Support Qwen/Hermes inline <tool_call> format when native function calling is unavailable

[Copilot]

Qwen models served via vLLM without native OpenAI function-calling support emit tool invocations as <tool_call> XML blocks inside the message content with finish_reason: stop, bypassing the structured tool_calls field entirely. The agent treated these responses as plain text and returned without executing any tools.

Changes

• src/openai_compat.py — Add extract_tool_calls_from_content(content): parses <tool_call>…</tool_call> blocks out of content, returning extracted ToolCall objects and cleaned content. Handles both "arguments" and "parameters" key variants. OpenAICompatClient.complete() now falls back to this when tool_calls is absent from the response.

• src/agent_runtime.py — Streaming path (_get_assistant_turn) applies the same fallback after accumulation, and patches the session message with cleaned content + properly-serialized tool calls so conversation history remains consistent on subsequent turns.

# Model response content (instead of structured tool_calls field):
<tool_call>
{"name": "write_file", "arguments": {"path": "./test_cases/TEST_QWEN_AGENT.md", "content": "test ok"}}
</tool_call>

# Now parsed and executed; file is created; content field is cleaned before storing in session history.

• tests/test_agent_runtime.py — Two new tests: unit test for client.complete() parsing the inline format, and an end-to-end test verifying the file is actually written when the model uses <tool_call> content blocks.