test(e2e): tighten toolset wire-name, pricing-reload precondition, and stream-ordering assertions

Addresses the three Greptile P2 findings on this PR.

- MCP toolset test: assert the served tool's exact documented wire name,
  built from the registered server alias and the '-' separator, instead of
  accepting any served name that merely ends in the bare tool. A wrong prefix
  or separator now fails, matching what a client calling the documented name
  actually gets.
- Pricing override test: assert the override reads as cleared before Reload
  Price Data, so the post-reload check proves reload preserved the clear
  rather than a later ordinary router refresh doing the clearing.
- Responses stream test: replace the mutable set accumulator with an
  immutable, order-preserving check of which item ids were announced strictly
  before each text delta. This also drops the fragile events.index(event)
  first-match lookup in the failure message for the enumerate index.

tests/e2e/llm_translation/test_responses_stream_text_start_e2e.py

@@ -16,7 +16,7 @@ issue describes.
 
 from __future__ import annotations
 
-from typing import Literal
+from typing import Final, Literal
 
 import pytest
 from pydantic import BaseModel
@@ -84,6 +84,19 @@ class _StreamEvent(BaseModel):
         return self.item.id if self.item else None
 
 
+_ANNOUNCE_TYPES: Final = ("response.output_item.added", "response.content_part.added")
+
+
+def _announced_before(events: list[_StreamEvent], index: int) -> frozenset[str]:
+    """Item ids announced by an output_item.added / content_part.added event
+    strictly before `index`: the ids a text delta at `index` is allowed to target."""
+    return frozenset(
+        item_id
+        for event in events[:index]
+        if event.type in _ANNOUNCE_TYPES and (item_id := event.announced_item_id()) is not None
+    )
+
+
 def _stream_post_tool_text(proxy: ProxyClient, key: str, model: str) -> StreamingResponse:
     call_id = f"call_{unique_marker()}"
     return proxy.transport.stream(
@@ -123,18 +136,16 @@ class TestResponsesStreamTextStart:
         deltas = [event for event in events if event.type == "response.output_text.delta"]
         assert deltas, f"the post-tool round streamed no text deltas: {[e.type for e in events]}"
 
-        announced: set[str] = set()
-        for event in events:
+        for index, event in enumerate(events):
+            if event.type != "response.output_text.delta":
+                continue
             item_id = event.announced_item_id()
-            if event.type in ("response.output_item.added", "response.content_part.added") and item_id:
-                announced.add(item_id)
-            if event.type == "response.output_text.delta":
-                assert item_id in announced, (
-                    f"text delta for item {item_id!r} arrived before any output_item.added/"
-                    f"content_part.added announced it; OpenAI-SDK parsers fail with "
-                    f"'text part ... not found' (#27671). Events so far: "
-                    f"{[e.type for e in events[: events.index(event) + 1]]}"
-                )
+            assert item_id in _announced_before(events, index), (
+                f"text delta for item {item_id!r} arrived before any output_item.added/"
+                f"content_part.added announced it; OpenAI-SDK parsers fail with "
+                f"'text part ... not found' (#27671). Events so far: "
+                f"{[e.type for e in events[: index + 1]]}"
+            )
         assert any(event.type == "response.completed" for event in events), (
             f"stream never completed: {[e.type for e in events][-5:]}"
         )

tests/e2e/management/test_model_pricing_override_e2e.py

@@ -122,7 +122,12 @@ class TestPricingOverrideClearAndReload:
                 response_type=NoBody,
             )
         )
-        _poll_override_cleared(proxy, model_name)
+        cleared_entry = _poll_override_cleared(proxy, model_name)
+        assert cleared_entry.litellm_params.input_cost_per_token is None, (
+            f"override must be cleared before the reload, else a later ordinary router refresh "
+            f"could clear it and let this test pass without proving reload preserved the clear: "
+            f"{cleared_entry.litellm_params.input_cost_per_token}"
+        )
 
         unwrap(
             proxy.transport.post(

tests/e2e/mcp/test_mcp_toolset_e2e.py

@@ -126,12 +126,15 @@ class TestMcpToolsetToolNames:
             )
         )
         served_names = [tool.name for tool in served.tools]
-        prefixed = [
-            name
-            for name in served_names
-            if name.endswith(SEARCH_LOGS_TOOL) and name != SEARCH_LOGS_TOOL
-        ]
-        assert prefixed, (
-            f"a toolset-scoped key must be served {SEARCH_LOGS_TOOL} under its "
-            f"'{{server_prefix}}{{separator}}{SEARCH_LOGS_TOOL}' wire name; got {served_names}"
+        alias = next(
+            (row.alias for row in client.registered_servers() if row.server_id == server_id),
+            None,
+        )
+        assert alias is not None, f"registered server {server_id} absent from /v1/mcp/server listing"
+        expected_name = f"{alias}-{SEARCH_LOGS_TOOL}"
+        assert expected_name in served_names, (
+            f"a toolset-scoped key must be served {SEARCH_LOGS_TOOL} under its exact documented "
+            f"wire name {expected_name!r} ('<alias>-<tool>'). A wrong prefix or separator still "
+            f"ends in {SEARCH_LOGS_TOOL} but breaks every client calling the documented name with "
+            f"'tool not found'; got {served_names}"
         )