feat(eval): Add post-run anti-cheating detection and transcript audit verifier

### Summary
Add an automated transcript verifier pass that audits execution logs (`AgentResult`) for unauthorized tool invocations or shell commands attempting to read test definitions or bypass evaluation.

### Scope
1. Inspect tool calls and command history for access to task files, `task.yaml`, or benchmark internals.
2. Automatically flag and invalidate tainted benchmark runs.

Ref: Internal bug b/552484220