[CI][MCP][Harmony] Heavy refactoring Harmony & MCP response tests and stabilizing with deterministic test infrastructure (#33949)

Signed-off-by: Andreas Karatzas <akaratza@amd.com>

vllm/envs.py

@@ -209,6 +209,7 @@ if TYPE_CHECKING:
     VLLM_GPT_OSS_SYSTEM_TOOL_MCP_LABELS: set[str] = set()
     VLLM_USE_EXPERIMENTAL_PARSER_CONTEXT: bool = False
     VLLM_GPT_OSS_HARMONY_SYSTEM_INSTRUCTIONS: bool = False
+    VLLM_SYSTEM_START_DATE: str | None = None
     VLLM_TOOL_JSON_ERROR_AUTOMATIC_RETRY: bool = False
     VLLM_CUSTOM_SCOPES_FOR_PROFILING: bool = False
     VLLM_NVTX_SCOPES_FOR_PROFILING: bool = False
@@ -1458,6 +1459,12 @@ environment_variables: dict[str, Callable[[], Any]] = {
     "VLLM_GPT_OSS_HARMONY_SYSTEM_INSTRUCTIONS": lambda: bool(
         int(os.getenv("VLLM_GPT_OSS_HARMONY_SYSTEM_INSTRUCTIONS", "0"))
     ),
+    # Pin the conversation start date injected into the Harmony system
+    # message. When unset the current date is used, which introduces
+    # non-determinism (different tokens -> different model behaviour at
+    # temperature=0). Set to an ISO date string, e.g. "2023-09-12",
+    # for reproducible inference or testing.
+    "VLLM_SYSTEM_START_DATE": lambda: os.getenv("VLLM_SYSTEM_START_DATE", None),
     # Enable automatic retry when tool call JSON parsing fails
     # If enabled, returns an error message to the model to retry
     # If disabled (default), raises an exception and fails the request