diff --git a/backend/app/workers/celery_app.py b/backend/app/workers/celery_app.py
index 2d34e118..0dfcd9da 100644
--- a/backend/app/workers/celery_app.py
+++ b/backend/app/workers/celery_app.py
@@ -8,7 +8,7 @@ import logging
 
 from celery import Celery
 from celery.schedules import crontab
-from celery.signals import worker_ready
+from celery.signals import worker_process_init, worker_ready
 
 from app.core.config import settings
 
@@ -209,6 +209,29 @@ def _build_beat_schedule() -> dict:
 celery_app.conf.beat_schedule = _build_beat_schedule()
 
 
+@worker_process_init.connect
+def _reset_db_connections(sender=None, **_kwargs) -> None:
+    """Dispose SQLAlchemy engine in each prefork child process.
+
+    Background: Celery prefork pool fork()'s child processes from the master
+    that already imported SQLAlchemy + psycopg. The child inherits open TCP
+    sockets to PostgreSQL, but psycopg connections cannot be safely shared
+    across processes after fork — first use raises:
+        psycopg.ProgrammingError: can't change 'autocommit' now:
+        connection in transaction status INTRANS
+
+    Fix: dispose the engine in each child. SQLAlchemy will lazily open fresh
+    connections per-child. Standard Celery + SQLA recipe.
+    """
+    try:
+        from app.core.db import engine
+
+        engine.dispose()
+        logger.info("worker_process_init: SQLAlchemy engine disposed for fork")
+    except Exception as e:
+        logger.warning("worker_process_init: engine dispose failed: %s", e)
+
+
 @worker_ready.connect
 def _resume_zombie_runs(sender=None, **_kwargs) -> None:
     """When a worker finishes booting (after redeploy/restart), find any sweep
@@ -221,6 +244,7 @@ def _resume_zombie_runs(sender=None, **_kwargs) -> None:
     >5 min, which skipped jobs interrupted seconds before redeploy — they
     stayed in 'running' status forever and required manual cancel/resume.
     """
+    logger.info("worker_ready: resume scan starting")
     from sqlalchemy import text
 
     from app.core.db import SessionLocal
@@ -333,3 +357,4 @@ def _resume_zombie_runs(sender=None, **_kwargs) -> None:
                 logger.info("worker_ready: process_nspd_geo_job enqueued job=%s", jid)
             except Exception as e:
                 logger.warning("worker_ready: failed to enqueue geo resume job=%s: %s", jid, e)
+    logger.info("worker_ready: resume scan finished (geo_jobs=%d)", len(geo_resume_jobs))