
 # Pause beenden um 06:00 Uhr
 0 6  * * * root /usr/local/bin/archivmail-ocr-pause.sh stop
 
+# ── GoBD-Retention-Purge (PROJ-56c) ─────────────────────────────────────
+# Löscht Mails mit abgelaufener Aufbewahrungsfrist (retain_until < NOW()),
+# entfernt sie aus dem Suchindex und schreibt pro Mail einen Audit-Eintrag
+# ("mail_purged") — analog zu Pilers purge.sh, nachts um 03:40 Uhr.
+40 3 * * * root /opt/archivmail/archivmail purge --config /etc/archivmail/config.yml >> /var/log/archivmail/purge.log 2>&1
+
 # ── Weitere Jobs (geplant, noch nicht implementiert) ────────────────────
-# 40 3 * * * archivmail /usr/local/bin/archivmail-purge.sh        # GoBD-Retention-Purge
 # 30 2 * * * archivmail /opt/archivmail/archivmail reindex        # nächtlicher Voll-Reindex
diff --git a/internal/storage/storage.go b/internal/storage/storage.go
index 8b31c16..ebff584 100644
--- a/internal/storage/storage.go
+++ b/internal/storage/storage.go
@@ -666,27 +666,57 @@ func (s *Store) Load(id string) ([]byte, error) {
 }
 
 // Delete removes a stored email by its ID, including its DB metadata row.
+// Delete removes a mail's metadata (emails row + dependent email_refs /
+// email_attachments rows, inside a transaction) and only deletes the file
+// from disk once that transaction has committed.
+//
+// This order matters: emails is referenced by email_refs and
+// email_attachments without ON DELETE CASCADE, so a plain
+// "DELETE FROM emails" fails with a FK violation whenever such rows exist
+// (e.g. any tenant-routed or multi-recipient mail). Removing the file
+// *before* the DB delete — as this function used to do — meant that
+// failure left an orphaned DB row pointing at a file that no longer
+// existed, silently, because the error was discarded. Doing the DB work
+// first and only then unlinking the file means a failed DB delete leaves
+// the file (and the mail) intact instead of corrupting metadata.
 func (s *Store) Delete(id string) error {
-	// PROJ-34: Enforce retention lock before any disk or DB operation.
+	ctx := context.Background()
+
 	if s.db != nil {
-		ctx := context.Background()
 		var until *time.Time
 		_ = s.db.QueryRow(ctx, `SELECT retain_until FROM emails WHERE id=$1`, id).Scan(&until)
 		if until != nil && time.Now().Before(*until) {
 			return ErrRetentionLock
 		}
+
+		tx, err := s.db.Begin(ctx)
+		if err != nil {
+			return fmt.Errorf("storage: delete: begin tx: %w", err)
+		}
+		defer tx.Rollback(ctx)
+
+		if _, err := tx.Exec(ctx, `DELETE FROM email_refs WHERE email_id = $1`, id); err != nil {
+			return fmt.Errorf("storage: delete: email_refs: %w", err)
+		}
+		if _, err := tx.Exec(ctx, `DELETE FROM email_attachments WHERE email_id = $1`, id); err != nil {
+			return fmt.Errorf("storage: delete: email_attachments: %w", err)
+		}
+		if _, err := tx.Exec(ctx, `DELETE FROM emails WHERE id = $1`, id); err != nil {
+			return fmt.Errorf("storage: delete: emails: %w", err)
+		}
+		if err := tx.Commit(ctx); err != nil {
+			return fmt.Errorf("storage: delete: commit: %w", err)
+		}
 	}
 
 	path := s.filePath(id)
 	if err := os.Remove(path); err != nil {
 		if errors.Is(err, os.ErrNotExist) {
-			return fmt.Errorf("storage: not found: %s", id)
+			// DB metadata is already gone (or never existed) at this point;
+			// a missing file is not an error worth failing the caller for.
+			return nil
 		}
-		return fmt.Errorf("storage: delete: %w", err)
-	}
-
-	if s.db != nil {
-		_, _ = s.db.Exec(context.Background(), `DELETE FROM emails WHERE id = $1`, id)
+		return fmt.Errorf("storage: delete: file: %w", err)
 	}
 
 	return nil
@@ -717,17 +747,18 @@ func (s *Store) GetReceivedAts(ctx context.Context, ids []string) map[string]tim
 	return result
 }
 
-// Purge deletes all mails whose retain_until has passed.
-// Returns the number of successfully deleted mails.
-// Mails that fail to delete (e.g. file missing) are skipped silently.
-func (s *Store) Purge(ctx context.Context) (int, error) {
+// ListExpiredMailIDs returns the IDs of all mails whose retain_until has
+// passed (PROJ-56c). Used by both Purge() and the cron-driven CLI purge
+// command, which additionally needs the IDs to clean up the search index
+// and write per-mail audit entries.
+func (s *Store) ListExpiredMailIDs(ctx context.Context) ([]string, error) {
 	if s.db == nil {
-		return 0, nil
+		return nil, nil
 	}
 	rows, err := s.db.Query(ctx,
 		`SELECT id FROM emails WHERE retain_until IS NOT NULL AND retain_until < NOW()`)
 	if err != nil {
-		return 0, fmt.Errorf("storage: purge query: %w", err)
+		return nil, fmt.Errorf("storage: list expired query: %w", err)
 	}
 	defer rows.Close()
 	var ids []string
@@ -738,7 +769,22 @@ func (s *Store) Purge(ctx context.Context) (int, error) {
 		}
 	}
 	if err := rows.Err(); err != nil {
-		return 0, fmt.Errorf("storage: purge rows: %w", err)
+		return nil, fmt.Errorf("storage: list expired rows: %w", err)
+	}
+	return ids, nil
+}
+
+// Purge deletes all mails whose retain_until has passed.
+// Returns the number of successfully deleted mails.
+// Mails that fail to delete (e.g. file missing) are skipped silently.
+//
+// NOTE: this does not remove entries from the search index or write audit
+// log entries — callers that need that (e.g. the cron-driven CLI purge
+// command, PROJ-56c) should use ListExpiredMailIDs directly instead.
+func (s *Store) Purge(ctx context.Context) (int, error) {
+	ids, err := s.ListExpiredMailIDs(ctx)
+	if err != nil {
+		return 0, err
 	}
 	deleted := 0
 	for _, id := range ids {