fix(proxy): retain known spend when no-DB reconcile fails

When BudgetReservationReconcileLogger cannot settle a reservation and
response_cost is positive, do not release_or_invalidate (which refunds
to 0.0 and finalizes). Without Prisma there is no ledger to restore
that spend; leave the reservation open so settlement can retry and
budgets fail closed. Zero-cost failures still release as before.

litellm/proxy/hooks/budget_reservation_reconcile_logger.py

@@ -63,7 +63,17 @@ class BudgetReservationReconcileLogger(CustomLogger):
                 budget_reservation=budget_reservation,
                 actual_cost=actual_cost,
             )
-        except Exception:  # noqa: BLE001  # a counter backend failure must still release the reservation
+        except Exception:  # noqa: BLE001  # counter failure: refund only when no known spend
+            if actual_cost > 0:
+                # No Prisma ledger on this path. Releasing to 0.0 / finalizing would
+                # permanently refund completed work and block retry; leave the
+                # reservation open so settlement can retry and budgets fail closed.
+                verbose_proxy_logger.exception(
+                    "Failed to reconcile budget reservation without a database logger; "
+                    "retaining known spend %s for retry",
+                    actual_cost,
+                )
+                return
             verbose_proxy_logger.exception(
                 "Failed to reconcile budget reservation without a database logger; releasing instead"
             )

tests/unit/proxy/hooks/test_budget_reservation_reconcile_logger.py

@@ -151,6 +151,36 @@ async def test_success_falls_back_to_release_when_reconcile_fails(
     release.assert_awaited_once_with(budget_reservation=reservation)
 
 
+@pytest.mark.asyncio
+async def test_success_retains_known_spend_when_reconcile_fails(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    reservation = _reservation()
+    monkeypatch.setattr(
+        "litellm.proxy.spend_tracking.budget_reservation.reconcile_budget_reservation",
+        AsyncMock(side_effect=RuntimeError("counter down")),
+    )
+    release = AsyncMock()
+    monkeypatch.setattr(
+        "litellm.proxy.spend_tracking.budget_reservation.release_or_invalidate_budget_reservation",
+        release,
+    )
+    logger = BudgetReservationReconcileLogger()
+
+    await logger.async_log_success_event(
+        {
+            "litellm_params": {"metadata": {"user_api_key_budget_reservation": reservation}},
+            "response_cost": 0.004,
+        },
+        response_obj=object(),
+        start_time=datetime.now(UTC),
+        end_time=datetime.now(UTC),
+    )
+
+    release.assert_not_awaited()
+    assert reservation["finalized"] is False
+
+
 @pytest.fixture
 def spend_counter_cache(monkeypatch: pytest.MonkeyPatch) -> DualCache:
     cache: Final = DualCache()