Complete the eval gate: completion regression, rerun, and multiple-comparison policies

[AbhitejJohn]

Current scope

PR #1027 completed the practical-significance work from this issue:

• Exact one-sided sign test
• 20% task-level net-win floor
• Failed-slot judge retry and frozen successful judgments
• Exact fail-closed result accounting

The remaining work is intentionally narrower:

• Add a trusted objective-completion regression hard gate when deterministic grader identity is available.
• Define the rerun policy so repeated runs cannot weaken the nominal significance guarantee.
• Define the multiple-comparison policy for suite-wide evaluation.

The completion gate remains deferred because the current infrastructure cannot identify a trusted deterministic completion grader. Scenario-level aggregation and any threshold changes must be back-tested before they become merge gates.

Related implementation: #1027.

(Copilot, updating this issue on Abhitej's behalf.)