[rshemet]

Summary

• align Gemma 4 assistant and tool-response turn closure with Google's updated canonical template
• continue consecutive assistant messages and completed tool loops within the same model turn
• add the thought-channel generation primer after terminal tool responses when thinking is enabled
• add native and Hugging Face golden regression coverage

Testing

• venv/bin/cactus build --python
• PYTHONPATH=python venv/bin/python -m pytest python/tests/test_chat_template_golden.py -k gemma4 -q -rs (10 passed)
• venv/bin/cactus test --component engine --suite llm --model /private/tmp/cactus-gemma4-template-verify-new --backend cpu (13 passed)