Guard contextual BFCL extra tool calls

[SonAIengine]

Summary

• suppress model-facing contextual extra tools for single-action BFCL queries with trailing impact/effect noun phrases
• keep explicit second-action queries intact by expanding multi-intent detection
• document the recovered multiple_7 Qwen smoke and representative 4-case smoke

Verification

• poetry run ruff check .
• poetry run ruff format --check .
• poetry run pytest tests/test_bfcl_tool_selection_llm_loop.py -q
• poetry run pytest tests/ -q
• qwen3.6-27b multiple_7 row smoke: /tmp/gtc-bfcl-extra-tool-fixed-multiple7-row-qwen.json
• qwen3.6-27b multiple_7 retrieved smoke: /tmp/gtc-bfcl-extra-tool-fixed-multiple7-retrieved-qwen.json
• qwen3.6-27b recovered 4-case smoke: /tmp/gtc-bfcl-extra-tool-recovered4-qwen.json

[SonAIengine]

Closing this internal stacked PR as superseded by the current main branch. The branch is preserved for history.