Endless Web-Search Loop

## Describe the bug
When using Agentic Chat / Chat Assistant to research a topic, the LLM enters an endless web-search loop — it repeatedly calls the web search tool instead of stopping to answer. This happens even when the prompt explicitly instructs it to use no more than 5 sources. There does not appear to be any user-facing setting to cap the number of tool calls/searches per turn, and no automatic loop/iteration limit kicks in to stop it. Proceeding

## To Reproduce
Steps to reproduce the behavior:
1. Open AI Playground and select the Agentic Chat / Chat Assistant preset with web search tool enabled, using Qwen3.5-9B via OpenVINO
2. Enter a research-style prompt: Research [topic] do not use more than 5 sources.
3. Observe the chat as it runs
4. The model repeatedly invokes the web search tool well past 5 calls and never returns a final answer — session must be manually stopped.
5. Any subsequent prompts in the session fail displaying an error: HTTP 400: {"error":"Mediapipe execution failed. MP status - INVALID_ARGUMENT: CalculatorGraph::Run() failed: \nCalculator::Process() for node \"LLMExecutor\" failed: Failed to apply chat template. The model either does not have chat template or has an invalid one."}. Possible chat history corruption by the aborted tool call?

## Expected behavior
The LLM should either respect the stated source limit, or (failing that) the app should enforce a hard cap on tool-call iterations per turn so a single agentic session can't loop indefinitely.

## Screenshots

## Environment (please complete the following information):
- OS: Windows 11
- GPU: Intel Arc Pro B50
- CPU: i5-13400
- Version: 3.2.0 beta
- Backend: OpenVINO
- Model: Qwen3.5-9B

The issue names no implementation files or tests. Start by reproducing the Agentic Chat or Chat Assistant flow in AI Playground with web search enabled, using Qwen3.5-9B through OpenVINO, then trace the tool-call loop and the failed session state after stopping it. Done means a turn cannot search indefinitely and later prompts remain usable after a loop is stopped.

Written by the indexing model from the issue text.