fix(tasks): share one sync engine per worker process to stop connection-pool leak

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

backend/app/tasks/_sync_engine.py

@@ -0,0 +1,36 @@
+"""Process-level sync engine for Celery task modules.
+
+Each task module used to call ``create_engine(...)`` on every invocation,
+which leaked engines (and Postgres connections) — high-fire-rate tasks
+like ``import_media_file`` would exhaust ``max_connections`` within
+minutes during a bulk migration.
+
+This module owns one engine per process. Celery prefork forks before any
+task runs, so each worker process lazily initializes its own engine on
+the first task and reuses it for the rest of its life.
+"""
+
+from sqlalchemy import create_engine
+from sqlalchemy.orm import sessionmaker
+
+from ..config import get_config
+
+_ENGINE = None
+_SESSIONMAKER = None
+
+
+def sync_session_factory():
+    """Return a process-wide ``sessionmaker`` bound to a single engine."""
+    global _ENGINE, _SESSIONMAKER
+    if _SESSIONMAKER is None:
+        cfg = get_config()
+        _ENGINE = create_engine(
+            cfg.database_url_sync,
+            future=True,
+            pool_pre_ping=True,
+            pool_size=5,
+            max_overflow=5,
+            pool_recycle=300,
+        )
+        _SESSIONMAKER = sessionmaker(_ENGINE, expire_on_commit=False)
+    return _SESSIONMAKER