feat(audit): admin audit log (Notification P3)

Append-only who/what/when trail for admin actions + scheduler run summaries,
generalizing QuotaEvent. Logged at ACTION / run-summary granularity — never
per-row (a scheduler run touching thousands of videos is ONE row).

Backend:
- migration 0054_audit_log + AuditLog model (actor_id FK SET NULL, action,
  target_type/id, summary, before/after JSON, created_at).
- app/audit.py: AuditAction constants (single source of truth, i18n'd on the
  client) + record() (adds to the caller's txn) + gc(retention_days).
- producers wired: role change / suspend / delete / invite approve-deny-add /
  demo add-remove-reset / discovery purge (admin.py); config set/reset — secret
  VALUES never logged, key only (config.py); scheduler interval + maintenance
  tune (scheduler routes); sync pause/resume (sync.py); and the scheduler _job()
  choke point writes ONE run-summary row per run — manual runs + errors always,
  automatic runs only when they changed something (no no-op-poll spam).
- retention: audit_gc scheduler job prunes rows older than audit_retention_days
  (sysconfig, default 180; 0 = keep forever).
- admin API routes/audit.py (/api/admin/audit): recent-window list (actor emails
  resolved, System for background) + clear (leaves one tamper-evidence row).

Frontend:
- new admin-only "Audit log" nav page (ScrollText) using the shared DataTable
  (first admin page to reuse it) — sort/filter/paginate/persist, action select
  filter, actor text filter, before→after detail, Clear-all with confirm.
- AuditLog.tsx + auditColumns.tsx + api.adminAudit/clearAudit + AuditRow type.
- trilingual audit + auditActions namespaces (HU/EN/DE) + nav + config-group +
  scheduler job labels; Audit config group (retention days).

backend/app/models.py

@@ -495,6 +495,30 @@ class QuotaEvent(Base):
     )
 
 
+class AuditLog(Base):
+    """Append-only admin/scheduler audit trail: who did what, when, and (where it applies)
+    the before->after values. `actor_id` is the acting admin; NULL means the scheduler /
+    background did it (SET NULL on user delete keeps the trail). Logged at ACTION / job-run
+    granularity — never per-row — so a scheduler run that touches thousands of videos is ONE
+    row summarising the outcome. `before`/`after` are small JSON snapshots (secrets excluded)."""
+
+    __tablename__ = "audit_log"
+
+    id: Mapped[int] = mapped_column(primary_key=True)
+    actor_id: Mapped[int | None] = mapped_column(
+        ForeignKey("users.id", ondelete="SET NULL"), index=True
+    )
+    action: Mapped[str] = mapped_column(String(48), index=True)
+    target_type: Mapped[str | None] = mapped_column(String(32))
+    target_id: Mapped[str | None] = mapped_column(String(128))
+    summary: Mapped[str | None] = mapped_column(String(255))
+    before: Mapped[dict | None] = mapped_column(JSON)
+    after: Mapped[dict | None] = mapped_column(JSON)
+    created_at: Mapped[datetime] = mapped_column(
+        DateTime(timezone=True), server_default=func.now(), index=True
+    )
+
+
 class AppState(Base):
     """Single-row global app state (admin-controlled)."""