fix: Subscription workers restart strategy, sandbox ownership race condition

test/support/integration_case.ex

@@ -30,11 +30,38 @@ defmodule Parrhesia.IntegrationCase do
         :ok = Sandbox.checkout(Repo)
 
       :shared ->
-        :ok = Sandbox.checkout(Repo)
-        Sandbox.mode(Repo, {:shared, self()})
+        # Delegate sandbox ownership to a dedicated process that outlives
+        # the test process. ExUnit terminates start_supervised! children
+        # (in reverse start order) before running on_exit callbacks. With
+        # the test process as sandbox owner, the connection dies the
+        # moment the test process exits — supervised children that make
+        # DB calls during their shutdown sequence would hit a dead
+        # connection. The separate owner stays alive through the entire
+        # supervised-child shutdown, and is only stopped in on_exit
+        # (which runs afterwards).
+        test_pid = self()
+
+        owner =
+          spawn(fn ->
+            :ok = Sandbox.checkout(Repo)
+            send(test_pid, {:sandbox_ready, self()})
+
+            receive do
+              :stop -> :ok
+            end
+          end)
+
+        receive do
+          {:sandbox_ready, ^owner} -> :ok
+        after
+          5_000 -> raise "Sandbox owner checkout timed out"
+        end
+
+        Sandbox.mode(Repo, {:shared, owner})
 
         on_exit(fn ->
           Sandbox.mode(Repo, :manual)
+          send(owner, :stop)
         end)
     end
 

test/support/runtime.ex

@@ -6,6 +6,7 @@ defmodule Parrhesia.TestSupport.Runtime do
     Parrhesia.Config,
     Parrhesia.Repo,
     Parrhesia.Subscriptions.Supervisor,
+    Parrhesia.API.Stream.Supervisor,
     Parrhesia.Web.Endpoint
   ]