test(proxy): stop the proxy budget from leaking between proxy tests

`proxy-infra / Run tests` failed on four parametrizations of
`test_real_proxy_child_auth_privacy_and_body_policy` -- the four that expect the
child request to succeed. The traceback ends in upstream code this PR does not
touch: `_user_api_key_auth_builder` calls `_fetch_global_spend_with_event_coordination`,
whose loader reads `prisma_client.db.litellm_usertable`, and the fixture installs
`object()` as the prisma client, so the read raises
`'object' object has no attribute 'db'` and authentication returns 401.

That block only runs when `litellm.max_budget > 0`, and the fixture never sets
it, so the value came from a test that ran earlier on the same xdist worker.
`test_add_proxy_budget_to_db_only_creates_user_no_keys` and
`test_add_proxy_budget_to_db_backfills_budget_reset_at` assign
`litellm.max_budget = 100.0` and `litellm.budget_duration` directly and never
restore them, and which tests share a worker changes from run to run, which is
why the same commit was green for these four cases and red for them later.

Two small test fixes, no production change:

- the two budget startup tests now set those globals through `monkeypatch`, so
  they are undone at teardown like the rest of the file's budget tests already
  do;
- the compaction fixture pins `litellm.max_budget` to 0.0 beside the other pins,
  so it asserts its own precondition instead of depending on suite order.

Validation: `test_native_compaction.py` plus both budget tests -- 16 passed;
`ruff check --config ruff-tests.toml` clean on both files, and `ruff format
--diff` leaves the added lines alone (the format check already reported this
file before the change).

tests/test_litellm/proxy/test_native_compaction.py

@@ -7,6 +7,7 @@ import pytest
 from fastapi import FastAPI, Request
 from pydantic import TypeAdapter
 
+import litellm
 from litellm.caching.caching import DualCache
 from litellm.exceptions import BadRequestError
 from litellm.litellm_core_utils.initialize_dynamic_callback_params import inherit_message_logging_privacy
@@ -118,6 +119,11 @@ async def test_real_proxy_child_auth_privacy_and_body_policy(
     monkeypatch.setattr(proxy_server, "user_api_key_cache", cache)
     monkeypatch.setattr(proxy_server, "llm_router", None)
     monkeypatch.setattr(proxy_server, "general_settings", {})
+    # Pin the proxy-wide budget: authentication only reads the global spend when a
+    # proxy max budget is configured, and that read goes through the stub prisma
+    # client above. A budget left set by an earlier test on the same worker would
+    # turn this fixture's child requests into 401s.
+    monkeypatch.setattr(litellm, "max_budget", 0.0)
     monkeypatch.setattr(common_request_processing, "route_request", route)
     with inherit_message_logging_privacy(True):
         call: Final = with_proxy_compaction_executor(

tests/test_litellm/proxy/test_proxy_server.py

@@ -3200,7 +3200,7 @@ def test_normalize_datetime_for_sorting():
 
 
 @pytest.mark.asyncio
-async def test_add_proxy_budget_to_db_only_creates_user_no_keys():
+async def test_add_proxy_budget_to_db_only_creates_user_no_keys(monkeypatch: pytest.MonkeyPatch):
     """
     Test that _add_proxy_budget_to_db only creates a user and no keys are added.
 
@@ -3217,9 +3217,12 @@ async def test_add_proxy_budget_to_db_only_creates_user_no_keys():
     import litellm
     from litellm.proxy.proxy_server import ProxyStartupEvent
 
-    # Set up required litellm settings
-    litellm.budget_duration = "30d"
-    litellm.max_budget = 100.0
+    # Set up required litellm settings. Through monkeypatch rather than plain
+    # assignment: `litellm.max_budget` is process-global, and any later test on
+    # this worker that authenticates reads the global proxy spend whenever a
+    # proxy budget is set, which needs a real prisma client.
+    monkeypatch.setattr(litellm, "budget_duration", "30d")
+    monkeypatch.setattr(litellm, "max_budget", 100.0)
 
     litellm_proxy_budget_name = "litellm-proxy-budget"
 
@@ -3258,7 +3261,7 @@ async def test_add_proxy_budget_to_db_only_creates_user_no_keys():
 
 
 @pytest.mark.asyncio
-async def test_add_proxy_budget_to_db_backfills_budget_reset_at():
+async def test_add_proxy_budget_to_db_backfills_budget_reset_at(monkeypatch: pytest.MonkeyPatch):
     """
     Test that _upsert_proxy_budget_with_reset_at_backfill issues a conditional
     update_many with `WHERE budget_reset_at IS NULL` to backfill the column on
@@ -3276,8 +3279,8 @@ async def test_add_proxy_budget_to_db_backfills_budget_reset_at():
     import litellm
     from litellm.proxy.proxy_server import ProxyStartupEvent
 
-    litellm.budget_duration = "30d"
-    litellm.max_budget = 100.0
+    monkeypatch.setattr(litellm, "budget_duration", "30d")
+    monkeypatch.setattr(litellm, "max_budget", 100.0)
     litellm_proxy_budget_name = "litellm-proxy-budget"
 
     mock_prisma = MagicMock()