feat: expand benchmark coverage and add BFCL v4

[aabbdev]

Summary

• complete Wave 1 coverage with C-Eval, full BBH, OCRBench v1, Global PIQA, current MathArena tasks, GSM-Hard, EvalPlus, and LiveCodeBench v6
• add hardened Docker execution and mandatory CI smoke tests for LiveCodeBench, EvalPlus, and BFCL
• introduce provider-neutral function-calling primitives for schema conversion, parsing, typed matching, parallel calls, and stateful execution
• add BFCL v4 single-turn (3,641 samples), official multi-turn checking (800 samples), frozen agentic memory/web tasks (665 samples), and a 5,106-sample offline weighted aggregate
• pin upstream revisions, release assets, checksums, and Docker base images; document protocol adaptations and blocked historical-score equivalence

BFCL naming

bfcl_v4_multi_turn uses the pinned official BFCL state/response checker. The agentic environment is deliberately exposed as bfcl_v4_agentic_offline, and the weighted aggregate as bfcl_v4_offline, because frozen evidence is reproducible but not numerically identical to the live SERPAPI leaderboard environment. No ambiguous bfcl_v4 alias is introduced.

Validation

• pytest -m "not integration and not docker" — 444 passed, 2 skipped
• pytest -m docker -v tests/integration/test_docker_sandboxes.py — 3 passed
• mypy . — 466 source files clean
• pre-commit run --all-files — all hooks passed
• uv build --quiet — wheel and sdist built
• one-sample end-to-end mock runs completed for BFCL multi-turn and agentic tasks

Limitations

Real-model transport checks across two providers could not be run locally because no provider credentials or local model runtime were available. The provider-neutral paths and mock transport are covered, and the tasks are ready for credentialed CI/manual evaluation.

[github-actions]

✅ Benchmark documentation has been automatically updated.