feat(metrics): add deterministic ToolPermissionMetric

[gh-raju]

What

Adds ToolPermissionMetric, a deterministic metric that checks whether an agent only called tools it was authorized to, given a permission policy:

• allowed_tools — an allowlist (least privilege); a called tool not in the list is unauthorized.
• denied_tools — an explicit denylist; a called tool in the list is unauthorized (a denial always wins over an allow).

Score = fraction of authorized tool calls (1.0 when no tools were called). Requires no LLM / API key.

Why

ToolCorrectnessMetric compares called vs expected tools and explicitly does not check authorization. Least-privilege / tool-permission enforcement is a growing production requirement for agents — an out-of-scope tool call (delete_account, wire_transfer, …) is a safety failure regardless of task success. This makes that a deterministic, CI-gateable check.

Relates to #2825.

Changes

• deepeval/metrics/tool_permission/{__init__.py,tool_permission.py} — the metric (mirrors ToolCorrectnessMetric structure and the BaseMetric contract).
• deepeval/metrics/__init__.py — export ToolPermissionMetric.
• tests/test_metrics/test_tool_permission_metric.py — 9 tests: allowlist, denylist, denial-wins, partial credit + threshold, strict mode, no-tools, policy-required, and sync/async parity. Deterministic, so they run without any API key.

Verification

• pytest tests/test_metrics/test_tool_permission_metric.py → 9 passed.
• black --line-length 80 and ruff check → clean.

Notes

• No new dependencies. No breaking changes.
• Deterministic (no model), so it is safe to run on every PR in CI.

[vercel]

@gh-raju is attempting to deploy a commit to the Confident AI Team on Vercel.

A member of the Team first needs to authorize it.