Summary

• add a default xgen-0.27 milestone gate to BFCL model sweep artifacts
• include per-category sweep rows and category repeat groups so parallel_multiple bottlenecks are visible without hand-calculation
• document how to use the gate before deciding whether to run another full expensive benchmark

Why

Full BFCL model runs can take hours. This makes the sweep artifact answer the immediate research question: did retrieved k=5 preserve enough of row-source performance for the 0.27 XGEN target, and is parallel_multiple still the blocking category?

Validation

• poetry run ruff check .
• poetry run ruff format --check .
• poetry run pytest tests/ -q -> 949 passed, 3 skipped
• CLI smoke: poetry run python -m benchmarks.bfcl_tool_selection.sweep --categories simple_python --limit 1 --top-ks 5 --tool-sources retrieved --model fake --llm-url http://127.0.0.1:9/v1 --timeout 1