Report BFCL equivalence-adjusted exact

[SonAIengine]

Summary

• add case/category/summary-level equivalence_adjusted_exact_match to the BFCL model loop
• aggregate the metric through BFCL sweeps and row-vs-retrieved deltas
• document that this is a research diagnostic, not a BFCL leaderboard replacement

Why

The near_duplicate_tool_surface subset has retrieval@5 = 1.0 but strict exact = 0.0 because BFCL contains functionally equivalent tool surfaces such as math.gcd vs number_theory.gcd. This metric keeps strict/evaluator exact unchanged while separately showing equivalent-surface value-correct matches.

Validation

• poetry run pytest tests/test_bfcl_tool_selection_llm_loop.py tests/test_bfcl_tool_selection_sweep.py -q
• poetry run pytest tests/test_bfcl_tool_selection_llm_loop.py tests/test_bfcl_tool_selection_sweep.py tests/test_bfcl_tool_selection_hard_cases.py -q
• poetry run ruff check .
• poetry run ruff format --check .
• poetry run pytest tests/ -q
• Qwen3.6-27B near-duplicate subset smoke:
  • command: poetry run python -m benchmarks.bfcl_tool_selection.llm_loop --categories simple_python --case-ids-file /tmp/gtc-bfcl-qwen027-tag-subsets/tag_near_duplicate_tool_surface.txt --top-k 5 --tool-source retrieved --model qwen3.6-27b --llm-url http://127.0.0.1:18000/v1 --disable-thinking --candidate-selection-guidance --cohesive-namespace-candidates --cache-dir /tmp/gtc-bfcl-neardup-cache --cache-namespace adjusted-metric --timeout 180 --progress --progress-every 1 --output /tmp/gtc-bfcl-neardup-adjusted-metric.json --json
  • cases: 4
  • evaluator_exact_match: 0.0
  • equivalence_adjusted_exact_match: 1.0
  • equivalence_adjusted_exact_case_count: 4
  • failure_tag_breakdown: near_duplicate_tool_surface=4

[SonAIengine]

Closing this internal stacked PR as superseded by the current main branch.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