feat(bench): add BFCL function-calling adapter

[drewstone]

Problem

The external-benchmark roadmap now selects BFCL as the cheapest public size-axis benchmark for the policy-vs-free-form contrast, but @tangle-network/agent-bench does not yet expose a BFCL adapter.

Desired adapter

Add a fail-loud bfcl adapter under bench/src/benchmarks/ that delegates to the official Berkeley Function Calling Leaderboard assets/evaluator where possible. The first useful scope is deterministic function-call categories suitable for weak-vs-gold calibration and small-vs-large model comparisons.

Constraints

• Do not fabricate BFCL scores.
• Fixture mode may test adapter plumbing, but official mode must require the real Gorilla/BFCL checkout or bfcl-eval project root.
• Keep BFCL V4/V3 naming current: V4 is the latest official line, while V3 multi-turn/missing-function categories are the immediate research target if they remain the best fit.
• Expose through the existing ADAPTERS map; no runtime-loop changes.
• Add preflight/load/judge tests that fail loud without official assets and pass on fixtures.

Research use

This enables the AppWorld + BFCL easy-subset-first tranche: AppWorld serves external validity through stateful code/API orchestration, BFCL serves the size axis for function-calling and missing-function/free-form contrast.

[drewstone]

agent-lab PR #79 now treats BFCL as the planned size-axis benchmark row and points here as the adapter blocker.

Desired next implementation remains narrow: add a fail-loud bfcl adapter under bench/src/benchmarks/, expose it through ADAPTERS, support fixture plumbing only for tests, and require official BFCL/Gorilla assets in live mode. No runtime-loop changes.