Attribute BFCL row vs retrieved deltas

[SonAIengine]

Summary

• add summary.row_vs_retrieved_deltas to BFCL model sweep artifacts
• pair row-source and retrieved-source runs by repeat/top-K/category/case ID
• report row-pass/retrieved-fail counts, breakdowns, and case IDs so retrieval-layer losses are separated from row-source/model upper-bound failures
• document the latest 100-case qwen3.6-27B attribution: row-pass retrieved losses are 11 cases, mostly candidate ambiguity

Why

The 0.27 gate is currently failing, but raw retrieved failures mix two different problems:

• failures that also happen with row-source tool lists, which are model/tool-schema upper-bound issues
• failures that only happen after graph retrieval/presentation, which are the actual graph-tool-call search/presentation losses

This change makes that split explicit in every sweep artifact so the next hard-case subset can focus on the cases graph-tool-call can actually improve.

Evidence

Cached rerun of the 100-case qwen3.6-27B middle sweep:

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 25 \
  --output /tmp/gtc-bfcl-qwen027-cohesive-guard-limit25-paired.json

Output includes:

row_vs_retrieved k=5 repeat=1 paired=100 row_pass_loss=11 retrieved_on_row_pass=0.883 retrieval_layer_failures=argument_value_mismatch:2,candidate_ambiguity:8,retrieval_miss:1

Validation

• poetry run pytest tests/test_bfcl_tool_selection_sweep.py -q
• poetry run ruff check .
• poetry run ruff format --check .
• poetry run pytest tests/ -q (955 passed, 3 skipped)