Guard BFCL cohesive namespace compression

[SonAIengine]

Summary

• guard opt-in BFCL cohesive namespace compression so it only keeps namespace families that are first-ranked or explicitly signaled by the query
• preserve singleton dotted namespaces when the query names that namespace, including simple plural forms like restaurants -> restaurant
• document the 100-case qwen3.6-27B middle sweep result and remaining 0.27 gate gaps

Model Evidence

Command:

poetry run python -m benchmarks.bfcl_tool_selection.sweep \
  --categories simple_python,multiple,parallel,parallel_multiple \
  --limit 25 \
  --top-ks 5 \
  --tool-sources row,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-qwen027-cohesive-guard-limit25-cache \
  --concurrency 6 \
  --progress \
  --progress-every 5 \
  --output /tmp/gtc-bfcl-qwen027-cohesive-guard-limit25.json

Result:

• row source: exact 0.94, strict 0.94
• retrieved source: exact 0.83, retrieval@5 0.99, strict 0.83
• milestone xgen-0.27: fail
• row preservation: 0.883
• parallel_multiple exact: 0.84
• failure categories: candidate_ambiguity:8, argument_value_mismatch:6, call_count_mismatch:2, retrieval_miss:1

Compared with the previous 100-case cohesive run, this removes candidate_not_present failures (2 -> 0) and lifts parallel_multiple (0.76 -> 0.84), but aggregate exact is still below the 0.27 gate (0.83 < 0.85).

Validation

• poetry run pytest tests/test_bfcl_tool_selection_llm_loop.py -q
• poetry run ruff check benchmarks/bfcl_tool_selection/llm_loop.py tests/test_bfcl_tool_selection_llm_loop.py
• poetry run ruff check .
• poetry run ruff format --check .
• poetry run pytest tests/ -q (954 passed, 3 skipped)

[SonAIengine]

Closing this internal stacked PR as superseded by the current main branch.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