style(auto-router): satisfy new ruff strict rules picked up from staging

Merge from litellm_internal_staging brought in RUF023 (sorted __slots__),
UP037 (unquoted forward refs where possible), I001 (import ordering), and
PLR0402 (proxy_server import aliasing). Auto-fixed with `ruff check --fix`;
no behavior change. Also regenerates schema.d.ts to pick up the endpoint
docstring wording from the prior commit.

litellm/proxy/db/autorouter_quality_signals.py

@@ -51,13 +51,13 @@ class Turn:
     """
 
     __slots__ = (
-        "session_id",
         "api_key",
-        "model",
-        "started_at",
         "client_disconnected",
-        "router_name",
         "has_client_session_id",
+        "model",
+        "router_name",
+        "session_id",
+        "started_at",
     )
 
     def __init__(
@@ -83,7 +83,7 @@ class Turn:
 class CohortSignals:
     """What one population evidences, and how much of it there was to look at."""
 
-    __slots__ = ("sessions", "escalation_rate_pct", "abandonment_rate_pct")
+    __slots__ = ("abandonment_rate_pct", "escalation_rate_pct", "sessions")
 
     def __init__(
         self,
@@ -97,7 +97,7 @@ class CohortSignals:
         self.abandonment_rate_pct = abandonment_rate_pct
 
 
-def rank_models_by_cost(router: "Router", models: Iterable[str]) -> Mapping[str, int]:
+def rank_models_by_cost(router: Router, models: Iterable[str]) -> Mapping[str, int]:
     """Order models cheapest-first, by what one fixed reference request would cost on each.
 
     Rank has to come from a request, not from a rate card: a model dearer per output token

tests/test_litellm/proxy/management_endpoints/test_auto_router_endpoints.py

@@ -21,10 +21,10 @@ from litellm.proxy.management_endpoints.auto_router_endpoints import (
     preview_auto_router_routing,
 )
 from litellm.router import Router
-from litellm.types.utils import Choices, Message, ModelResponse
 from litellm.types.management_endpoints.auto_router_endpoints import (
     AutoRouterRoutingTestRequest,
 )
+from litellm.types.utils import Choices, Message, ModelResponse
 
 ADMIN = UserAPIKeyAuth(user_role=LitellmUserRoles.PROXY_ADMIN, api_key="sk-test", user_id="admin")
 
@@ -55,7 +55,7 @@ def _request(prompt: str, **config_overrides: object) -> AutoRouterRoutingTestRe
 
 
 async def _route(prompt: str, monkeypatch: pytest.MonkeyPatch, **config_overrides: object):
-    import litellm.proxy.proxy_server as proxy_server
+    from litellm.proxy import proxy_server
 
     monkeypatch.setattr(proxy_server, "llm_router", _router())
     return await preview_auto_router_routing(
@@ -119,7 +119,7 @@ async def test_tier_model_missing_from_the_proxy_is_reported(monkeypatch: pytest
 
 @pytest.mark.asyncio
 async def test_llm_classifier_call_is_billed_to_the_calling_key(monkeypatch: pytest.MonkeyPatch):
-    import litellm.proxy.proxy_server as proxy_server
+    from litellm.proxy import proxy_server
 
     router = _router()
     calls: list[dict] = []
@@ -164,7 +164,7 @@ async def test_llm_classifier_call_is_billed_to_the_calling_key(monkeypatch: pyt
 async def test_a_key_that_cannot_call_the_classifier_model_is_rejected_before_it_is_called(
     monkeypatch: pytest.MonkeyPatch, config_overrides: dict
 ):
-    import litellm.proxy.proxy_server as proxy_server
+    from litellm.proxy import proxy_server
 
     router = _router()
     calls: list[dict] = []
@@ -194,7 +194,7 @@ async def test_a_key_that_cannot_call_the_classifier_model_is_rejected_before_it
 
 @pytest.mark.asyncio
 async def test_a_key_over_its_budget_cannot_run_a_classifier_config(monkeypatch: pytest.MonkeyPatch):
-    import litellm.proxy.proxy_server as proxy_server
+    from litellm.proxy import proxy_server
 
     router = _router()
     calls: list[dict] = []
@@ -228,7 +228,7 @@ async def test_a_key_over_its_budget_cannot_run_a_classifier_config(monkeypatch:
 
 @pytest.mark.asyncio
 async def test_a_heuristic_config_does_not_need_a_budget(monkeypatch: pytest.MonkeyPatch):
-    import litellm.proxy.proxy_server as proxy_server
+    from litellm.proxy import proxy_server
 
     monkeypatch.setattr(proxy_server, "llm_router", _router())
 
@@ -249,7 +249,7 @@ async def test_a_heuristic_config_does_not_need_a_budget(monkeypatch: pytest.Mon
 
 @pytest.mark.asyncio
 async def test_no_llm_router_on_the_proxy_is_a_500(monkeypatch: pytest.MonkeyPatch):
-    import litellm.proxy.proxy_server as proxy_server
+    from litellm.proxy import proxy_server
 
     monkeypatch.setattr(proxy_server, "llm_router", None)
 
@@ -261,7 +261,7 @@ async def test_no_llm_router_on_the_proxy_is_a_500(monkeypatch: pytest.MonkeyPat
 
 @pytest.mark.asyncio
 async def test_non_admin_without_a_team_is_rejected(monkeypatch: pytest.MonkeyPatch):
-    import litellm.proxy.proxy_server as proxy_server
+    from litellm.proxy import proxy_server
 
     monkeypatch.setattr(proxy_server, "llm_router", _router())
 

ui/litellm-dashboard/src/lib/http/schema.d.ts

@@ -820,8 +820,17 @@ export interface paths {
          *     model mid-session, and how often they hung up mid-stream, for auto-routed traffic and for
          *     the same keys' directly-addressed traffic.
          *
-         *     Reads LiteLLM_SpendLogs rather than the per-session rollup, because escalation is a
-         *     question about turn order and the rollup folds order away.
+         *     Both cohorts come from one scan of LiteLLM_SpendLogs rather than the per-session rollup,
+         *     so they cannot drift apart -- same window, same session grouping, same disconnect test --
+         *     and because escalation is a question about turn order, which the rollup folds away.
+         *     `router_name` is NULL for a directly-addressed request, which is what separates the two
+         *     populations downstream. Abandonment reads `error_information.error_code` rather than
+         *     `status`, because a client disconnect still bills its partial streamed spend as a success
+         *     and so does not show up in `status`. `session_turn_count` counts, per api_key, how many
+         *     rows in the window share a session_id; a fallback uuid minted by the spend writer
+         *     (`_get_session_id_for_spend_log`) is always unique to its one request, so a repeating
+         *     session_id can only have come from the caller, which needs nothing beyond columns every
+         *     deployment already writes, prompt storage on or off.
          *
          *     The two cohorts self-select, so this is directional evidence and not an experiment: a
          *     deployment that pins its hardest prompts to one model and routes only the easy ones will