fix: add test_id to normalize_tool_request() to avoid hash collision

[msager27]

What does this PR do?

Fix an issue with the recording system in which a hashed recording might collide. Here's some context:

While working on PR #5216, which adds vllm provider coverage (using a qwen35 model) to the responses test suite, I was running some tests and creating recordings using vllm+qwen35. In one case, I noticed an existing recording with gpt was changed. It seems the model generated the same web search query causing the collision. For reference, I'm including the diff of the changed recording below.

I originally included the fix in PR 5216, but that PR will take some time to resolve as it ties into the discussion of how to add vllm+gpu provider coverage upstream. So, I'm submitting the fix separately. If accepted, I believe the new integration test record action would need to be run to update existing recordings.

Note: the other normalize_* functions already include test_id.

diff --git a/tests/integration/responses/recordings/54aa690e31b5c33a0488a5d7403393e5712917253462292829b37b9320d6df82.json b/tests/integration/responses/recordings/54aa690e31b5c33a0488a5d7403393e5712917253462292829b37b9320d6df82.json
index a8e1e861..8f5af108 100644
--- a/tests/integration/responses/recordings/54aa690e31b5c33a0488a5d7403393e5712917253462292829b37b9320d6df82.json
+++ b/tests/integration/responses/recordings/54aa690e31b5c33a0488a5d7403393e5712917253462292829b37b9320d6df82.json
@@ -1,7 +1,7 @@
 {
-  "test_id": "tests/integration/responses/test_tool_responses.py::test_response_non_streaming_web_search[client_with_models-txt=openai/gpt-4o-llama_experts]",
+  "test_id": "tests/integration/responses/test_tool_responses.py::test_response_non_streaming_web_search[openai_client-txt=vllm/Qwen/Qwen3.5-35B-A3B-llama_experts]",
   "request": {
-    "test_id": "tests/integration/responses/test_tool_responses.py::test_response_non_streaming_web_search[client_with_models-txt=openai/gpt-4o-llama_experts]",
+    "test_id": "tests/integration/responses/test_tool_responses.py::test_response_non_streaming_web_search[openai_client-txt=vllm/Qwen/Qwen3.5-35B-A3B-llama_experts]",
     "provider": "tavily",
     "tool_name": "web_search",
     "kwargs": {

Test Plan

Run something like this before the fix:

[Note: vllm-qwen35 isn't checked in, but this is what I ran locally to repro the issue]

uv run ./scripts/integration-tests.sh --stack-config server:ci-tests --setup vllm-qwen35 --inference-mode record --subdirs responses --pattern "test_response_non_streaming_web_search and openai_client and llama_experts"

One of the gpt recordings will get updated.

Rerun after the fix. Rather than updating an existing recording, a new one is created.

[iamemilio]

there is a known flake in the ollama suite btw