diff --git a/src/backend_server.py b/src/backend_server.py
index 2f85ddc..9788236 100644
--- a/src/backend_server.py
+++ b/src/backend_server.py
@@ -1000,6 +1000,19 @@ def _ensure_schema(conn):
     ).fetchone()
     if exists is None:
         schema.create_schema(conn)
+    # 2026-08-06 (real incident): create_schema() above only ever runs for a
+    # BRAND NEW db file. Every schema.migrate_*() function is documented as
+    # idempotent/safe-to-call-unconditionally (including against a fresh
+    # create_schema()'d db, which already has whatever they'd add), but
+    # nothing ever actually called them against an EXISTING, already-
+    # populated database until now -- caught for real when this exact
+    # process restarted against the live production db and every
+    # GET /api/tasks/:id started 500ing with "no such column: parent_run_id"
+    # (D's migrate_add_work_graph_fork_columns had shipped and been tested,
+    # but never wired into the one code path that runs on every real
+    # startup). Run all three every time, not just on first-ever creation.
+    schema.migrate_expand_artifact_types(conn)
+    schema.migrate_relax_org_node_constraints(conn)  # also runs migrate_add_work_graph_fork_columns internally
 
 
 class _Server(ThreadingHTTPServer):