Fix local model tool-call loops

[TerminallyLazy]

Summary

• Add a small-local-model system prompt extension for local/Ollama-sized models such as Qwen 9B.
• Repair common malformed local-model tool JSON into canonical Agent Zero tool requests, including root-level tool args and response-intent fallbacks.
• Route explicit chat-completions tool-capable local model turns through native chat tool calls and parse returned tool calls into Agent Zero's existing execution path.
• Record successful tool calls, finalize canvas-opening text-editor writes/patches locally, and block semantic duplicate tool executions before they run again.
• Sanitize invalid Unicode surrogates in extracted tool args and text-editor file writes.

Testing

• git diff --cached --check
• docker exec charming_villani /opt/venv-a0/bin/python -m py_compile /a0/agent.py /a0/extensions/python/system_prompt/_10_small_local_model_prompt.py /a0/helpers/extract_tools.py /a0/helpers/litellm_transport.py /a0/models.py /a0/plugins/_text_editor/helpers/file_ops.py /a0/tests/test_small_local_model_prompt.py /a0/tests/test_stream_tool_early_stop.py /a0/tests/test_text_editor_context_patch.py /a0/tests/test_tool_request_normalization.py
• Direct container smoke for helpers.extract_tools.extract_tool_request malformed/local-model shapes.
• Direct container smoke for Agent duplicate-completed text-editor write detection and canvas completion finalization.
• Restarted local agent0ai/agent-zero:ready container and confirmed UI returned HTTP 200.

Note: full pytest was not runnable in this checkout because the container framework runtime does not include pytest, and the host Python environment is missing project dependencies such as litellm.