Boost high-confidence semantic phrase near-misses

[SonAIengine]

Summary

• add guarded semantic phrase boosts for near-miss BFCL-style queries where the user phrase and tool description both provide evidence
• cover phrases such as genetic similarity, population density, highest common factor, Biot-Savart magnetic field, lawyer specialization, instrument availability, grocery criteria, historical population by state/year, and public preference
• document the updated deterministic retrieval baseline and remaining hard-case counts

Results

• deterministic BFCL recall@5: 0.94200 -> 0.95000
• deterministic BFCL all-tools@5: 0.920 -> 0.928
• deterministic BFCL MRR: 0.8217 -> 0.8305
• deterministic BFCL NDCG@5: 0.8421 -> 0.8509
• hard-case count: 80 -> 72
• expected_present_below_top_k issue count: 67 -> 59
• case-level recall changes: improved 8, worsened 0
• X2BEE scale acceptance: pass, 1084 unique tools, hit@3=1.00, selector exact@3=1.00, avg candidates=2.16, max candidates=7, avg latency=41.81ms

Validation

• poetry run ruff check .
• poetry run ruff format --check .
• poetry run pytest tests/ -q
• ARTIFACT_DIR=/tmp/gtc-bfcl-near-miss-current make research-check-deterministic
• REPORT=/tmp/gtc-bfcl-near-miss-current/bfcl-deterministic.json OUT_DIR=/tmp/gtc-bfcl-near-miss-hardcases FAILURE_CATEGORIES=retrieval_miss REPORT_TOP_KS=5 TOP_K=5 INSPECT_DEPTH=20 make bfcl-hard-cases
• make xgen-scale-acceptance OUT=/tmp/gtc-bfcl-near-miss-x2bee-acceptance.json

[SonAIengine]

Closing this internal stacked PR as superseded by the current main branch.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