feat(replay): add quality assertions

[luoye520ww]

Summary

• extend the existing HTTP/SSE replay benchmark with correctness assertions
• score required output, forbidden behavior, changed files, and cost budgets
• include quality violations in the existing run failure path

Why

The experimental v2 branch had a separate pure replay evaluator that was never connected to a runnable benchmark. This change folds the useful scoring behavior into the existing replay runner from #645 instead of introducing a parallel evaluation system.

Changes

• add requiredOutputs, forbiddenBehaviors, expectedChangedFiles, and maxCostUsd to replay expectations
• derive observations from real assistant and tool-call SSE items
• use file_change tool arguments for changed-file assertions
• emit a weighted quality result while retaining the current pass/fail report contract

Tests

• npm.cmd --prefix kun test -- replay-benchmark.test.ts (8 passed)
• npm.cmd --prefix kun run typecheck
• npm.cmd --prefix kun run build
• focused ESLint
• git diff --check

[XingYu-Zhong]

Reviewed and approved for merge.