Pin the HTTP target's unpinned error checks (closes #285)

withRetry's check on a failed result write could be removed with
every test still passing, and so could six other error checks in
target_http.go. Each now has a test that fails without it: the
circuit breaker learning the answer to a send whose result went
unrecorded, the result write for an invalid config, building the
request, reading the response body, decoding the target config, and
decoding the stored inbound headers.

The two backoff lookups' error checks stay unpinned: without them a
failed lookup leaves a zero time, which gives the same answer, so no
test can tell the difference.

Model: opus-5-5

internal/delivery/engine_integration_test.go

@@ -1425,6 +1425,32 @@ func TestDeliverHTTP_InvalidConfig(t *testing.T) {
 	)
 }
 
+// TestDeliverHTTP_InvalidConfigUnrecordedStaysPending: a delivery is
+// failed for an invalid config only once the reason is recorded.
+// Unrecorded, it stays pending, where the sweep finds it again.
+func TestDeliverHTTP_InvalidConfigUnrecordedStaysPending(t *testing.T) {
+	t.Parallel()
+
+	db := testWebhookDB(t)
+	e := testEngine(t, 1)
+
+	event, del := iSeedEventAndDelivery(
+		t, db, `{"config":"invalid"}`, "",
+	)
+
+	task, d := iHTTPTaskAndDelivery(
+		event, del, "bad-config", `not-json`, 0, 1,
+	)
+
+	require.NoError(t, db.Exec("drop table delivery_results").Error)
+
+	e.ExportDeliverHTTP(context.TODO(), db, d, task)
+
+	iAssertStatus(t, db, del.ID,
+		database.DeliveryStatusPending,
+	)
+}
+
 // --- Notify batching ---
 
 func TestNotify_MultipleTasks(t *testing.T) {