Add BFCL candidate selection guidance ablation

[SonAIengine]

Summary

• add --candidate-selection-guidance to BFCL model loop and sweep runners
• record the option in report/cache keys and expose it in sweep console output
• document the qwen3.6-27B smoke result: guidance fixes the call-count mismatch class, but not parallel_multiple sibling ambiguity

qwen3.6-27B Smoke Evidence

Via go165 vLLM qwen3.6-27b, category limit 5, row/retrieved k=5.

Artifact: /tmp/gtc-bfcl-qwen027-guidance-smoke.json

• row exact: 1.00
• retrieved exact: 0.90 (0.85 -> 0.90 vs baseline smoke)
• retrieval recall: 1.00
• failed gates: row preservation 0.90 < 0.94, parallel_multiple 0.60 < 0.75
• remaining failures: parallel_multiple_2, parallel_multiple_4 candidate ambiguity

Hard-case bundle: /tmp/gtc-bfcl-qwen027-guidance-hardcases

• cases: 2
• failure categories: candidate_ambiguity:2

Interpretation: prompt selection policy can remove call-count over-decomposition, but the remaining 0.27 blocker needs candidate equivalence/grouping rather than more rank wording.

Validation

• poetry run ruff check .
• poetry run ruff format --check .
• poetry run pytest tests/ -q -> 951 passed, 3 skipped