fix(gemini): treat rolling -latest aliases as Gemini 3+

`_is_gemini_3_or_newer` matched the substring "gemini-3", so Google's
rolling aliases (`gemini-flash-latest`, `gemini-flash-lite-latest` and
`gemini-pro-latest`) were not recognised even though they resolve to
Gemini 3.x. Every Gemini 3 behaviour keyed off that check was therefore
silently disabled for them, including the documented
`skip_thought_signature_validator` fallback.

The visible symptom is a multi-turn tool call failing with
`400 Function call is missing a thought_signature in functionCall parts`
whenever the assistant turn's tool call does not carry LiteLLM's
`__thought__`-encoded id, for example history replayed by a client that
uses its own tool-call ids.

Matching is by exact alias name rather than an `endswith("-latest")`
check, because versioned names such as
`gemini-2.5-flash-native-audio-latest` also end in `-latest` and are not
Gemini 3.

litellm/llms/vertex_ai/gemini/vertex_and_google_ai_studio_gemini.py

@@ -106,6 +106,19 @@ else:
     StreamingChoices = Any
 
 
+# Google's rolling aliases, which always resolve to the newest release of each
+# tier and today serve Gemini 3.x. They carry no version number, so they must be
+# listed explicitly for Gemini 3 feature detection to see them.
+# https://ai.google.dev/gemini-api/docs/models
+GEMINI_ROLLING_LATEST_ALIASES = frozenset(
+    {
+        "gemini-flash-latest",
+        "gemini-flash-lite-latest",
+        "gemini-pro-latest",
+    }
+)
+
+
 class VertexAIBaseConfig:
     def get_mapped_special_auth_params(self) -> dict:
         """
@@ -266,11 +279,21 @@ class VertexGeminiConfig(VertexAIBaseConfig, BaseConfig):
         - gemini-3.1-pro-preview, gemini-3.1-flash, gemini-3.1-flash-lite-preview
         - gemini-3.5-flash
         - Any future Gemini 3.x models
+        - The rolling `-latest` aliases, which Google keeps pointed at the
+          newest release of each tier and which currently serve Gemini 3.x.
+
+        The `-latest` aliases are matched by exact name rather than by an
+        `endswith("-latest")` suffix check, because versioned names such as
+        `gemini-2.5-flash-native-audio-latest` also end in `-latest` and are
+        not Gemini 3.
         """
         # Check for Gemini 3 models
         if "gemini-3" in model:
             return True
-        return False
+        # Rolling aliases carry no version in the name, so the substring check
+        # above cannot see them. Strip any provider prefix (`gemini/`,
+        # `vertex_ai/`) before comparing.
+        return model.split("/")[-1] in GEMINI_ROLLING_LATEST_ALIASES
 
     @staticmethod
     def _forward_gemini_function_call_id(model: str) -> bool:

tests/test_litellm/llms/vertex_ai/gemini/test_vertex_and_google_ai_studio_gemini.py

@@ -2275,6 +2275,74 @@ def test_is_gemini_3_or_newer():
     # Edge cases
     assert VertexGeminiConfig._is_gemini_3_or_newer("") == False
 
+    # Rolling `-latest` aliases resolve to Gemini 3.x and carry no version in
+    # the name, so they must be detected explicitly.
+    assert VertexGeminiConfig._is_gemini_3_or_newer("gemini-flash-latest") == True
+    assert VertexGeminiConfig._is_gemini_3_or_newer("gemini-flash-lite-latest") == True
+    assert VertexGeminiConfig._is_gemini_3_or_newer("gemini-pro-latest") == True
+    assert (
+        VertexGeminiConfig._is_gemini_3_or_newer("gemini/gemini-flash-latest") == True
+    )
+    assert (
+        VertexGeminiConfig._is_gemini_3_or_newer("vertex_ai/gemini-pro-latest") == True
+    )
+
+    # A versioned name that merely ends in `-latest` is not Gemini 3.
+    assert (
+        VertexGeminiConfig._is_gemini_3_or_newer(
+            "gemini-2.5-flash-native-audio-latest"
+        )
+        == False
+    )
+
+
+def test_thought_signature_fallback_for_rolling_latest_alias():
+    """
+    A tool call whose thought signature did not survive the round-trip must
+    still get the dummy signature Google documents, otherwise Gemini rejects
+    the follow-up turn with:
+
+      400 Function call is missing a thought_signature in functionCall parts.
+
+    Regression test: `gemini-flash-latest` serves Gemini 3.x but was not
+    detected as such, so the fallback was skipped and every multi-turn tool
+    call that lost its signature failed.
+    """
+    from litellm.litellm_core_utils.prompt_templates.factory import (
+        convert_to_gemini_tool_call_invoke,
+    )
+
+    message = {
+        "role": "assistant",
+        "content": None,
+        "tool_calls": [
+            {
+                "id": "call_abc123",  # no `__thought__` suffix -> signature lost
+                "type": "function",
+                "function": {
+                    "name": "search_website",
+                    "arguments": '{"query": "Iceland"}',
+                },
+            }
+        ],
+    }
+
+    for model in [
+        "gemini-3-flash",
+        "gemini-flash-latest",
+        "gemini-flash-lite-latest",
+        "gemini-pro-latest",
+        "gemini/gemini-flash-latest",
+    ]:
+        parts = convert_to_gemini_tool_call_invoke(message, model=model)
+        assert any(
+            "thoughtSignature" in part for part in parts
+        ), f"expected a thought signature for {model}"
+
+    # Pre-Gemini-3 models must not gain a signature they never needed.
+    parts = convert_to_gemini_tool_call_invoke(message, model="gemini-2.5-flash")
+    assert not any("thoughtSignature" in part for part in parts)
+
 
 def _tool_call_messages(tool_call_id: str):
     return [