bug(server-https): mutual recursion between stream_llm_response and handle_tool_follow_up causes unbounded stack growth

Bug Description

stream_llm_response and handle_tool_follow_up in server-https/app.py call each other with no depth limit, creating unbounded mutual recursion.

Call Chain

stream_llm_response()
  └─► handle_tool_follow_up()
        └─► stream_llm_response()       # follow-up triggers another tool_calls
              └─► handle_tool_follow_up()
                    └─► stream_llm_response()
                          └─► ... (no stopping condition)

Root Cause

When the LLM returns finish_reason == "tool_calls" in a follow-up response, stream_llm_response calls handle_tool_follow_up again. There is no recursion depth counter or maximum hop limit — so a model that repeatedly requests tools will cause a stack overflow and crash the server process.

This is a real failure mode in agentic workloads where the LLM issues multiple sequential tool calls — exactly the multi-hop scenario this system is designed to support.

Impact

• Server process crashes under multi-hop agentic queries
• No error is surfaced to the client — connection drops silently
• Affects all users sharing the same server process

Proposed Fix

Introduce a max_depth parameter (default: 3) passed through both functions. If depth exceeds the limit, return a graceful fallback response instead of recursing further.

I will submit a PR with this fix.