τ³ task-reference fixes

编译摘要

The official audit documents tasks whose expected actions contradicted the domain policy or available tools.

This is direct benchmark-maintenance evidence that a false negative can originate in the reference contract, not the agent.

Other fixes changed ambiguity, impossible constraints, fallback behavior and policy loopholes.

Therefore the before/after score shift cannot be attributed only to a corrected answer key.

A benchmark score without task/evaluator version can become uninterpretable after task definitions change.

The minimum comparison identity should therefore bind:

benchmark version
  + task id
  + policy / environment version
  + evaluator / expected-action version
  + model / trial

对 Verifiable-Agent-Engineering：reference truth needs versioned task/policy lineage.
对 Evaluation-Integrity：benchmark maintenance can materially change measured capability without changing the model.