Add BFCL retrieval rank hint ablation

[SonAIengine]

Summary

• add an opt-in --retrieval-rank-hints ablation for BFCL model loops and sweeps
• record the option in reports/cache keys and prefix retrieved-source tool descriptions with graph rank hints only when enabled
• document the qwen3.6-27B smoke evidence and the next bottleneck: candidate ambiguity / call-count mismatch, not retrieval miss

qwen3.6-27B Smoke Evidence

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

Baseline artifact: /tmp/gtc-bfcl-qwen027-smoke-live.json

• row exact: 1.00
• retrieved exact: 0.85
• retrieval recall: 1.00
• milestone status: fail
• failed gates: row preservation 0.85 < 0.94, parallel_multiple 0.60 < 0.75

Rank-hint artifact: /tmp/gtc-bfcl-qwen027-rankhint-smoke.json

• retrieved exact: 0.85
• retrieval recall: 1.00
• parallel_multiple: 0.60
• aggregate improvement: none on this small smoke

Interpretation: simple rank wording is not enough; next work should focus on candidate equivalence/grouping and multi-call set construction.

Validation

• poetry run ruff check .
• poetry run ruff format --check .
• poetry run pytest tests/ -q -> 950 passed, 3 skipped