Fase 1a: microservicio tcef-notifications (workers FCM v1 + email)

Sustituye el envío de push por la API legacy de GCM (muerta desde jun-2024) y
saca el envío del observer de Meteor a una cola BullMQ controlada.

- poller: consume 'notifications' pendientes (campos correctos; el cron viejo
  tenía un typo nofitied/emailNofitied que nunca casaba)
- fcm-worker: firebase-admin (FCM HTTP v1), payload compatible con la app
  Flutter (FLUTTER_NOTIFICATION_CLICK, collapseKey=_id, data); purga tokens
  muertos sin reintentar
- email-worker: nodemailer + plantillas new-fire portadas verbatim
- idempotencia persistente: notification_sends, indice unico (notificationId,
  channel) -> sobrevive reinicios y replays
- salvaguardas anti-spam: modos dry-run/shadow/canary/full, exclusion mutua por
  canal (OWNED_CHANNELS + notifDisabledChannels en Meteor), kill switch,
  limites por usuario/dia y circuit breaker por batch
- docs/legacy-behavior.md (caracterizacion) + docs/rollout.md
- 47 tests (vitest + mongodb-memory-server), typecheck y build OK
- deploy: systemd + pm2 (Node 22 standalone)

Falta el service account de Firebase para envio real de push (pedir al usuario).

src/queue.ts

@@ -0,0 +1,46 @@
+import { Queue, type ConnectionOptions, type QueueOptions } from 'bullmq';
+import type { Config } from './config.js';
+import type { Channel, SendJob } from './types.js';
+
+export const QUEUE_NAMES: Record<Channel, string> = {
+  fcm: 'tcef-fcm',
+  email: 'tcef-email',
+};
+
+/**
+ * BullMQ connection options. We hand BullMQ plain options (not a shared ioredis
+ * instance) so it manages its own blocking/non-blocking connections and sets
+ * maxRetriesPerRequest:null on workers itself.
+ */
+export function redisConnection(cfg: Config): ConnectionOptions {
+  return {
+    host: cfg.redis.host,
+    port: cfg.redis.port,
+    password: cfg.redis.password,
+    db: cfg.redis.db,
+  };
+}
+
+export function createQueues(cfg: Config): Record<Channel, Queue<SendJob>> {
+  const opts: QueueOptions = {
+    connection: redisConnection(cfg),
+    defaultJobOptions: {
+      attempts: cfg.retries.attempts,
+      backoff: { type: 'exponential', delay: cfg.retries.backoffMs },
+      removeOnComplete: { count: 1000 },
+      removeOnFail: false, // keep failures = dead-letter for inspection
+    },
+  };
+  return {
+    fcm: new Queue<SendJob>(QUEUE_NAMES.fcm, opts),
+    email: new Queue<SendJob>(QUEUE_NAMES.email, opts),
+  };
+}
+
+/**
+ * Job id = `${channel}:${notificationId}` so BullMQ itself dedupes enqueues:
+ * re-polling the same pending doc will not create a second job.
+ */
+export function jobId(job: SendJob): string {
+  return `${job.channel}:${job.notificationId}`;
+}