Summary

When _tool_guardrail_halt_decision fires the turn loop appended the halt
message to messages (conversation history) but never called
_fire_stream_delta, leaving SSE / TUI clients with a silent stream close
indistinguishable from a crash.

Root Cause

# Before (conversation_loop.py ~3432)
if agent._tool_guardrail_halt_decision is not None:
    ...
    final_response = agent._toolguard_controlled_halt_response(decision)
    agent._emit_status(...)
    messages.append({'role': 'assistant', 'content': final_response})  # history only
    break

final_response was written to internal message history but never pushed to
the stream consumers (stream_delta_callback, TTS, gateway SSE).

Fix

Add a single _fire_stream_delta(final_response) call before messages.append:

agent._fire_stream_delta(final_response)   # ← emit to all stream consumers
messages.append({'role': 'assistant', 'content': final_response})

This matches how the normal (non-guardrail) response path delivers text to clients.

Testing

• Added test_guardrail_halt_emits_final_response_to_stream to
  tests/agent/test_tool_guardrails.py — verifies _fire_stream_delta is
  called with the halt explanation text whenever _tool_guardrail_halt_decision
  is set.

python -m pytest tests/agent/test_tool_guardrails.py -v
# 1 passed

Fixes #30770